OPTICAL IMAGE SCANNERS AND CHARACTERRECOGNITION DEVICES:

A SURVEY AND NEW TAXONOMY

Amar GuptaSanjay HazarikaMaher Kallel

Pankaj Srivastava

Working Paper #3081-89

Massachusetts Institute of TechnologyCambridge, MA 02139

ABSTRACT

Image scanning and character recognition technologies have maturedto the point where these technologies deserve serious considerationfor significant improvements in a diverse range of traditionallypaper-oriented applications, in areas ranging from banking andinsurance to engineering and manufacturing. Because of the rapidevolution of various underlying technologies, existing techniquesfor classifying and evaluating alternative concepts and productshave become largely irrelevant. A new taxonomy for classifyingimage scanners and optical recognition devices is presented in thispaper. This taxonomy is based on the characteristics of the inputmaterial, rather than on speed, technology or application domain.

2

1. INTRODUCTION

The concept of automated transfer of information from paper

documents to computer-accessible media dates back to 1954 when

the first Optical Character Recognition (OCR) device was

introduced by Intelligent Machines Research Corporation [1]. By

1970, approximately 1000 readers were in use and the volume of

sales had grown to one hundred million dollars per annum [3]. In

spite of these early developments, through the seventies and

early eighties scanning technology was utilized only in highly

specialized applications.

The lack of popularity of automated reading systems stemmed

from the fact that commercially available systems were unable to

handle documents as prepared for human use. The constraints

placed by such systems served as barriers, severely limiting

their applicability. In 1982, Ullmann [2] observed:

"A more plausible view is that in the area of characterrecognition some vital computational principles have not yetbeen discovered or at least have not been fully mastered.If this view is correct, then research into the basicprinciples is still needed in this area."

However, ongoing research in the area of automated

reading (conducted as part of the wider field of pattern

recognition) is leading to new products, and the market for

scanning devices has expanded dramatically. Concomitantly,

there has been tremendous growth in the market for third-

llI

3

party scanning software, as well as software and peripherals

for specialized tasks such as handwriting recognition and

data indexing, storage and retrieval [3].

This paper analyzes the technology used in the image

scanning and character recognition industry and presents a

taxonomy for evaluating the capabilities of current systems.

Although the main focus is on optical scanning and character

recognition technologies, an attempt is made to identify

auxiliary technologies and products as well. In addition,

the paper discusses future trends and characteristics of

systems which are likely to become available in the next ten

years.

Inclusive of this introductory section, there are nine

sections in the paper. Section Two presents a discussion of

the principal recognition approaches and technologies. The

third and fourth sections classify the scanning industry by

market segment and function respectively. Section Five

discusses ancillary technologies such as scanning software,

data storage, indexing and retrieval products and

handwriting recognition software. In Section Six, future

trends in products and technologies are discussed. In

Section Seven, a new taxonomy is developed for evaluating

the capabilities of current scanning systems. Section Eight

describes a set of representative scanners and applies the

4

taxonomy to judge their performance. Finally, the

conclusions of the paper are presented in Section Nine.

III

2. RECOGNITION TECHNIOUES

The recognition of text, the scanning of images, and

the raster to vector conversion of technical drawings have

usually been considered independently of each other. The

technologies corresponding to these three areas must be

integrated in order to accurately scan and process complex

technical documentation. In the case of most non -

technical applications, only the first two areas need to be

considered. A framework that allows recognition of both

text and images is presented in this section.

The three major stages in the processing of a document

are preprocessing, recognition, and post-processing. These

are discussed in the following subsections.

2.1 Preprocessing

Preprocessing is the conversion of the optical image of

characters, pictures, and graphs from the document into an

analog or digital form that can be analyzed by the

recognition unit. This preparation of the document for

analysis consists of two parts: image analysis and filtering

[4]. The significance of these parts is described in the

following paragraphs.

(a) Image Analysis -

The first stage of image analysis is scanning [5,9,25].

Scanning provides a raster image of the document with

sufficient spatial resolution and gray scale level for

subsequent processing. In the case of a picture or a graph,

the issue of gray scale level is more important that in the

case of text. For text, this phase consists of locating

character images. With the exception of high end scanners,

which employ contextual analysis as well, reading machines

are character-oriented. Each character is treated as a

unique event and is recognized independently of other

characters. This implies that the document must first be

segmented into distinct characters, and then the identity of

each character recognized separately.

The process begins with the optical system taking a

raster image of the area that is supposed to enclose the

character. Alternatively, the raster image representing the

character is "cut out" of the image of the document. In

either case, the image is transmitted sequentially to a

single-character recognition subsystem. If the image, or

the information on the features of the character

constituting the image, possesses characteristics which are

significantly different from the characteristics maintained

by the character recognition subsystem, then the particular

area is deemed to be either an unrecognized character or

ill

noise. Depending on the system, the output is expressed as

a flag or a blank. In some systems, any graphic or

character input which is outside the size limitations is not

flagged but skipped. The formatted output, in such a case,

contains a blank zone corresponding to the input of the

improper size.

(b) Filtering

After the image is analyzed, filtering takes place.

Filtering minimizes the level of noise. The latter may be

caused either in the source document or by the opto-

electrical transformation mechanism. The process of

filtering also enhances the image for easier recognition.

One filtering process that eases the extraction of the

features of the character in the recognition phase was

proposed independently by Lashas [5] and Baird, et al. [6].

They presented two OCR readers in which black marks

constituting the character are transformed into quantized

strokes. Their approach is depicted in Figure 1.

Preprocessing: New Approaches

The preprocessing phase consists of deriving a high

level representation of the contents of the image. The

scanned document is seen as a set of blocks corresponding to

independent ways of representing information such as text,

line drawings, graphs, tables, and photographs.

a.At4

Is tr .*A, t_#4*1 w"O-4 I +RIO-.. , aw 44 f wltv

ot- ,; ,

.PPdB*- -- -- -ALM

CCR-A font numerals and horizontal and verticalquantized stroke segments

.C

..or

CCR-B font numerals and fourquantized strokes

-"

_ 11111

=ill i

11111t d-- -

detected directions of

FIGURE 1: QUANTIZED STROKES

Lashas, A., et. al. "Optical Recognition Based OnAnalog Preprocessing and Automatic FeatureExtraction". Computer Visin, Graphics andImage Processinq.vol. 32. pp. 191 - 207.

t,- u-~~~

1a

"NO&

r-

III

understanding of this document involves the following:

(a) Identification of the major blocks of

information.

(b) Identification of the spatial relationships between the

different blocks (for example, the reading order must be

recognized so that the logical connections between different

blocks of text or graphics can be easily derived).

(c) Identification of the layout features such as the number

of columns, margins and justification.

(d) In the case of text, further identification of

headlines, footnotes and other similar characteristics.

Typically, the textual portion is first distinguished

from other information, and then columns of text are

recognized. Next, these columns are split into lines of

text which are in turn segmented into single character

images.

A reader which is able to accept a free format document

was described by Masuda et al. [7]. Their scheme of area-

segmentation uses projection profiles obtained by projecting

a document image on specific axes. Each profile shows the

9

structure of the document image from a particular angle.

The projection profile is very sensitive to the direction of

projection. In order to check for skew normalization, the

image of the document is incrementally rotated and the

horizontal and vertical projection values are noted.

Through an analysis of the intensity of these values and an

examination of different areas, general text and headlines

are separated (Fig. 2). Another reading system based on

this approach is described by Kida, et al. 8].

Commercial system today require areas to be manually

segmented [2,5,7]. While segmentation methods are fairly

rudimentary in low-end scanners, methods that permit the

operator to define several text and graphics "windows"

within a page are available on high-end products. These

methods include the use of a mouse or a light pen to define

text, graphic, or numerical zones.

2.2 Recognition

Recognition occurs at the level of characters in most

commercial page readers. However, high-end scanners are now

complementing character recognition with sophisticated

contextual analysis. Techniques used for character

recognition and contextual analysis are discussed in the

following paragraphs.

_-- _ .r__ , ---_~ == Lw

_w= == -_ _ nm - - . -

-J a __ 5~mWmm~ dS a .4mg _ _ _ _- -r~· -- mm - -i iIii . ii-

m am

- __ ._ _

'" ~-- '~ . - ,.mmN - ¢ws-_mm,^-cs,~~_ 3E3 , E_* ps

Projection Profiles

Text line areas extraction

FIGURE ':

Ma suda,

SEPARATION OF TEXT

-I. "Approach to a Smart DocumentRead,-- ' F? Conference on Computer Visionand ? ~ r--n Recognition. 1985. pp. 550 - 557.

S0

__.· ._ __ _ __i__

- W

I

III

I"a

10

Character Recognition

The character recognition problem is essentially one of

defining and encoding a sequence of primitives that can

represent a character as accurately as possible. The most

common approaches to character recognition are:

(a) template matching, and (b) feature extraction [9].

(a) Template Matching Technique

Among the oldest techniques for character recognition,

template matching involves comparing the bitmap that

constitutes the image of the character with a stored

template [9]. The amounts of overlap between the unknown

shape and the various stored templates are computed, and the

input with the highest degree of overlap is assigned the

label of the template.

The primitive in this case is a very simple function

that assigns one value to a point of the bitmap if it is

black, and another if it is white 5,9,25]. The performance

of different readers using this technique depends on the

decision algorithms. This method offers high-speed

processing especially for monofont pages. Even in the case

of documents containing several fonts which are extracted

from a limited set of fonts of a given size, this technique

can be very effective. Moreover, this method is relatively

immune to the noise generated during the opto-electrical

Ill

11

transformation.

However, this method is less effective in the case o

unknown fonts, multiple fonts and multisize characters.

Further, it imposes constraints on the format of the page in

areas such as the spacing of the letters and the position of

the text. If these constraints are relaxed, template

matching becomes slow and costly, because of the need to

translate, scale, and rotate input images.

(b) Feature Extraction

In contrast to the template matching technique which

emphasizes the importance of the overall shape, the feature

extraction approach focuses on the detection of specific

components of the character [5]. This approach assumes that

local properties and partial features are sufficient to

define every character.

Feature extraction techniques identify local aspects

such as pronounced angles, junctions and crossing, and

define properties such as slope and inflection points. In

one method of recognition, a boolean or a numerical function

that characterizes each feature is calculated and then

applied to the given image [5,7]. Another method involves

defining a partial mask that can be systematically

positioned on the pattern [7]. In a more recent strategy,

12

recognition is based on the analysis of the direction and

connective relationships of the strokes of the character

[7].

Structural Analysis Methods

The use of structural analysis methods is a recurrent

theme in the literature [1,4,5,6,7,10]. These methods are

frequently utilized in commercial reading machines. Each

character is defined as a set of topological primitives such

as strokes, segments, holes, and arcs, and these machines

analyze the characteristics of the scanned inputs to detect

such primitives.

Isolated, unbroken characters are first recognized

using a structural description 9]. These descriptions are

independent of the position and the size of the character.

The shape is then parameterized so that the results of the

structural analysis can be compared with a stored list of

shapes. Next, the shapes are clustered to discriminate

between characters and to classify them. The power of

structural analysis depends of the number of features of

each character used by he system. An example of the

structural analysis -f a character is shown in Figure 3.

I*

]

+

K

C4

lm

C'O

C0;

aI(A

oL

a-

00I)O

0_

La

(n

II[

4Lu.0Co

)-i

f-4n

0z4nI-

I)U,0rr

.xm

(3

U)U.LU0i

0

LI I

0

13

Character Classification

So far, the discussion has centered on single

characters in a single font. Most documents contain

multiple characters with considerable variation across

fonts. A statistical approach is used for dealing with

these variations. In reading machines equipped to handle

multiple fonts, tests of character recognition are based on

statistical data constituted by a training set and a test

set. When the term "omnifont" is applied to a reader, the

implication is that the reader has been trained on a large

set of fonts [6].

Feature extraction is generally complemented by

classification. In order to be classified, characters have

to be discriminated from each other. The classifier is

designed to construct a discriminant function underlying a

complex probability distribution, taking into account the

non-linear variations in the characters [26].

The two essential steps in recognition - feature

extraction and classification - have different optimization

criteria. It is felt that feature extraction should be

optimized with respect to the reconstruction of the

character, while classification should be optimized with

respect to recognition 4]. Consequently, feature

extraction should not be directed towards recognition

14

through definition of features that minimize the

classification process.

Classifiers are usually adapted to different fonts by

the manufacturer. However, some machines provide a

classifier which allows for the adaptation of new fonts by

the operator. An on-line training capability for multifont

recognition can also be provided 4]. This concept of

trainability is similar to that employed in modern speech

recognition systems, which overcome the variability of voice

between speakers by requiring that adaptation be performed

during a training phase prior to regular operation. The

main disadvantage of trainable systems lies in their slow

training speed, as the training set contains several

thousands of characters.

Limitations of Current Recognition Techniques

Ideally, feature extraction techniques should be able

to generate a set of characters that are independent of font

and size. However, the wide variety of fonts encountered in

office environments results in a huge number of possible

characteristics. Furthermore, ambiguities occur due to the

similarity in the features of different characters across

fonts. For example, the distinction between 1 (one , I

(capital I) and 1 (el) across and even within fonts is not

obvious. Moreover, there are a number of characters which

15

cannot be correctly identified without further information

about their size or position. For example, the character

"O" (capital o) in one size corresponds to the lower case

"o" in a larger size; further, this character is the same as

0 (zero) in several fonts [11].

The user can control the accuracy of the system by

adjusting the error and reject rates [12,25]. The error

rate is the percentage of the characters in a document that

are incorrectly identified by the system. The reject rate

is the probability that the system cannot identify a

character at the desired level of confidence (i.e., if the

reject rate is 5%, the system will flag 5% of all

characters, on the average, as being unrecognizable).

Clearly, there is a tradeoff between the error rate and the

reject rate - if a lower reject rate is selected, the error

rate will increase and vice versa [12,25]. Consequently, a

reader may possess a very low error rate, but it may flag or

reject every character that does not offer a high

probability of correctness as defined by the decision

algorithm (13].

There is naother aspect to accuracy. For example, a

reader may recognize the characters in "F16" correctly, but

then flag them as rejects because the word is not in its

built-in dictionary. Also, the inability of readers to

16

distinguish between ambiguous characters such as "I"

(capital ai), "1" (el) and "1" may or may not be considered

to be an error.

The error rate of the entire recognition process is

dependent not only on the functioning of the single

character recognition subsystem but also on the level of

preprocessing [12]. For example, different kinds of

segmentation errors may occur, and lines of text may be

missed or misaligned in the case of a document containing

several columns. Situations involving such missed or

misaligned lines can be minimized by preprocessing

(described in Section 2.1).

With respect to performance evaluation, character

recognition is sometimes deemed to be a branch of empirical

statistics [4]. There is no reliable way of modelling the

accuracy of a reading machine except by comparison with a

standard set of norms. The impracticability of statistical

modelling is due to the fact that the pattern generating

process and its multivariate statistics are influenced by a

number of barely controllable, application-dependent

parameters.

Recognition techniques are prone to ambiguities, apart

from involving heavy computation. The sole reliance on the

17

physical features of the characters results in a high rate

of errors and rejects. To combat this problem, recognition

methods are complemented by contextual and syntactical

analysis aided by customizable lexicons, especially in high-

end systems [9,11,13].

In the case of low quality documents with different

kinds of noise and broken and touching characters, the error

rate is high. Contextual analysis becomes a necessary

condition for reducing the rate of errors and rejects in

such situations.

Separation of Merged Characters

The breaking of images into character blocks is based

on the assumption that each character is separated by

horizontal or sloped line of blank spaces. However, in the

case of tight kerning, inadequate resolution of the scanner,

poor quality of the document, or high brightness threshold,

adjacent characters spread into each other. Separating

merged characters involves the following three processes:

(a) Discriminating multiple characters from single character

"blobs";

(b) Breaking the blobs into components that are identified

as single character blobs;

(c) Classifying the different characters and deciding

18

whether to accept or reject the classification.

Defining a set of criteria for distinguishing between

adjacent characters is difficult because of the many ways in

which characters merge together and the fact that merged

characters contain misleading strokes. Separation of

characters is a computationally intensive task that

significantly slows down the overall recognition process

[2,4,8,121. Separation is usually possible in a limited

number of cases and for pairs of characters only.

Contextual Analysis Techniques

Contextual analysis is of two types: layout context

analysis and linguistic content analysis 9]. The former

covers baseline information on the location of one character

with respect to its neighbors. For example, it generates

formatting information and is usually language-independent.

Linguistic analysis on the other hand involves the use of

spelling, grammar and punctuation rules. For example, a

capital letter in the middle of a word (with lower case

neighbors) is not accepted. Layout context analysis

capabilities are available on several systems available

today. However, only the high-end systems offer some degree

of linguistic analysis capabilities.

19

Dictionary Lookup Methods

Several contextual methods for word recognition are

based on the analysis or comparison of a word or string of

characters with stored data for type and error correction.

Spelling checkers are commonly used to complement character

recognition devices [13]. Such checkers may be part of the

recognition software, or they may be part of the text

composition software that the operator uses for correcting

the processed document. In either scenario, when the size

of the dictionary is large, the search time can be very

long. Furthermore, if the contextual information is not

considered, the scanned word may be incorrectly converted

into some other word (13]. Several high speed correction

methods use the concept of "similarity measures", which are

based on the fact that most errors are due to character

substitution, insertion, or deletion [22]. The similarity

between two words (the correct word and the garbled word) of

equal length is measured using the Hamming distance. The

Modified Levenstein distance [9] generalizes the similarity

measure for substitution, insertion and deletion. The

minimization of this distance forms the optimization

criterion. Tanaka [14] proposed two methods that yield 10 -

35% and 35 - 40% higher correction rates than a typical

dictionary method and reduce computing time by factors of 45

and 50 respectively. One of these methods is based on the

assumption that different characters can be classified into

III

20

classes or groups that are independent of each other, so

that a character in one class is never misrecognized as a

character in another class. This categorization helps to

significantly reduce the probability of errors.

Use of Statistical Information

The frequency of occurrence of different characters

varies widely. For example, the letter "e" has the highest

frequency of appearance in an English document. Also, the

letter "t" is the most frequent first letter of a word, and

the letter "q" is always followed by a "u." Further,

several sequences of character combinations are more likely

to appear than others. The frequency of occurrence of a

character within a given string of text can be efficiently

modeled by a finite-state, discrete-time Markov random

process [22]. The use of statistical distributions of

different combinations of N characters (N-grams) allows the

implementation of error corr tion algorithms such as the

Viterbi algorithm. According to Hou 9], this algorithm can

reduce the error rate by one half.

Linguistic Context Analysis

Characters are primitives of strings constrained by

grammatical rules. These rules define legitimate and non-

legitimate strings of characters. Based on the recognition

of some words, the class (noun, verb, etc.) to which they

21

belong can be identified along with the applicable syntactic

analysis to identify misspellings, misplacements, and other

syntactic errors. A string of words or a sentence can be

decomposed using a parsing tree. There are several

efficient parsing trees for different types of grammatical

structures [9,12,15].

Syntactic analysis assumes that the text is constrained

by one type of grammar. This assumption need not hold in

all cases. Several technical documents contain uncommon

words: serial numbers, technical words, and abbreviations.

Such situations require interaction between the technical

operator and the system, or the use of dictionary lookup

methods. Except in such special situations, linguistic

context analysis techniques are well suited to the task of

identifying and correcting reading errors [15].

Future of Recognition Technology

Conventional techniques for character recognition based

solely on geometric or analytical properties are not

sufficient to accurately process complex documents at high

speed [22,25]. While the use of contextual analysis

improves accuracy, it does not increase processing speed.

To improve speed, i- becomes necessary to simultaneously

analyze the document from several points of view. For

example, one unit can analyze the image, another can

III

22

recognize characters, and a third can perform contextual

analysis of the text.

The use of a single recognition technique is

insufficient to solve all ambiguities encountered during the

recognition process. As such, general algorithms must be

complemented with customized rules for special cases, such

as distinguishing letters that are easily confused (e.g.,

"o" and "0" and "6" and "b"), or recognizing characters

printed in pieces (such as "i", "jn", and ";"). Ideally, a

single process control unit should collect all the pieces of

information from the subsystems, order them, and then make

appropriate final decisions [4]. Such a strategy

(functionally depicted in Figure 4) allows the recognition

of different parts to be done in parallel.

2.3 Postprocessing

After the text and graphics are recognized, they must

be encoded for transmission to, and processing by, a word

processor, a graphic editor or a desktop publishing program

(5]. In the case of text, the image is most commonly

converted into the ASCII format. In the case of graphics,

the storage requirements for the document vary greatly,

depending on the extent of compression. The storage

capacity may pose a major constraint in some cases. For

example, in systems which scan at 300 dots per inch (dpi),

one single 8.5" x 11" page requires over one megabyte of

LUI-

z I-w O

LU -

n CC0

mC )

W

E

CCS0

m

<LU;. -LU,,=

U) cn*

gC

a

LI.

v coCC

o'ar- C

,;0,.2.-_ __S.

C

0 _

SU'.

IOLL_4_)

t._J

C

SW GS.

Co3 ca 0b.P cc

:_ PuJ m

Ia cC; ,a I-S

*r LIC,"S_

O0

0'I;

0

S

o

C

I-S

&6

Ch.0.

0a

a0L

r)

oO

0

E

o

0Q

U

w K

0 U o

--

O3

IAFq

-0

op*,r

�

I�S,\1 r-'---'-i ! t

cx -

-)1

23

memory if stored in raw bitmap format. While the advent of

optical disks with storage capacities in gigabytes tends to

mitigate the storage problem to some extent, it is usually

desirable to use more efficient storage strategies.

24

3. CLASSIFICATION OF SYSTEMS

For analytical purposes, the scanning industry can be

divided into four segments: (a) Low-end scanners; (b) High-

end scanners; (c) Integrated systems; and (d) Software. The

characteristics of each segment are explained in the

following paragraphs.

(a) Low-end scanners

These are small scale systems, usually desktop-based

with relatively low sophistication in dealing with page

scanning and image processing applications (1]. They are

dependent on a host computer system for processing and

storage (commonly an IBM personal computer or a Macintosh).

The capabilities of these systems are enhanced by third

party image and text processing software that is generally

executed on the host computer.

Low-end scanners are either automatic feed or flatbed,

and can handle between 20-100 documents per minute (or 20-50

pages per minute). The recognition technology tends to be

simple - generally the matrix matching technique is used.

Costs are low, ranging from $1,800 to $7,000. The software

cost averages around $2,000 [3]. These scanners are geared

towards low volumes (10 - 100 pages or documents at a time),

and relatively low accuracy (1 - 10 letters per page misread

II1

25

or flagged for error on the average). The Scan-Optics Easy-

Reader 1720 and the Datacopy 730 are examples of popular

low-end scanners.

Most of these machines can scan a page or a portion of

a page as an image or as characters of text, depending on

the intent of the user. In some cases, the user can specify

which zones within the input page contain text and which

zones include graphic images. Some of the systems use

separate phases for scanning text and images respectively.

Hand held scanners have entered as the lowest end

products in the scanner market 3]. These usually offer

limited page and text reading capabilities without requiring

any additional software. They are either plugged directly

into the personal computer like a keyboard or linked by a

simple interface (such as a parallel card). They are

transparent to the host computer, its operating system, and

its application software. The volume handled is very low,

because of the need to physically move the scanner with the

hand. The scanning speed ranges from about 50 to 200

characters per second, and the number of onts recognized is

limited to 5 or 10. Their costs are quite low ($200 -

$1000) [161. The Caere Corp OCReader is an example of a

widely used hand-held scanner.

26

(b) High-end scanners

These are large, sophisticated image/page/document

processors with stand-alone or host based capabilities [1).

They offer mature hardware and software devoted to the tasks

of text recognition and image scanning. These systems

include sophisticated dedicated processors, editing and

correction workstations, as well as customized recognition

software that uses sophisticated techniques like feature

extraction [30). They support their own storage systems and -

can be readily integrated with other common systems.

High-end scanners feature multi-font and multi-size

flexibility to handle multiple fonts and multiple sizes

along with automatic feeding capability that can cope with

high volumes. Large numbers of documents (1,000-10,000) can

be accurately processed at high speeds (200 or more pages

per minute or 750 or more documents per minute). Costs

range from $15,000 upwards with the median falling between

$25,000 - $30,000 [16]. Examples of high-end scanners

include the Calera CDP Series and the Kurzweil family of

scanners.

(c) Integrated Systems

These systems serve as automated image processing

facilities for the office, and are designed to network with

III

27

the existing computer and communications facilities. A

typical system, controlled by a common operating system such

as UNIX, consists of an enhanced office network that

provides multi-media, multi-user, and multi-tasking access

to its diverse resources 171.

An integrated image system offers various data entry

facilities such as optical character recognition devices

(both low and high end), image scanners, keyboard

workstations and facsimile equipment. The system is usually

managed by a dedicated processor (a mainframe or a

minicomputer), although LAN-based systems are becoming

increasingly popular. The processor is procured from a

vendor, or the system is configured to operate within the

existing computational facilities. Data storage is

centralized and is accessible from all the workstations.

Software is executed either on the processor or on a

decentralized basis [17].

Typically, an integrated system will support a variety

of software packages, including "off the shelf" word

processors, database and statistics programs, and other

professional software tools. Similarly, at the output level

there is support for a variety of printers, monitors, and

facsimile equipment.

28

Several integrated imaging systems are available,

including the Scan Optics ScanEdit 3200, the Wang Integrated

Image Processing System, and the DRS 4000. IBM has

announced its own system, the Image Plus, which is expected

to be released in 1990. Figure 5 provides a schematic

representation of a typical integrated imaging system. This

paper does not offer a detailed treatment of integrated

systems, as they represent an amalgamation of new imaging

technology with conventional computing and communications

power, rather than a new technological innovation.

(d) Software

Third party software tools offer powerful scanning and

processing capabilities for both images and text. As

mentioned earlier, they are substantially enhancing the

capabilities of low-end scanners. In addition, software

packages utilizing new approaches from areas such as

database management, expert systems and neural networks are

being developed for specialized applications [1]. Most

packages are compatible with the popular desktop scanners as

well as with Macintosh and IBM personal computers.

The orientation of this paper precludes a detailed analysis

of software products. However, Section Seven provides a

brief treatment of currently available products.

0n (00

4441

0eaz

tit

C

zUl

I

(.

zLUI-

LL

CL-

0.0

-]

OC0t ,o-0

oCC"

c

0 00 Q00a-

40

0

=

* 0

zoZ c

D-4Aa:'UA XX -L <K

0

0O

oz

I-0 0cc(L:I

I-0occ0

z0PI-4I-

O0

z0

4cI-

be

o3.

III

0 a

a" a

(jo. co

I

I

29

4. FUNCTIONAL CLASSIFICATION OF MACHINES

Apart from classifying products by market segment, it

is feasible to classify the image and character recognition

industry by function. Thus, "reading machines" can be

divided into four major categories: (a) Document Readers;

(b) "Process Automation" Readers; (c) Page Readers; and (d)

Image Readers. These groups are described in the following

paragraphs.

(a) Document Readers

This category of readers was the first to be

introduced. Developed during the sixties and seventies,

these machines are oriented towards transaction processing

applications such as billing and form processing [4,8,251.

In earlier machines, the source document was prepared in a

rigid format using a stylized type font and the character

set was limited to numeric characters plus a few special

symbols and sometimes alphabetic letters. Contemporary

machines have greater flexibility in font recognition and in

coping with material of poor quality generated by high speed

printers. The speed of processing is very high, typically

between 400 and 4000 characters per second. An important

feature is on-line correction with the help of bitmaps of

unrecognized characters.

30

(b) Process Automation' Readers

The main goal of these readers is to control a

particular physical process 4,8,9,25], such as automatic

sorting of postal letters. Since the objective is to direct

each piece of mail into the appropriate sorting bin, whether

or not the recognition process results in the correct

decision for every single character is not critical in this

case. Since a reader of this type is designed with specific

applications in view, it is not explicitly considered in

this paper. However, the class of integrated systems

discussed in Section Three above is capable of performing

"process automation" functions.

(c) Page Readers

These reading machines were originally developed to

handle documents containing normal typewritten fonts and

sizes (4,8,9,25]. Initially intended for use in the

newspaper and publishing industries, these machines were

designed with the capability to read all alphanumeric

characters. Until the late seventies, devices of this type

were quite restrictive in terms of their requirements for

large margins, constrained spacing, and very high quality of

printing. Further, they accepted only a small number of

specially designed fonts such as OCR-A and OCR-B. The

reading speed for a monofont page was several hundred

characters per second and the price of the machine itself

31

was around one hundred thousand dollars per unit. The above

situation was significantly altered by the introduction of

the Kurzweil Omnifont Reader in 1978. As its name implies,

this machine was able to read virtually any font. Since

then, several other machines belonging to this category have

been introduced, and their capabilities have been

substantially enhanced. For example, in many cases the user

can specify the particular zones on the page to be read, and

the format of each zone. On-line error correction is

another common feature. In addition, prices have fallen

steeply to under twenty-five thousand per system.

(d) Image Readers

These machines were originally designed to meet

Computer Aided Design (CAD) needs, and subsequently came

into more generalized use. They capture items such as

drawings in the form of their mapped image and then

reprocess that image to generate an equivalent vector

graphics file.

The functional classification described above is fast

becoming obsolete. As the matrix in Figure 6 indicates, one

of the main trends in the scanning industry is the

convergence of separate functions within a single system

[1,3]. One machine can now perform functions that

previously required multiple machines. An equally important

II'

Z

o-cn

un 2 E:

00 a

i 0 .mQtu~

tLUWj a

LUW0cw a.ee 3 U

LLDZOI-- OZ<(.

LA

LU

ot2

WI-

z

U)

LUI-Co

U)x

z0

o(0U.

(I

u.

0~

X X X X

X X,=,~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

0

I-

zZ

z

z

z

U)

I

LUI...

Zr

ZrZcZ3

C zZ wLU Z

Zi

otn./-J~nCl) Z Lw Z

Z3 4o (nJi cn

zLU

Ou O w

32

trend is the growth in the market for ancillary products

such as third party scanning software, storage-end devices

and data indexing and retrieval software 3,17,18]. The

next section explores this trend.

III

33

5. ANCILLARY PRODUCTS

The rapid growth of the scanning market has been

accompanied by a proliferation of ancillary products. These

fall into two broad groups:

(a) Low-nd Enhancement Products

This group consists of low-cost products which are

designed to augment the capabilities of low-end scanning

systems. They interface with the scanner to perform one or

more functions of a high-end system. Third party scanner

enhancement software and low-end interfacing devices fall

into this group.

(b) Complementary Products

This group consists of products which complement and

support the capabilities of scanning systems. Storage-end

devices and software for specialized scanning applications

are examples of complementary products.

The following paragraphs briefly describe the trends

and products within each group.

5.1 Low-End Enhancement Products

These products emerged in response to the substantial

costs of high-end systems. By performing one or more

34

functions of a high-end system, they enable relatively basic

systems to emulate high-end systems at lower costs. Scanner

enhancement software and peripheral interfacing devices are

typical examples of this phenomenon, and are discussed in

the following paragraphs.

(a) Third Party Scanner Enhancement Software

These are page recognition packages like OmniPage,

Advantex, OCR ReadRight and SPOT. They perform tasks such as

creating scanning templates, specifying key fields, spelling

and grammar checking, and separating columns, graphics and

text. Ranging in price from $600 to $4000 3], these

relatively user friendly packages are compatible with most

popular scanners and support the more widely used word

processing, spreadsheet and image compression formats.

Although scanner enhancement programs enable users to

perform functions beyond the capability of low-end systems,

the claims of some vendors that these programs can replace

high-end scanners are only partially true. In comparison to

high-end systems, such programs are slower and less

versatile in the number of fonts they can recognize. They

are also less acurate, and are not very good at separating

columns, graphics and text. However, they represent a

serious alternative for many users who do not need or cannot

afford a high-end system [16]. A good example is the desktop

III

- 35

publishing industry, where user friendliness is a primary

concern. High-end system vendors have had relatively little

success in accessing this market, but third party packages

used with low-end scanners have been extremely popular.

According to International Data Corporation [18], overall

sales for these products have grown by 215% in the period

between 1986 and mid 1989.

(b) Low-End Interfacing Devices

This group of scanner ancillaries is the hardware

equivalent of scanner enhancement software. These products

are usually plug-in boards for personal computers, and they

interface with the host system and its low-end scanner to

perform high-end scanning functions. Like scanner

enhancement programs, these devices enable users to

recognize text and graphics and "read" them into word

processing and other formats. Specification of key fields

and template creation are also supported. Ranging in price

from $1000 to $5000 [16], low-end interfacing devices suffer

from the same drawbacks as their software counterparts, such

as slower speeds and limited recognition of fonts.

5.2 Com1lementary Products

This group of products complement and support the

overall capabilities f scanning systems. Storage-end

hardware and data indexing and retrieval software are

36

included in this group. In addition, there are a number of

products which combine scanning technology with technology

from other areas to perform specialized scanning

applications. One example of this trend is the advent of

neural network based programs for handwritten character

recognition 191.

(a) Storage-End Devices

Scanning systems require much more storage space than

is available with traditional magnetic storage media [28].

Using an optical disk system with several gigabytes of

storage space is one way of meeting this need. A survey by

Frost and Sullivan in 1989 [3] found that between 1986 and

1988, sales of optical disk systems grew by over 150%.

These systems consist of one or more optical disk drives

(usually Write Once Read Many or WORM drives) and sometimes

a "jukebox" to perform disk storage, retrieval, loading and

operating tasks [21]. However, these systems are very

expensive and a typical system with two disk drives and a

five platter jukebox costs between $27,000 and $33,000 [18].

In addition, one must pay for sophisticated software

packages which store, index and retrieve data on CD-ROM's.

One alternative to optical disk technology is being

marketed by Scan Optics, Inc. This is the Image EasyFile

System, which uses 8mm cassette tapes (on up to 28

37

concurrent drives) as the storage medium. Each tape has a

capacity of 2.3 gigabytes. As data on cassette tapes have

to be accessed in a serial manner, this system is much

slower than an optical disk system. However, at $ 17,000

for a 28 drive unit, Image Easyrile is considerably cheaper

than an equivalent optical disk storage system. A cassette

based system is a viable alternative to an optical disk

system in situations where the access frequency per data

point is low.

(b) Software for Specialized Applications

An interesting development in the scanning industry has

been the joint application of scanning technology and

technology from other fields to handle specialized scanning

tasks. A good example is the development of neural network

based scanning software for handwritten character

recognition. Neural network computer models are based on

models of the functioning of the human brain. They emulate

thought processes such as learning, and can be taught to

perform complex tasks [19]. The handwriting recognition

application emerged as a result of the use of neural

networks in the wider area of pattern recognition.

A number of packages are currently available, either

off-the-shelf or as part of a customized service that

configures handwriting recognition solutions to customer

38

needs. NestorWriter of Nestor Inc., Providence, Rhode

Island is an example of the former, and Inscript of

Neurogen, Inc., Brookline, Massachusetts is an example of

the latter. The main attraction of these packages is that

they can be trained to recognize a much wider range of

handwritten characters than is possible with conventional

scanning software.

Most products in this area are designed for DOS, OS/2

and Macintosh environments. Data are entered through a

scanner or a digitizing tablet, and recognition speeds range

from 1 to 3 characters per second. As in the case of

conventional scanning software, there is a tradeoff between

the accuracy rate and the rate of rejection. The accuracy

rate can be made arbitrarily high if tolerance for error is

low, and vice versa.

The main problem with these packages is that it takes a

very long time to train them. In addition, they scan at

much slower speeds than conventional systems do. Until

these problems are resolved, they are unlikely to be widely

used. However, they have excellent long term potential, and

should eventually c me nto widespread use by the scanning

industry. Short term :rospects are brightest in areas such

as scanning numeriJai -cmunts on checks and signature

verification.

39

The next section discusses broad trends in the optical image

scanning and character recognition industry.

40

6. TRENDS AND PROJECTIONS

Rapid advances in hardware and software have created

a major demand for scanning technology in diverse industries

ranging from banking to defense and from engineering to med-

icine 171. Instead of focusing on each application

separately, this section focuses on broad trends. Figure 7

presents a schematic representation of the evolution of the

capabilities of scanning technology.

6.1 Falling Costs

one of the primary reasons for the growing demand for

scanning technology is the fall in the cost of products. As

technology improves, costs will continue to decline. Figure

8 depicts this trend of falling costs in terms of the system

costs per font.

6.2 Coalescence of Text, Image and Graphics Processing

Graphical information can be stored either as simple

bitmaps or as a set of standard geometric entities. The

latter option allows for easy editing and storage in a

library of symbols. Currently available scanners do not

offer these capabilities. However, developments are taking

place in the area of Computer Aided Design (CAD) for

converting raster images of line drawings into vector

graphics files, which can be easily modified with graphic

CI

C O

C c

* -tL :c,..

I0a). .5;

Q Bu

a

0 .

2..__ ~ - -o o

,-a

as E_o ;.* x

s E.

O lo -o

_, C0 a .: z

_ DE 5

-~~1 z~~ ~ o _ :s _ " .EEc

OCI~~~~~~~ CC C ...0a 0UlIt ~ ~~~~~~~~~~~~~~ - -- '"E

O0~~~~~~~~~~~~~~~~~~~~~~~

LU

I-m

CL

C,)0

zZZ

o

-0

LU

LU

.

a 0a

@00 O

"O

iv

LUz

0

Zo

u( m

- a

LLU

- aNZ

C OO

Z -

Z

LL

co

w ca XUlCC

aCI-

0aoWa

E 0

_m _a mc

r * i

-U0

0

o o- c- Sa a* C*) 3

U

w0

rr a 1n ~ 0

- 0r-LI E ° o -

III

41

line editors. Raster-to-vector conversion systems are now

available from several companies including Houston

Instruments and Autodesk Inc. However, at present only

approximate solutions are supported. For example, all

curves are decomposed into line segments. It is expected

that ideas from the area of raster to vector conversion will

be combined with scanning technologies to enable text,

images, and line graphics to be edited and processed through

a composite package.

6.3 Integration and Networking

one of the most significant trends in the information

processing sector as a whole is the increasing integration

of disparate data processing equipment into a single

composite system (17]. This trend is reflected in the

scanning industry by the emergence of the integrated imaging

system (described in Section Three).

The trend towards integration can be observed in other

areas of the scanning industry. For example, optical

character recognition capabilities are being increasingly

integrated into facsimile equipment 29]. At the same time,

facsimile capabilities are also becoming available on

document readers such as the Datacopy 730.

42

6.4 Impact of Artificial Intelligence Techniques

Advances in Artificial Intelligence techniques in areas

related to character recognition and image analysis will

lead to faster, more accurate and versatile reading machines

[231. Semantic analysis and natural language parsing aids

for contextual analysis will improve the process of

identifying letters and words, reducing the error rates, the

reject rates and the need for operator intervention [24].

Eventually, reading machines will be able to handle

virtually all printed material [18,271.

6.5 Use of Secial Purpose Hardware

A highly accurate omnifont reader requires

sophisticated algorithms for vectorization and

classification, separation of merged characters, contextual

analysis, and syntactical analysis. These algorithms can

benefit from microprogramming and special purpose hardware

which is optimized for specific applications. Sophisticated

readers such as the Calera CDP 9000 and Kurzweil 5100 use

special purpose hardware.

6.6 Limiting Factors

Although it appears that ongoing research will enhance

the capabilities of scanning systems, several factors

continue to impede a wider acceptance of the technology.

These are as follows:

III

43

(a) The accuracy provided by current systems is still

inadequate for several applications. The editing of errors

and the presence of unrecognized characters continue to be

major bottlenecks. The time taken by the operator to

overcome these bottlenecks limits the throughput of the

overall system.

(b) Wide acceptance of the new technology will occur only

after it becomes possible to automatically handle complex

documents containing both text and graphics [13,17,20]. The

need to hire and train expert operators continues to inhibit

the widespread adoption of scanning systems.

(c) Broken strings and touching characters in a document

still constitute a major hurdle for even the most

sophisticated machines. Scanning systems remain highly

sensitive to variables such as paper quality and thickness,

clarity and font. Major developments in syntactical and

semantic analysis are needed before reading machines realize

their full potential.

The major trends and projections are summarized in Table 1.

g111111111'S~~~~i

I I I1 f

I is

t1aiii

Qe*-

IIi,

aIIEj 'iLX it II

If

'11

III

44

7. A TAXONOMY FOR PERFORMANCE EVALUATION

The conventional criteria for evaluation of a reader

are speed and accuracy. However, by expressing the speed

simply in characters per second and the accuracy in a single

figure, one is overlooking the fact that the performance of

a scanning system is heavily dependent on the

characteristics of the input. For example, it takes longer

to process a document with multiple fonts and multiple

character sizes than it does to process a monofont, monosize

document [9]. The presence of graphics and the formatting

of the document also affects the reading speed. Further,

the quality of the print is a major factor affecting the

accuracy of the system; broken and touching characters, low

contrast, and skewed text result in high error rates and

reject rates as well as in a significant reduction in the

speed of reading. The speed and the error rate presented in

the technical documentation supplied by the vendors consider

the characteristics in only one case - usually the perfect

one.

In order to assess the capabilities of a system, one

must consider not only the scanning speed but also factors

such as time spent in editing and correcting, time spent in

training the operators and time spent in training the system

itself where applicable [13]. In the case of a document

45

containing several graphs and multiple columns, the time

spent in editing the document can be significantly greater

than the scanning time [12,13]. Consequently, it is

difficult to obtain an accurate estimate of the overall

speed by simply observing the elapsed time for the scan

operation.

With rapid evolution in the technologies used in the

new generation of products and their broad functionality, it

becomes necessary to use a larger repertoire of evaluation

criteria that gives appropriate weightage to the major

factors that determine the efficiency of the scanning

process, such as: complexity of document; quality of

document; recognition technology used; and man - machine

interface. Unfortunately, there is no framework that

currently meets this need. In order to fill this void, a

new taxonomy for the evaluation of scanning systems is

developed in this section.

7.1 Document Complexity

The above facts highlight the need for a measure of

document complexity that takes into account the diversity of

fonts utilized, the sze of characters, and the proportion

of images in the docment. Since no measure exists,

a five-part document :mplexity classification scheme is

proposed to facilitate systematic analysis. This

III

46

classification groups documents into the following classes

(in increasing order of complexity):

(a) Class 1: Basic Text - Only Documents: All material is

in a single font, with a single pitch and uniform spacing.

An example of this class is a typewritten document, as shown

in Figure 9 (a).

(b) Class 2: Single Column Documents with Multiple Fonts

and Mixed Sacing: This covers text-only documents with

proportional spacing, as well as typeset and laser printed

documents with multiple formats (such as bold or

hyphenated). A sample Class 2 document is shown in Figure

9 (b).

(c) Class 3: Single Column Documents with Segregated Text

and Images: Such documents contain all material in a single

column format. The text is justified or hyphenated and

there are some images. These images can be easily separated

from the text (separate zones for text and images), as in

the case of the example in Figure 9 (c).

(d) Class 4: Multicolumn Documents: Such documents contain

two or more columns on a page. Although they have mostly

text, there are some images and tabular material. A printed

page from a newspaper will fall under this category. A

Dr. A. Gupta

- Figure references are not consistent (some with, some withoutparentheses).

- The references quoted on page 6 are nowhere indicated in the text,except for the * on page 5, line 14.

- The figures are of poor quality. Figs. 2 and 3 are unnecessarilycomplicated, Fig. 4 lacks one legend in.the ordinate.

All the above could be corrected by a complete rewrite. The idea isinteresting, having a minimum configuration microprocessor displaycurrent, power and maximum demand by a TLU method on current valuesobtained safely by a non-invasive current transformer. I would stillquestion the following:- Only current is measured and the data that are displayed are basedupon, previous, actual current and power measurements, stored in thetable. As power is still a function of current and voltage (and phaseangle if the load is not purely resistive), the results will be in errorif the line voltage changes. In addition, unless it is the intention touse this only on one particular load, the type of load will influence ^the results, making this system not quite general purpose. -- No mention is made of the actual hardware used. From the referencesI presume it is an Intel machine. The type of ADC (speed, number ofbits) and the method of display (printer, 7-segment displays) are also-not mentioned.- Page 3: "for every possible digital value, the current and power aredetermined and stored ... ". In what format? The description on page 4and the flowcharts of Figs. 2 and 3, leave much to the imagination, unlessthe reader knows exactly the configuration of their microcomputer. Thereferences to the "status", "addresses" and "data" field suggest to methat some type of single board computer, that has these fields as displays,was used. Can it be assumed that all readers are familiar with this?- If I follow the flowchart in Fig. 3 (maximum demand display) correctly,when a subsequent current value is determined to be less than the previousone, a search through the complete table is made, using the same, old,input data before displaying the, same, results. I don't know how muchmemory is included in their system, but surely the previous value could bestored somewhere to avoid this unneeded search.- Similarly, if the next value is determined to be higher than the previousone, again a complete table search is made instead of starting at theaddress of the last (lower) value.- In the three phase measurement system, a simple two comparison scheduleshould be added to indicate the highest load, rather than leaving this upto the human operator.

SAMPLE DOCUMENT - CLASS 1

-2- May/86

FIGURE-9 (a):

III

HIGHLIGHTS OFKNOWLEDGE-BASED INTEGRATED INFORMATION SYSTEMS ENGINEERING

AMAR GUPTA AND STUART E MADNICK

Large organizations must necessarily rely on multiple computer systems for a number of reasons, such asthe increasing size of the organization and the growing reliance on computerized data. In virtually allcases, dissimilar and incompatible hardware and software systems are operating on a concurrent basis.While these systems may meet the objectives for which each was designed, their heterogeneity presents amajor obstacle to ready access and assimilation of the information they contain.

The objective of Integrated Information Systems, or Composite Information Systems (CIS), is to mitigatethe problem described above. Such systems must be geared to span applications, functional areas,organizational boundaries, and geographic separations in order to present a unified picture to the user.While designing such systems, it is necessary to look at a number of inter-related strategic, technical, andorganizational issues. The goal of the Knowledge-Based Integrated Information Systems Engineering(KBIISE) effort, highlighted in this report, is to survey the state-of-the-art of methodologies foraddressing these needs and identify areas needing increased research focus.

Strategic issues include motivating cooperation between multiple organizations, each with its own glispriorities, and security needs. One critical success factor for such cooperation is participant consensus onthe issue of access to each others' technical and non-technical information. There is an urgent need toclearly define the domains of shared information, the potential benefit to each group that participates,and the role and the responsibility of each constituent.

Under technical issues, the evolution of distributed heterogeneous information systems is studied. Beinginherently more complex than conventional databases, such systems require powerful semantics andupdate capabilities, as well as sophisticated concurrency control and recovery mechanisms. Both thesemantics and the syntax of individual queries and updates must be mapped across systems. In additionto physical connectivity issues, it becomes essential to develop new techniques for incorporating logicalconnectivity across systems. Such techniques combine ideas from the fields of database technology,communication technology and expert systems technology.

Organizational issues cover the process of making controlled changes in complex organizational environ-ments. The prevailing theory of inter-organizational networks explains the multiple forces that modulatebehavior of individuals, groups, and organizations. There is a need to develop focused standards to serveas the foundations for neutral representation as well as for the development of more elaborate standards.

The problem of distributed heterogeneous inform; tion systems occurs in many disciplines ranging frommanufacturing to banking and from maintenance to logistics. A number of government organizationsincluding NASA, NBS, and different agencies of the Department of Defense are faced with this problem.Within the U.S. Air Force, several efforts are directed at finding solutions in this area. In order to attainsuccess, it is necessary that government, industry, and academia work together to develop commonstandards and to fill major voids that exist. A plan for concerted action is developed in this report.

SAMPLE DOCUMENT - CLASS 2FI~GURE 9 (b)-:

ATTACHMNT C

Materials for IES AdCom eeting (May2 4 ,1985,Toronto) M aterial--A

Suggestion For IE Society Enhancement

Objective

To promote enhancement activities for the purposes of increasing IES meter-ship and improving and strengthening IES functions, image, and effectiveness.

Overall Plan…aw i.. .1 .1. .. _ ;__ .. … ;..: _.._ 1 ….....

t1e overual plan o nnance our soc3le;y InVOIveaS

addressing four areas as follovs:

1. Method of Promotion and Scope

2. Goal3. Enhancement Activities4. Assessment of IES (status of IES)

Each of the above areas will be addressedin the following sections.

1.Method of Promotion and Scope

The enhancement activities are aimed at invigorating

our society. These activities encompass all aspects-.A ! 1 - I 1 I- -1…,. L - - -… - - - &E , -- - - __

wnlcn are lkely Co improve cne managemenc ezfciency

of the IES AdCom and they extend over a wide range.

To promote enhancement activities,

there are 3 promotional methods to be used singly

or in combination, depending on the nature of what we are doing;

(1) Top-down managerial activities from IES AdCom

(2) Technical Committee activities which extend horizontally

(3) Section/Chapter activities which have a bottom-up effect

§ Society activities with team work and cooperation; see Fig.l-1

SAMPLE DOCUMENT - CLASS 3

IES EMBERSHIP

8,000

By 1990

The IES's GOAL is a member-ship of 8,000 by 1000.

This barometer will trackprogress toward that goal.

...

i

.

FIGURB A (c):

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. IE-31, NO. 12, MAY 1964

TABLE VIMICROPROCESSOR OPERATING SYSTEM SUPPORT BNCHMARKS--STACK

EXEA!SER [81

Ta N d Pu pb Tia p TAvim- tablu h (ds(c) hWE$& (e r (ddU

I I0 2 Il 13 4 =0 434 53362 laO I 2715 1 2U O 3332 79383 200 2 343 3440 44 4236 13 10320

Total 12427 971 23S9

(1)+(2)+(3) 3Ptrfenanow nuum Ho 1.00 ,. 3.4

TABLE VIGENERAL CHARACTERI OF 32-Btr MICROPROC!MO ([1

MLOW-.h Mt ago CM

TICnoLO I.as t. sm.r, U nS1OMWO CMOs

O. OP1F uASON 4.lIIS 4.3 219.0.0 O ISIZ OF C 10.003 MII 4.4 n MAI .rm ,l$ L

i05l 0.SIPAI T ./@ WATT A I IMS 4 WATo 2.5 IAVAoIWPICOMIT U ACTIM 83 o4 P

,4 TOTALW4ASI C C X o r1INC 1II l II m II

1O00. Qf IlOIt t4 , 6r A zSM LL l"lmuA IN vlSIIM VW

GIOINSf V156 GEIP u.IAI0o TO vUs"h0 0F 0 16 2 2

NSITnCTO4Smo OF ACOOMs S i1 10 S

'Curnteuy for iteral use only.

TABLE VMIEXECuTm N TM S (21,1 [61, [171

inMAC14 LANGUAi4 IWN TIM( I ILLISICOIOSIZI S*nCN S4EvlE PLZI. ACs

C 32 14 20 9400 4l~.11t710 PASCAL tUNIX) 32 I 4 220 11 w 70

PBSC (VMS1 32 1 4 20 1 .53

C 32 4 7 ?4 37.100 710I6 Is tMM ) PASCAL 14 5 3 10 32.47 11 0

PASCAL I2 50 960 32.320 12.323

61000 itS 6MI) PASCi 5 3 Itg 910 27t0PASCHAL 32 1 5 244 9200 200O

80*60 C MWl PASCAL 14 7 3 7?4 44 000 111

'32'4EL 3 II Mtl ) DA 1 6 * "s 4700 t 47 00

502346 NltM , PASCAL IS 1 at 91138 2213

02"r IC UNlt) PASCAL 1 11 t35 7311 1774

NP 32-' CPU, PASCAL 32 Na "A 7 259(Is MNHt

NIs I1Q3 (?7 mtNl SCA. 32 NA NA 24 "ot

'Indicates expcrimento prolowpe."Indicates vendor-pro'ided nfortiatio.

be monitored. Consider, for example, the machine tool appli-cation shown in Fig. 1. There are several independent variablesto be controlled and monitored as follows:

a) rotation speed of parts,b) lateral movement of parts,c) lateral position and angle of each of the cutting tools.

To increase productivity, itmust operate concurrently;tool interfere with anothervariables can be monitoredbut the overall concurrency

is essential that multiple toolsalso, at no instant should anytool. Each of the independentby a separate microprocessor,

of tool operation will be deter-

mined by the ability to maintain noninterfering profiles. Thelatter aspect is dependent on the communication bandwidthbetween the computing elements. The overall performanceof such multiprocessor configurations is affected by thenumber of processors, the communication mechanism amongthe computing resources, the characteristics of the computa-tional workload, and the control program. Whereas the majorconstraint in single processor systems is the speed of theprocessor itself, the major constraint in multiprocessor systemsis usually the speed of the interconnection mechanism usedfor communicating between the computing elements.

Fig. 2 shows an example of a multiple microprocessor-

SAMPLZ DOCUIMENT - CLASS 4

112

GUPTA AND TOONG: MICROCOMPUTERS IN INDUSTRIAL CONTROL APPLICATIONS

Fi. 1. A typical machs tool pplicaoa with mudi_ too,.

LOGIC OUTPUt TOCaT ON PANEL

TO MACHINE

Fig. 2. Computer system configuration for machine tool application.

based multi-axis machine tool control that ses a sharedcommon data bus. Machine tool throughput s imited not somuch by the number of microprocessors attachable to theshared bus, but more by the utilization under load of theshared data bus and data memory. The memory utilizationfor two typical functions executed under the multi-axismachine tool control environment is summarized in Table IX.The total utilization figures have been broken up by principal

microprocessor tasks. In each case, the program code residentin each microprocessor subsystem has been optimized to makeminimum use of shared memory. For nonoptimized programsin which flags and status registers are maintained in sharedmemory, frequent accesses to shared memory are essential.This often has the undesirable result of driving up memoryand bus utilizations to 70-90 percent. At this level, queuingcontention will have a severe adverse impact upon system

SAMPLE DOCUMENT - CLASS 5P'IGURS.9- ():

113

III

47

sample Class 4 document is shown in Figure 9 (d).

(e) Class 5: Integrated Documents: Such documents contain

both text and images. A typical document of this class

contains multiple columns, with several charts or

illustrations within each column, as shown in the example in

Figure 9 (e).

The above five - tier complexity classification scheme is

utilized to evaluate a broad range of products in Section

Six. The five sample documents shown in Figure 9 were

selected after extensive study; they are used for the

purposes of evaluating products and validating the taxonomy.

7.2 Document Ouality

Since the performance of a scanner is highly dependent

on the quality of the input documents, it became necessary

to carefully control the variations in the quality of the

documents used in the benchmark tests. Although it is

difficult to establish rigorous measures, documents can be

broadly grouped into three classes, based on their quality:

(a) Low noise documents: This category comprises original

typewritten and typeset documents, with normal leading and

clearly separated characters. Skewing is absent or

negligible. In addition, these documents have no

48

hyphenation or kerning.

(b) Medium noise documents: This category comprises

original laser printed documents or high quality dot matrix

printed documents, as well as good photocopies of such

documents. The contrast is good and skewing is low (under

2%). Further, the characters do not touch each other.

There may, however, be some instances of kerning,

hyphenation, and uneven leading.

(c) High noise documents: This category comprises second or

later generation photocopies with broken segments of text

and characters touching each other. Usually, there is low

contrast and skewed text.

The above three - tier scale represents one measure for

defining quality. Strictly speaking, quality is a

continuous variable with multiple dimensions (e.g., quality

of characters, quality of background, contrast ratio and

amount of skewing). Consequently, this scale represents a

first level quantization of this variable.

7.3 Recognition Techniques

Recognition technique is a qualitative variable that

tries to capture the sophistication of the technique used in

the recognition process. Various recognition techniques

__�11___1______________________

III

49

such as matrix-matching and feature extraction offer

different capabilities for reading. The implications of

using different recognition techniques were examined in

Section Two.

7.4 Man - Machine Interface

In order to minimize the total cost of scanning and

editing documents, one important factor to consider is the

interface with the reading machine. A higher-speed system

that requires special skills and training of dedicated

operators may, at times, be less desirable than a lower-

speed system with a very user friendly interface. Two of

the evaluation criteria that represent this variable are

trainability and document handling. In addition, it is also

important to examine the interface between the reader and

other computational equipment.

The concepts outlined in subsections 7.1 and 7.2 above are

independent of the equipment utilized. Also, they do not

require an understanding of technical terms. Document

complexity and document quality are therefore good variables

for an evaluation exercise. Such an evaluation is

documented in the next section.

50

8. EVALUATION OF PRODUCTS

The number of scanners available today runs into the

hundreds, and it is very difficult to generate an accurate

and exhaustive comparison; even if it could be generated, it

would soon become obsolete. Consequently, in this paper

only a few representative products have been analyzed. The

list is as follows:

Datacopy 730

Caere Corp OCReader 500 Series

Scan Optics EasyReader 1720

DestScanners PCScan 1000/2000

Sharp JX-450 Color Scanner

Truvel Truscan TZ-3BWC

Kurzweil Discover Model 30

Calera CDP 9000

Scan Optics ReliaReader

Kurzweil 5100

Low-end scanner

Hand-held scanner

Low-end scanner

Low-end scanner

Low-end scanner

Low-end scanner

Low-end scanner

High-end scanner

High-end scanner

High-end scanner

8.1 Results of the Evaluation

The results of Phe tests on these products are

presented in Figures 1-12. Figure 10 depicts the relative

performance of the systems in terms of a document complexity

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)

(j)

III

0 Ao

0*

o0 *

z;

LU X ~ ~ t CL~·C

LL 0~~~0 y~~~~~~ UF U

0

I- -

C 0a: * ~~~0 gI.->. - z0

U

00 -

u~~~~~x ~

COI X

a: t _

<:~~~~~~ Z * 0O *t 0C~~ ~~~~~~~~~~~~~~~~ Co *

- 2wV~~~~~~~z~0 * 2 LLI~~~~~~~ 0*

0c 000 ~ ~ ~ r -

LUI L I I_ zE o

u >E z "e z

=0 · o0 0 -E ,

L

.2

0

0E

- o

0ao

0.

C 0

V'" E1*

0

a,

51

versus document quality graph. In Figure 11, relative

performance is measured by graphing document complexity

against scanning speed. Finally, Figure 12 maps the

performance of the systems on a column and graphics -

handling matrix. A more comprehensive comparison appears in

the Appendix.

At the low end, all the scanners were able to read

Class 1 and Class 2 documents without much difficulty.

Class 3 documents were read most successfully by the Scan

Optics EasyReader and the Kurzweil Discover Model 30. None

of the low-end scanners was very successful with Class 4 and

Class 5 documents.

All the high-end scanners were able to handle Class 1,

2 and 3 documents. The Scan Optics ReliaReader was not as

accurate as the Calera CDP 9000 and the Kurzweil 5100 in the

case of the Class 4 document. The Class 5 document, with no

clear separation between text and graphics, was handled most

effectively by the Calera CDP 9000 system. In the case of

such complex documents, the process of editing and

reconstituting the original format is time consuming,

exceeding ten minutes for the sample used in the benchmark

study.

Characters that touch each other and those with broken

01IZ z :

Z :_ _CE iAS... ,,

O

OCLC a 0

0 F ' o anow0 cJu y r ; 0 Q

zo °L Uo

Z * (3O U.~0 -

00 0~~~~~~~~~~~

1 I

LU . -

eO c

U.

L

to

a

U

O

C0

6 _

oQ 0U

0

Ea

C4 0

0

a

0IQ

e ·

a o ca

0 . ' _ Ca > 0 O

000

o.m00

4)

N 0

Ul~Sd Sd

x

aI:

z-J

Zz:

o(00.

-j00J

z

o

a.I-

LI:

0

U.w

I--.i

a

cJ

Lu

U-wlP=

mr

o0

O --

0 _ .c.OS

!Qa

S0

,od

A0z

c a

m

-m

. o

=

(0 '

00

t-

00

I-

a0

0 Car

am=c

o_ Q

> a ca a a

O S .a

0 .- -

= C I l l-S aC

52

strokes were the major sources of error for all the systems

tested. The complexity of the document being scanned

severely impacted the scanning speed and the accuracy. The

accuracy of all the scanners was found to be highly

sensitive to the quality of the document. Even a

typewritten document caused a significant number of errors

in cases where the quality of the documents was low. A

single handwritten mark or even a speck of dirt is a

potential source of error for the reading mechanisms

employed in most systems.

53

9. CONCLUSION

The image scanning and character recognition industry

is undergoing a period of rapid growth in which different

technologies and concepts are being amalgamated. The

traditional frameworks for classifying various approaches in

this discipline have lost their relevance. In this paper, a

new taxonomy was proposed. This taxonomy is based on the

characteristics of the input, rather than on the application

or the price of the product. Based on the quality and

complexity of the input material, it is feasible to predict

when off-the-shelf technology will attain the maturity

threshold needed to motivate the adoption of these automated

technologies in a particular industry.

III

REFERENCES

[1] All About Optical Readers, Datapro Reports, Datapro ResearchCorporation, Delran, NJ, 1989.

[21 Ullman, J. R., "Advances in Character Recognition", inApplications of Pattern Recognition, Ed. K. S. Fu, C. R. C.Press, Inc., Boca Raton Florida, 1982, p.23 0.

(31 Market Report on Imaging, Frost & Sullivan, Inc., 1989.

(4] Shurman, J., "Reading Machines", Proceedinas of theInternational Conference on Pattern Recognition, 1982, pp. 1031-1044.

[51 Lashas, A., et. al., "Optical Character Recognition Based:onAnalog Preprocessing and Automatic Feature Extraction", ComputerVision. Graphics and Image Processing, Vol. 32, May 1985, pp.191-207.

([6 Baird, H., Simon, K., and Pavlidas, T., "Components of anOmnifont Reader", International Symposium on Pattern Recognition,John Wiley & Sons, 1986, pp. 471-493.

(7] Masuda, I., et al., "Approach to a Smart Document Reader",Conference on Computer Vision and Pattern Recognition, McGrawHill, 1985, pp. 550-557.

[8] Kida, H., et al., "Document Recognition Systems for OfficeAutomation", Proceedings of the International Symposium onPattern Recognition, John Wiley & Sons, 1986, pp. 343-362.

[9] Hou, H. S., Digital Document Processing, John Wiley & Sons,1983.

[10] Suen, C. Y., Berthod, M. and Mori, S., "AutomaticRecognition of Handprinted Characters - the State of the Art",Proceedings of the IEEE, Vol. 68, No. 4, April 1987, pp. 469-487.

[11] Harris, B., "Characteristics of Imaging Systems", Byte,March 1988, pp. 76-87.

(121 Mason, J., "Error Rates of Imaging Systems", Imagina,February 1987, pp. 42-46.

(13] Johnson, R., "What High-End Scanners Offer You", Computerworld, April 1989, pp. 17-29.

(141 Tanaka, Eiichi, Kohasigushi, Takahiro and Kunihoka,Shiamura, "High Speed String Correction for OCR", Proceedings of

the International Conference on Pattern Recognition, John Wiley &

Sons, 1986, pp. 340-343.

[15] Wilson, P., "Error Correction in Sophisticated Scanners",

Imaging, May 1988, pp. 78-89.

(161 "Survey of Scanning Products", PCWorld, May 1989, pp. 43-49.

[17] "Networked Systems in the Automated Office", Business Week,

August 1989, pp. 45-56.

[18] Report on Scanners, International Data Corporation,Framingham, Ma., 1989.

[19] Buffa, Michael, G., "Neural Network technology Comes to

Imaging", Advanced Imaging, November/December 1988, pp. 42-50.

[20] Kapoor, Ajit, "Electronic Imaging and InformationProcessing", The Office, December 1988, p. 56.

[211 Dollard, Tony and Odorisio, Linda, "State of the Art - State

of the Industry: A Review of Commercially Available Optical Disk-

based Engineering Data Systems", Proceedings of the ATI 87

Conference, Monterey, California, February 1987, pp.4 56 -4 7 8.

[22] K. S. Fu, Syntactic Pattern Recognition and Applications,Prentice - Hall, Englewood Cliffs, NJ, 1982.

[23] Nandhakumar, N. and Aggarwal, J., "The ArtificialIntelligence Approach to Pattern recognition - A Perspective andOverview", Pattern Recognition, Vol. 18, No. 6, 1985, pp 383-389.

[24] Rich, E., Artificial Intelligence, McGraw - Hill, New York,1983.

III

(251 Mantas, J., "An Overview of Character RecognitionMethodologies", Pattern Recognition, Vol. 19, No. 6, 1986, pp.122-134.

[26] Ward, Jean Renard and Blesser, Barry, "InteractiveRecognition of Handprinted Characters for Computer Input" IEEEComputer Graphics and Applications, Vol. 23, September 1985, pp.231-240.

[27] Berman, Ari P., "Image Processing With Embedded ExpertSystems", Advanced Imaging, No. 15, November/December 1988, pp.51-60.

[28] Jones, Tony and Safdie, Elias, "Scanner Selection - MoreImportant Than You Think", Proceedings of the InternationalElectronic Imaging Exposition and Conference, Prentice - Hall,1987, pp. 341-351.

(29] Brobst, Paul L., "Fax Networking", Proceedings of theInternational Electronic Imaging ExPosition and Conference,Prentice - Hall, 1987, pp. 512-522.

[301 Stern, Randall, "From Intelligent Character Recognition toIntelligent Document Processing", Proceedings of theInternational Electronic Imaging Exoosition and Conference,Prentice - Hall, 1987, pp. 236-245.

APPBNDIZ

CHARACTERISTICS AND CAPABILITIES OF THE

PRODUCTS EVALUATED IN THE STUDY

III

(U1D.,.4-4

>4

u

O '

wU.4.1a4M

m

4)

(U

Xw34

m 01m Cre *

Ca V

o u0 w

(1,),

.fl.4 UN

4)24

4) Q £.j.)4U0.p4 4 .

4-) C

- a'

LP £ Op4 CU U,- 1 .i (

C Uo QJ

c4.) .

m. 4_ Q U I.1-s 0) I m

CU E -4 Q) Cc s4 cn M: M - N=

0z

I .1

.4 .,4 ,' 1443 0

24 arlu- 1) . - 02

0 4J0] , 04) -.F4 C

4. O

0 . > 4E, a

0z

I .,0 54 U

0 0> N t_

"· C ,,

. L.e : U M ,t - ) .-

_13 A 14 (U . M a Q1W Q aQ4 > ) E-4 4C 0 - 3u

>44.4

U04

0 m3

ea

0 0

z.40

4)IQ1.4

a02

(U)

-4)

. 4 )

z 1: t

I

4 000

mC. )C4 M

-4CW cU

-,(U

.. i

o e

.94.,4

o I

" 4· ,'

41I

0202.0

4.)1.4

024

1.4ad

1

SJm(0

C W

Q4)

P Ow SU

.,4 · .. :

c

c02

4J24

4-4

II C

1 4 .C .-s 4)

U -4 C

m0 0

C O0

l .4 C ., -

· O :

r4

o

a

Ca

00.4Lnw

4)313N1.4

0

000

02

1.441)00

(Ucn

4)a

z0

E-4

;3

Wr~

w'-4E.44:

I

U)

H~zw

E43J04I-sW'-4

c:

U

C)E4

4:

Um

134

u

u

onC.)

0

zU

0wtul

oVCa

0

o

14oU

C

024)

04

In

C.,-t

(lcn

a4

CN

eaQQ4)

a (nzz n w -n .C'

U H0 C

u0

0 W ua0; w

O

-.a4

4).0

OO(UN0

1,I

OO

Uare

a

4

,-09oa>1

;3I

c(U..4

·(U

O

O3aUa

0O

cJcUo

,.4)a

a43o

02

oaw4)

4-4)a

4)3a

Uo

Urn

3

4-4

u41)

4)

OwZ u

02.0

-,1X

0) CC -i, _ =

.e =Z

't . -4

FO00m

O

4)UW4.1.11.43W04>4E-4

3O

4.1

4)

4)

r

E

cI _O

cX 1I _ ~

u-- : .,- -":

E -J CL- C I-

- I - =C" I :

02

0

0

CU

02azI

0Z

z W

$UCC0

0 2; V Qu 4)

W4

= l I l

T l . - --- --

------ l l .

j k ' j j I m I(U

0 0 uV.I-co -

00 ,.~-4 " 41 ,.L4 '-4 a L n 41 31 j4 44 -

4 u r C C- Q4 .) n

X ( U (U U U 4 (1 0 0 0 ' 4;. 0

r, II 4 Us r X 4 *I: I Y t z a0 O V E-4 4 O o Q t

4 0 4., e x 4-) O J xO S -~e O V ( O 1 Q O x Col $ .. m 0 . C0 U 0 OX1 4 0 0 :32 ~ U; Q 44 >4 E-4 Q 4 .

41

I v OL ' c 3.-P4 1 S4 U M Q

W * oc o aQ a "

u cS O C D .- 4 0o ; -H u v C m si a oo (

o .u . e, 0 v f i 0 ra 04) Q c Q0 X X C 4 O Q4 :· ~ 4 Q1 (U (U r- 41-N >O o 4 ' -0U - ) 4f., C O - x u X .* " c o _ 4 V 4 0 n co

3 . *V E- : C 051 QO4 24

L O 4]UL3a OQ O O . ...Z U U; 0 4 > i o 4 41 X Z4 Q. c X (.

CO U 44 4 S W 4

41 O 4 L a tieU S.. U 4- 1 .j H

a . 3 a i 1 Q1 4 :

Q0 O -X 4 .3 c3U ·- ~ (0 ( a U :- -

C UI 4 o X0 UO 2 , U 0

U *H0 z: C v 5 ' X O Ia 0 Z O

U X U0 'Q..- u4 X ( U* x

c r0 N m o b ______ >4 E o m , \

n~~0 I IH !1 3 a I Wi m 00; rl i Qf 1 3 ~4 I; I

: il * z _ c I 3 41 ) i U -4 X I Q IW ii n Xt <* O 3 c 0 > I 3 i 3 0 U 3 0 NiE- (U| m0 " 5 4 6 3 3 U 0 4 1 i U i > 0a ( .

[ >I U Q- xU tD ~ W Ji;|XXEv | V ~ COc |u = OCi|

W

E-I

4U)

U)

0WcnI,

w1%CZon

00m0II

I4U

11mu~Um04

E-4,cwu

40:4-

u

I -,H

0

.4 43 = 43 ) 0la c w

x 2C-r 4 O 0s Qs

43 0 $4Ec Q4 0E C E P

i I 4

2 04

a C'N

·-i i 4.1

., a3

~ . O m

O 4 C

0-- a : 3

'040

2044

CO4.C0

44

43o>4

43

4-

cn

eU

e1 .,.-H

' C 0

-H O 0

I

Q) $.

04 0 .

I

U w

.,

.r4

.,-.IQ L4O4 f3-

04 3

(U

u

L0

0O0

(u

C -3 *,4

O

L4O I

0

>4-)-cH

L4 44C

>LI 3

o Q 00

O - VO

z l I

H 0- , .C:

OE4;(U C O

EC4 Oc 4(O~Cz OE20M

111

C0

O4-440

04.

4

C

r.l0.-4-4

C000043$4

O

N Q

-H >

O Q003

431..

raI-,tE-44CO04Zw

CA04

W

04

OwW3d0cnH

E-4H

4HMI4

4

4ciz

EnC-)H

CO;E-4

MHC-

4

W4u.

$44)

PC

(lo4)

-H43

CZ0cn

o0

9-H

ill

31N

ooo(4

04

(O

$4

43CCrU

0CO43m

(v

HW4

E-H

U

1

I

3a

04

U

E-r

w03

I0MO

14

n430.43

oa40uCo

I$4

3

3

.0wu

-H4

'U,)0rq

-H0M

43

$4*,

4i$404

4.)

.-H$4

-W-3s

COw

O a0$4 -H04)

*,i uam

EO ciC o

:044 )C

0

04-i

N 4-P4 Q(

COO wC Ca 0

0 4-4

0

43$41

m3:4-440uO

U=0Oz0$4

04-40

-,. 4-i

04 3

- (

0- 0cu

CN

I

cq

ON .H1

C) MI jH -H

C-;O~..

I

0$4

..)

O 40OEOU.

C40

CO

0 $4EO .MI C

I

o

r-1

U)C

u >4

a4 ~3 4--H - CO

O O

.$4 .qo O >4iH 04 4.

U (l

.-4

V 4

0 O

U O

Q .

OCO43

- 0Qr. z

·- U4.) 0

43 Cmv cM 44 IC

4 0II m w

-4-)m ea W

43$4Z E; 4Q400

Z b

O

0U a

- H

COCo -Ho

CO

'UOa0$4o C 44

43.0

Z(U

C

.,D>1-H-H

..H(U$4

I

04-4·- HC

0O

-H O- -H

O -H

O Cz 04 U

0z

CJ,4

OC0

.,t--4O

3 m

I 444 - o43 04(n 0

$40

cM

4-n

rO

-a u

.,-0

O C

4. to

4$4

344

C

_ 0

0' C

>14=).-H

,i C -Q

o

uCO.F4,-4

0.)H

(U

0--0

M,

I XSm

Q 4 4s 1-H

H 4 -H

$4 I4H C(1) H-

V LJ ssu < lu U - a)re cr O -li 4 0 O(t m $4 0 0o-u~

l I- I w --l

| __ I l l -

l . l-

It

V I41 u

4P I

0 0oO O

0 1 0

> 04

I - e 4 4 i

0 0

·n

4) 02 44 ro-4.4 u

4) 0 4 14I O m si Lc )U4;

I '1-2 Il l ) . a04- .0 (U 0 I

.a - > C IT

0 2 0(U C I U 1 O1V I

) 0 102

-4 L C a 0 = a t 1e m 4

0

4)

I 3

a0 3 '' CEu .,. i <O = 1 i

= L)

= ~~ 3 w m (Ot i¢Ei I = .

H

E-Ez

04

C3Vcu

HH

E4

V:4c

rd<

u

E-4U

a0

E.Z

0

CdCa

E-4

0 a

a.4O O

L)4

4401HE-4

14

4-43 UIX O

E-40dc i

I .IO C

U *- 4

E. L

a _I a

l m e

Iu a7 Ir .r;(

n EH Oi _ Mr

. O

Z L

04m O

14

C4I

QCZ

0

03

'-4

¢43N

O

00N

4

'U

9:C,

u

4U

zHH

<

= E-4:D "> => w

i 4'UQk

O

0 a; a

~4)ON- *4 02

E-~4)

C O u" U cH

4)uz

m f

cm

la

0

05-1.44)s ccNI- C)

n m <

HU ICu 4 4-J)4 1.

4).04.1

C. 4.1 X041

H c)I

4)) -

-i ,,

L 4: .'-O4. ,<

¥CI]

02 4)

0 4 a4- .

4 £ m

0 0

0 33n 01.41.'.04)

=4.1

a) W

4.1w 0a

3 94

2:

-W4.1

4

8~20l ;1-3 j M m

4-2.0IOS; n ZnI 'r

' az 1-5 -I.,4.1Z" " W1 C4 "

I -a

CO04441

,,- 4

C 04. ·.

04

(04n O1 a4.a

!;C)

*0C

.Q1

!-

4IC: _3 C O

3) a)

3- 0

03o

¢J01I.4S4

0o 0

C

0

r-o

ui

4

C

I

O

UH 0

-

-

.

I. t u , J

X, _

I~v _

$4~~~~~~ I I

· ,. " 0 M 0 C 4i S.0 0 O ~J·~~.'a a O 10 MUa 3> a Q .H a r a V03 ; e CT Ca c a a iJ aa

-¢ .- 04 ** O m a a a o s a

C- .- P .,-r de C . e d O Ooa a~~~~u r L r- a u 0

41 411 O) 1 44pt e ^ Q U Q,'rL 4 C4 C. OIA: ur a *H c, U Cc PC I _ a re

0 w 0 0 . e O O .o 0 a . . - ·^ a r j u e

CA z a 4 a z r ;r a C3 V 31

U W C P1a .H cJ o C

si · ~ ·I- a HU 4 -J 3

4~~~~~~~~~~~~~~~~~~~~~,o - .H Q) ·c cnI02 W- ,.Co ffi Q ol o . E

C tv 8 °

Yr ·r O a a c v_~ *H o~ a e re · 4 ~, F )., U 0 .S

k c c Q) a Il 0 o k a 0 4

M v v Q aa v>O. X a ·*H o c

Z r4 H -W u

o ( s1 4 O (

o I e c s Eou su | ; a

c e O ro a

U a or C C; crc X~~~~S _ . £ >4 (

m~~L c x _ .rl aQ ^4 (1i a% u o

L - m .H E

a X; X H - 1 ~3 aC c x - r-4 4 -J Mc

a O O CT a) c, _, a 0 C

fi X X O: Q3 £ 4 i n

or · ti Ua ·~~~~~~~~~ · 4 a M u

a Q) a a x o 3- ·c Q aZP Z 3 V W 4 - C

Zr v Z ' SE0< E· E

z c, S:; I O O~~~~~~ C ~~~~~~~w z E-4 M UX V

<: w a O Ez c) LI U

OmO~ WWLa mm3 H 0 32 W1 X1 a, m o 374: : n- > a4 m O C w> C; f 1 X (1 :>4 O > O

Ed : Mw

III

rar

E-44

Pcn

wo

DA

rnwc1-E-4HI

*4

HIgo

4

az4EnuQH1E-

ocn0

3Cn0W

PE4z

raUu

Qru0E-4wen

u

m

L4m02(U

I

0m

E2.-42

-

c00'(1ro~

L4C

V-.

4r

o I,-4r-qmr

(U)2-.4

9-0mC

1-4

02mm

C4.E

,t

-4

O

'V-4.,Q

a

'Vla

L4'V

02

-0'3' a)

tis .-4

i

a CL 'r a$4 (U-I

o

O (

'O'

02wCmu02

U 90'VX-i 4)

-4 0MX ha.,W ._ :r1: Qo~~~~~~~~C o~~~~~~~~~~~~~r r+ sw

:0

Zn

--4 IJ

. -4 .9-s

4j 91--4 0411 A

I _n I -

.z i C~~~~~~~~~

! z ,

MO C.) rZ _ ' 'n

PILI~~ ';t

. 1

4-)V

3 r4 02 -".4 -m4 m

(a a 4

w 4C0 omQj N m

4 o

m 4. 0 ad

'Vi 14 C > 3 o V'u o4 "4 oC, 0 4 'V

.4o

(1Uw a

V 2

.,.

IVO

O C

O 41

.i

o0N

"C

0

4. 4 '

O1 VO0,c. 0 u w

, = m-

Cm U0 N

ft M M - N·ta g

0Z

0Z

0Z

I

>iz N 5-

' PL a , V-- i V .4I _ ;, Q.

14 o) >C ( a 4-4 a )(

( O0 - M3! .

044

Q-4.0

W

-4

ocE-4zw

Ww

ru5.;'z0cn

m4Ecl: -4

I-s:

PC)

uCZ)z4

v

cn

u4ul

E-4u4:

4

C)

U

U

wl

I

·P4 Ia

(4P-4 n

3 bN a

0O

040

(U

(Uu)a

0

IVuu946U)

IVMU

04

0u-

Z

< C= a

z V

I N

.^

V'1,-4ol

(U

".400U

C

.I4

I.

IUC)w9

0la

I I*4I q

2'V

I U

m-Cat.O

.VmCtv0w

014C

'VI-4.XI 0

.-4v2

(UC-i5

Io

w

L4

a,I 4

44

aA

mI I u

I Mil I I aI I

4.CIm

.9.4

"-4a,4

'V4o

4-_ 'V- 02-(U-4 (C

4 0

m

-40u

F

0Z-4

4.,wel02

(U

IC0t.-4I"ru

C

ra(U

'V

o140

IU

00U

0

I

4

O 'C

0 449-coIV CV .

'V'Vj

.r41

0

0

0 ,la.~

o a,

co

C

-4143

'V

>44

O (

4-

CO

,,~ >.,.4I Z

4 0

Iz Z 4

<I S V ,,,

- .

III-

- --

- - II - - II

Sm' ' ,,,,, I

�u

---

_r W

C s

at

0O

D

.94

00laO.

04-Wa4

.94

0

00

4.3adM

uCL)

4)

4-4),

mC 30

a 3C 0

Q MI

c

3:4

a(U3

0024.3)

o\o

OLnU-,0

CD1

4)

(UW

-4.0m(a4.0z

4O0

z

41)

-(

(U0

z

0 4a ·c ! i C) O : ,^ ~ ~ ~~~~I ! V W

Z - -_

5: 3 ! i- I o 3O - ' O ·r 'r (d

v] .) 1 , : 3- 4- _4

'4 ~ .4 .v11 5 > ;.< aP O ·H _ - -> r= hr

'¢ 4 -) _ O < H

t3 C

III

E

W1-4E-42

S:

E4HH

m

PcW

U,-4W

-4rucn

HE4HA

UHwX0,uIn

E-4VE

Xw

;

:u5:

V

U,

UCDa0

X*

E-4z

lz~

C.)

0;;UZE-4WEnr

c

0.r44-0m

1.4

4.)X0

4)x3.~4-z(U4)

0

.,4J2:9434

43

1:

4c

V

3-4L4

ro

I 0ma

*1- 1I

:10

0

o

00

4-)aU

0

V)

U

U,04C)

2CZH t¢E-4 H

:C =m Wc- E-4

3 "> CZ

4.40w z

CQ q

_ CP4) p-(U

MI 4-4

.14

c

rl

I

C U

41)N

C 0

4)

O 00

O (3

4.10

aLc0I tvH aCU1-0

.140 >44U

W- l

I Qo .om 441)(Un

-o A

Po4).0

4-)(U.--ow

U

2

O40.3

4-:

4.3zatCH )

14

C 0

O(V Sa ua4

02

>4

a

>4

024)>4

4)

oUO

_ a

( C

4 VV Um

44-0co

020

aec.r4(U.4a'

:4i

4)

D,S

0E(U1.40c4.):4X

E-4

zC)u0a a

L4I.A a 0rQ: Z X

HH I4 cn am -W>4 2 t= :Z C0: =:

>

O .~ OO W-a O

Z mO O0

E- 0X s

.I I .

-

- - - - f l

a3z

J------ I

__i mnE4

4- ) Q a xO IIN 1 . O0 m 4 )

! : 0 4) X ¢) J

_ , _ II( 1 > 1 Q4 3 _. .' _j :> ) .4 0 a , a 4- C

4r0 o 1 . 4 O ( 0 't 0 0 w - >1 '0 '4 ' oQ) 4im 3 0 (U 0-. *' C

m3 11 I N>4 C 0 0.40.) 1.~4.) 4 I .1 : 4, ilN O . ; a 1 U: > P 4 0 .Z0 S l 4 , . a t 4

,'0 4)YC :3 E 01 a a 1 U Z a P O C 0 0C) C 4- 0 C , . l4.C0 a 0 0 .'- e4 - 0. 4 4 (U 04O ' I I 3 U 0 4

o (U' 0 , ' .J ::.i ( :(U N O ~ 0 0 ) 0 0 C .C c

U , , , ( I . 4- 4 ___ '-4 U V U U ZO 4>x~ ' 9 4

U : r I H x S DIo I : · o 1.4 Q4 4 0)I -4

IX~~r I Qi -_4

> La w 0 w z 0 .,, 0 o U U Z' E C C a 0.02I 4 oc_ . % 4 (U 02I Q) -Z ; E 4 O t I U X e

: n: 1 3 V = Z J a) C O I X£ O - u4 *

E4 ., I H 0 : n 4 ) -' 4-'

E I ( Q ' 0; c - V 4 U * .4O O -4 1. 4

U I( Nn UCS4)( H -o U ( UZ 3 C

__ _ .' VE-44 OA 3 U

H

4:

az

cJE4Cl0

tnI-4H

4:

Edz4U

cn

E-4

4:

U

U)uz

U0Ima0;04

E-4z

U

rd

0

Cl)

III

I . '~I

Ir~ U ~ , .0· rll~~~ Le 3 Q,~~*'> U V 0 J.

·o ~ I 0 0 C O C . · 'w k-4 ,...~ e w c: CO M,-I (U .V 0 ( 44 o.^ I O Q0 O V ( L Q ' o V -i 3· CO O 0 L. .i . ' m C QQ0 4J L

I. 0,. , r .- Q- 0 Qw o

O 41 I 4 x a C / W C O C . o

NO Q m H ffi (fi H E l ; o .~ 3 14 -4 &0(V 02 0- V4) 0 4) a d P002 0, c W r .4 > I i u4-£ L0 m r4 c k._c w= O | OQJd H m < m o m o o o a0 o X s

0 I - : Ic 0 . ,-H 0 0

a * q O I - _ XO e0 l c £ °

0 W a t..~ (U- 00.. a . 0 0 .

o . , -4. QU ~ S ., .OU 0 (U ~.J m 1 0 0 (

~c~ 0,-0(n 4C ) o Q Q 4 a ) a, " ·r a) < * 4

04 I M O 0 0 V .V04 1 r. O C O 4- O , -. _ .O I C 0 2i L QQ 0

4l H m 0 H0 . O m

04 0 V 3 0 e 02 ' 0 4 -( I U u cn4V o 4 - Po O O 44 o

4-) U -) S U 4,0 V 0.. C(U '-4H DU . 0 V 0 0

o ( I*4- I - 4e X O E-44 I O

V .4-)I Z 0 > 0o o C)0 I O0 0 , : o 0 , 4. a w 0>, ae CQ c~( m C 0 o

C : .Ha) O V

U ( E- l 4VC I .4-J S4 0 n H£ cr ; V n a tD . i 4i O I O uz u: O uz O O m 0 C

I 1 _ MIu2N la

I_ _ 04 0- u 0 _ _ ul 1_ _4 > _I Z - OL O a s WI n w O Z.CO E Z O 4I4 E - a sW 0 4i 3 3 a0 W C 3

Z< Z O 4i MI = h V 0 C O ' M n

H > I , C -'C Ul :D WI m

Eq I , v N 0 O 4 Q C 4> uz EV V ; *r

U0

04uZ

0E-

w;

C4

Uw

HE4

0

U,

4Hw

I-..

I-t

ww

0W

C4

w

WHE,,HV4

UctUaz

uE-4

HZrtj

u

T 0o a O

oil r=~~~~~~~~ ~ .rla raCJ, o C: .,4 .,.. ~ 1 ,. ~

~~~~4~(· , II · . 4 3 o

. C 4 4 w o r 0 ao .,, c c: v wmO, i,mu ;. a. , V 3 IC '

-*4 f- 4 Q$ . 4 4 V PC 0 > e 4- 4) U Vo tvH X

= 0 3 % = z C N a a L4 -W C 4i V 4-) I 404 > 4 ~ 0 . O c 0 O t_1 O II a O * H C a v a C a, O P eMOI V Z a rn U] r a a a C Y vl 3

0 .)3 ·1~ · ~~~~~~~~~~~L ,4 Om; C X V e~~~C

0 E u -a, c Q a iC C C o0 0 Q to >4o o ) u V - -

b Q C t_ ~ a U V O o Z .a ¢o C © :. . 1 ·

a . , . ~ Qq > 4.a t VX~~~~~~~~~~~~~ ~ ~~ ad 4) 4- -P4 ; o .r4

O 0 .,< L4 :4 r C -- 41 L4 1 > m Uu LI i O) 4i 0-- c) · C

Q > W 4 0 L4 J *a a w Q f X I VU Ef0

4 3 a a C c m Li 3 Z :

J u a z 0 0 V 0 F O o=

_ .z - .~b O Q 0 Cn C(Y 3 c 4) a a

4>;~~a Q X I _- I1 O Q v 0 3-' V ·-r L)I O £ O u SC3 4 C HNf > sw : : zcn n (1 u 0 0a .d X > Q , O LC i i I 3 3 r:a)ca 3 4 3 fi ·̂ 3 _r az v c av Q O o ·rl _) *H - W -

a ¢@ C) 3 O fi 4 6 U > U IV .CI a . O v- O 4 f W 04

z I4 m U E a

S t4 C (l s0 z 4 :O ) A S L4 4 (du Cd O > ;I

E-4 4 > aJ04 IV E-1 3 l n4 E-4 0 M

> :J Zx 0 4- ) oW z : 4 a o U

C:

En

ZE-4

we

O)134w=lcn

Io 0"O

C=Lc lo

4wri E-

MecOc

E

ul

v34UqXcnEdue

w

E

C.

III

U O 08 0 84 I 0 :)

E-0 0 0m I > I 0 o .Cc I ·m la a .ove I0, 0 0 0 m O ·· HI V ,1 V O V

E- Um z co z n c v E t:

02 0 2n.,~ Q2O 0 Ch~ 0 0· . I , I 0I -s 1

0 I Q 4) 0 0 . 0 E

01 ' 01 i m 4 0 0 0:::; 4) : ( C 4) ' - C .,,~o on I - , i 0 z :X I m .~ 3 . m 4) 0 O u s- 00u ct M I · uz .l , c0 Q Qcn I u 0 a c 0 m mO

u I U - 3 u 41 ,4-j 44 "a >1

Zo It o - c- x4i e x ., q

m o I ou o o, re a a)e u - Cao ^

a I cl a E-4 4

'4: I ox O k

04 4- 4 - 44 0Co cn- O 4..T 4.0 0-l< I w c -- sC I 4 0 1" a U Q 4 )

> I u 0

O O m Cf) zr o o P o O c 4 z I w e- 3 c; - E 0 ; ·0 0uz :z vC |1 C Er *- w N 0 3 0 U

I I _ I C

u accJ ii I u z z :; Uz

ZU

Vd1-454-4:

W54

Co

OH5.4

wd

Hrn-4:U

E-4cO

5.4CoHu2z

54UCcnE-4

u

u

CoEd4

0

mU

0

E4

0c4

N

O

4J0.4)

O.,4

C(UO

4)

01

4)4-

U 0;= I (S C E- n

(o .14 4 02 H -1 4- 0 m-" . > .v 4J^3t I *a _ Ei a) 'r 4) >, . . a u o I a Ac14 Ii 4 ~ c, m 43 n I _ -t C a>

cn I 0 ep 3 C Im C m al

E4 C121.4sie C _ U4 Z > H 0 4

#¢ ·Ilr a C a Crr a tCPo4O

4 (a C4 l _

OI U O U ocn t I U Z

r: r I > > >U . e CUt

a a a 0 4 4.)Z I 4 4., 4, 40 4: 4i. 02 X

0 ' . 0'' ' 0 0 di tU

c) 0 .4 OCO E- 0us 4-) ) 4u0 O C41 (- > S 4)

) C C) 4.) 4

a a 4w EZ * C Q O c 1. 4 O

= W .J O 4) 4) 4) Z(n 0 410 C CL >4 -44

H 14 02

0 O . 4-

C i n V I m4 Q- a )

M > 4 > -4 C >i 0 z m i -4 O 3>4 O m ( H Q4)0 t V. O 0 40 O * - " o3 o o 1 4H 44 "-_ U- * 4 U4-W 4 0 4- _ - a C1 3 4-E4 H 44 *-4 u .4L a a * O o ) 4Z Z Po

; H 3 P1 h h Q1 .4 C 4CU C H' s O

.4C Om a4 .4 - a 0 mi 4 U 0.)45 Z OU u .4> a a a X u u u u E4 E = 0 0 awU 3 [3 3: w a: E-4r < ad :n1 a

n

f3h1-4

PC

E-

z1nCU)

n

I-I04n

E-4ndA0C.)

cnH

14HH

C.)4I=4C.)uHr14U)

co-4

u

0~D0a

14

ra

0

En

klc4l C all z Tn' 1 ° m m 3 v I

V I ·v 1n^Ca O a f C - Om3= I,~ " .,-+ a4 ~ ~ O o ~ .- C o

cX I x ee^^4 a 0 V w u i 0mo a U I _ U P ~ O a ca 21 ._ C . ,

cx t n oc v g c us a Xa rN a 14 C a O C E I ., 0 I 0 4. ; e .,- , C

I P1 Pa a ·r m 'r a m o sVJ I m c, c, si o O WX so s ee I O H oOa o c4 ·3 C 04 o CD 0 C

is Z .~ 0s: 2: ~ W 3 = .i 4 . X -W ) 4

,.= .,. 0 -'. ra 0 0 04.4- m 41 u 3 14 z - 4 3 Q i.4 ·- I I U 0 .-I n n3 4 a : r4 - -4) -- 4 -r-4 --4 u ~ ~ ~ ~ ~ - - 3ra a) 04 m C r4 o

O E- cw W W C m C 4 O ('Oe 1 Z I 0

,, _1 E . . E 4 (*. Q) c ,V a --q C ; Q0 4- 3 i 0 Q H '

~ *< O C a o p, a U O f a s r 3 rn m a s 4 I s Q u 0 k-CS: X 4 I ` m et C S4 W $ 4 w e w F4 I Q 4 w C a o e-Ha 4o 4 tV >

CO a - A H 4 41 Ora4: O iQ 4i U ( > U 4) Li In 4 oe o I X *9 c a cn'r *H a t/ co O O Q c :z: tn 3 *· u U ^£ gX H ff P1 a w H a

W ii I I I Im a 3, I > C I S:

4) -4 O L4 C-4 r -4 4) 4ia I CT ·r P1 P1 a

>c -4 la 0 · 4v) a Pr O I -4 OaaUZ I *H E 'X ;XX I E; a a W 0 a ) X o

I r U c Q0 Q O o Q rr: I XX a a a m aC 0 q at racn 4 Wa PO CU 14 5 4 CD 11 = to s

H m w E-4 1-4 4J 4-) c, WH- | 4-4 I (1 -') x M -4 0)Ev I O ) <5 I Q Pc vt x e *HM 4uz O (UA s4 Q 3 O I m-H ; U m 4t & a 4 --4

1 a n e 3 0 w-Z 4 5s H at c, ·c a Q Uz : a)4 9 U 4-) 4-) 4 a M 0(U I U;; v z £ o m > v I0 m o m o o o o ms (U w 4

I V V] Orc XI 0 s4 iH Y O 4 a O Ef a e ;

> ; S aS 9 m 3 ( 71 w3 c

c 4 a £ X 5 £ Q .i c :: UC 4> z | ., Z 4i JW 4 U ' 0 O - 4 4 >

° 1 O O O Q 0(V = O r

H r 44 . 4 34 M 1 w 9 Sk V3

Eq r n a o a - a a ce<: P; I e cP , a) u S a a c, I a ·I : * 3 U 3 ~~~~ 3 4 a ql ~~~~~L4 L ( cr;1 P } .~ ~~~ Pi Pi X ~~~4- CN 4i la 4 4 Q 4- I: EV <*

L3: Pt I|L z rn Z $4 M V 3 m OW W I 5

" ft E- D 3- 3 C O2 3 - D H 0 2w u E4 0 O 0 0 01 a:

III

W

HEd

zwEnwm,

0at,To

wHE-4H

inE~Pd0

4uCd

a

V]c)U

HE-4C)E~0

W4

vc

0

aa

CZ

04zw

To

EW

I3X Q 4 O 0 ,)

C Ix C a E _ _ C Pi '"0 4 X >, C 1 .ad > 4

) 0q Mt 0 0 4 .) to > 0 O.I u 0 _ 0 c,-, s - .,

b C N 0 C V C C V V VOE _4 tU: <a I 3 a X4.)4 4 0 0

"0 ~ - 0 4- ::I2 0 " 0 w 0 -J4r0 H -i io PIEt C r j4)

4) * >4e t >1 >' . 02o r t w C -

0 X -4 LI <QO2 4 2 C1 04 4 4J

5. L 4 a 4 ) ' .U O 0 W

0 4 &4 - P J ',~ ,' - 0 C) tu Eu (a - cC 0 N e -I. C C r- X

02 r 0 1 0 'U 02~ 00Eu *"m0 .i v I 0 ' W c4a e H I I a o I i C U -

o ·eo C 0 4) 0 0 0

I n o~ o_ _ ______

4 a . Q) C u * I I > C C

4.1 tCT i- -.

Eq4 II a- _ , ., ·r * a " 00 V X _.1 40 (a 04

O 0 u3 O '4 C 4 - u c. _ 4z X :T a U : C 40 a 4) 0

O _ t S V I O . : C I, .1C. )3 _ _ 00 .: .- , t .4) ,'02 c.e- _ ' O >4 _ _

I3 4 4 . "4X a V I 0 E V .

4 Z 4 0 CL L _ a * 2 toz t0 0 4 co0 C 44 0 4 W .1-4 440 > H 0 -W N 4-1 .4 L4 1H C

EH I C E 1 02f W)4 4j- OtE4 0 I U Z £ 1 o I I 4 £ V 4>

tz I£u LI 02*" 4 . *- 3 - .e .. _ -i ;-c|-4 X N o u x ! - V O O M -- M

> U w O 0 0 0 0 O I O - O UC) (- OL > aM0 C) V. _4 _

H

-

02E-,

z

HC)Ccn-t

UE-,

Z

E -4CuE-

rC)0aO

P1

0:E-4zCarz

0

.3

II'

0 Io _

Q .,.I 1i- : 0I) e 4) 0 u ·x I m *i e _ 4) r

U- 0 ~)(U ( . 44J 0. C )O 00 e C U r~r *- O5 a m~( O

'1 a 3 a 0 LL e 4- V

2.l.1 .) I O *w-4C,, I H O Q. ·aw O -0 0

a2 0 1:402 U 0 t i

0 0 a 4 ) )4.) 0 (aO

:4: (U 0 w C ) 0a I rr a c, ;r 4 Qa I .U 9 c c( e 00'60 0 ( 2 0 o ( >p 0

: z1 £ 0 Q C)3 04 ) OCo ° | *, X c £; sw m ouao I O u £ o e> O#4: Z O 0 Ps x Z

C I m ( O u W m o

H I c C( c * J C 4.4) £ 4

VI 02(Um 4 3 4 -. ci L- O2 0 c z (UO H i 4 C

CO Z0 I4 v o v o CEn z mY a Z 3 4J 4 U -

ZiI = a U 2 O J E- 4

'-,4 ( E U4 0

E-4 I C v C

Z 0 SZ4 J #E H I -4 > - 0 C.)

¢ H - z W 0 1___

H

:

EI

o

H

H'4

V)C-

0C

4:

4:C)

a(o

avEzuin0

CoE-4zw

'440

w 4

u a

-( Cu

0-04vsu

Q hZ :>4144)4)

.,4

4)0

w m

0

0O

u

Q02I C

'"4 Q 14 4w ' e

_ 1402

4I 3 n dE O5w P, O N

4)0 02 ' Cu

'"4C Cu C 4a k a C1 NmC £_- V Q O c

C n P ,·

0z

CE

0U

v =

_JC-4 .-_ X,,

0o

C Q1n 4)

'01-a Cux4 04 44 4)C 4) e

4.

'-40 14a . 4

-.i '"4 0 Z

4.2 Cu02

C1.4C o00

4)

4) 3

L4 02

N N 04

0z

oC 0

*H-E 02

U 0 4 4

· . oi I 03

et 0)m - Ne m-

0

4) .-4C 4i 4a 4)4.200

4.44) ,-Cc, ;

s.4 4) .)O 3

4) 0mV

> 014 ve 0 ms sad

C. N CP

0z

I 0 Q

3 0> N a

> Ct 0

·- 4 - 4- X ,.a .,4 a

C4 > 4 )E-4 14

o 0 O- MC.)

CO

4.4

4)

ZA

4

E4m4-

140

~4-0

4 wC

'4Cu(O(2

0

- k0 01

,4 4

C

e

U I

02 m

0 e

0 I

cu

.I

03

'I-N

.0

4)

4,I,

M02

Ct14M

14m

Cu

PC

'..4)"44.,-

Co 0

00

0.,4U

>

4 )

S14

0O) PO

0

L4 4

9 OCS1 0CA 0N 44

rnE-4

aO

0

E40

W

E-4

4cnE

adW

E1-441

I4U02ma

Iu

cnH

E4U

WHE-4u:0

4u

144)Cu

4)

0k~

0

Oi

U

5.4

4)cr

CuU

S

-VItj

u

004)

.40'4

ci.r4z

ZO031E-:D

044

30

14)

>1

Ct4)L40%

4J

'4

m02

4)4.)m02

4iC

,

4.)"4

0z

04

O0

0

u=a0

0z

N".4

alO0u

1..

004

0'O

=

r'4

Cu41o

~4O4

02co4)

0W'4C)

l

.,

a

O

O

ooo4.)'"4

3

c)14

u

0CD

C

C)

Q4

0=To

'"4

4-.4.)

To~~~~

C1404-Iw

.)4.02w4)140

4.U,

_

3

,It

0) CC n.f-4 l

14

0 0<1 o cwS

.,so4)4.24.)'"414341)

4

C0

00oCD

4.

4)

C0o

30

i

Q)

CD 0nz

E - o- I J0 uI u:

zO 0

C 0 4Z 4-

= V C

= Le 0

4VA: V 4

--

III III --

__

l | II Ill Il -

: I II I . -- Jll I

I>i"

a.4l"

O

CD

CI

01440

C.)

o4

o U00t-La

.-4

a.

00n

03

4

E,1-4,

Cm

E-4E-4C.)Cu

w0On

:30

E-4 'OE-4 C" c

V Swre E C

U 01

C,

4J

00

Cn

OZ

OZ

z

I ;

2 , - J 2 1. V3 · cl ~~~~n -

C

X,44.)1.4-W

m

CO

I I

4 - 4

O - >4 .,

u ai C r

4 : C )4) 4 4 ) u

H H XL Q 1 .

0 O C C 0O C) ::. O

;~~~cau4

02,14

C02O >

.H C

14

Xl

4.11.44 hi

>1>40 0 e.- O

O -H

Z C -C: 0 u

U. 4Ec:;

III

4)

CA

0C=0

E-4

C;

Ez.zc,

L44)ra

qi

0

0o40

U

Vci

.i"-

C.)

00

C"'!

C.4)4.)0

u1..0

H=

HW4u

w:

0

-4

O C 4-e

COI

Q Qu LrOr

U.) C

C L4W )

4 O 44) C1.4

I 4.

C Cm-4 a C-H OM 4)

0 44 r

C

z >m 4(I -4

-.4 *l"

14

0 04(aO Pi

L4 L Q

tZ

C24E-z(c

CO

<::

vH

E-4

cv5:

02>4

u

c14

4.)

x4)

0z04

4)

0Q)

QU

(U

(

4.)03m

Z

0u

IO ..a -a

I . I

>nZ 1

l E-,

I =

1.4

.1-4 Q4-) 44

I

O 44. ,: .

14

04)

)

o04

1.4

u.,44)I

O0

4.

4;

C

O.4I )n

a I 0I 4004IM

t_

a:>r

* rr'-42 ICIs.*HcoX

c 2

1:

in

03

I 4

'-4

ul)

co

N

.,I

X:

co4.)4)4)Q,020('1

02

4.)4)4)co

0U,0

4.)

ad

1e

CO

O

43ca040

.JZ

(3) C4

In

14

a4i04

0

= >.

-1

X e

>4

m

C

Ea Ia)v Iu0

QCf4uu M]S <

l l lI J - -J --

--

I II . S -

I - - -- -

-

~~~- ---

I M 4

., .e

~; *4O0

40 U 400 0

4 ' 4J 0 4C r) 9 0 00 - to 4"o o C i oa

-0-Q) C)4 .4 0 ., _ C i

0 0)x 4) 2 I 4 . .J 0

m 0 04j.) 4) 4 - 0 C'40 4 54 W o4 d4 .i i o1 4.)

00 3 0- 4 .0 0

4.

%404.4

0

4)

.).94

'4

4E I0"0o Ci 4..4Q0 w 0 0 -

I4 OE.4

>_4

C.,4.9-m1.4.)4)

0

4).,4

3mw

4>4

U)E4

Ca

0

14raPe

W

C)pi

0

14uC4

HE414E

3d

0

HWC4Hlz

'4

wEz

U]:00

ra

C.)

E4

9:

C.H4U)H

14

C.)H9-4cnE-4

4:C4xu0

1.44)I

4)

C.)0H;o

0000%

a$au

M.4

4)#"4M)

C.)O

0o

4)

054C).9I

. -

m

4H

-4w

H

.

i;1.1 I

al

..44-·rl4)

040

U

0 -4 ,

4i ac(d O C

u 3 _.,. 4

U 4).9-4 w

M - t

0O.,-44.U4)

Q

4)H00

.944.

04.)

a

U.,944.flI2

04.z4

0V

O v O

lu4 XrLd

C4,-4

I

041-4

-4

'IO

O4 0N -4·94 4.

I -W4i .-4C:

O ~U

0

c m C >O C

., .

aC C(O4.) 0

e Vj -m 0 o

4) 4)0c

o o0

-4 C 4v i

4i C 4 )."I - > . ;

- .4. L. 4) -

ra (L >'r4)02>4-I 4) L C 4 xm I4 I-

0 4.C4.)

0244 4)

o =

e 4n an( 02 . )= 0 2 042

I U

, I>4 w w

. 4 4.) .-4 $4 "

(- 4 ,.,-iJ i) -

O Z,M 0.-

>-

o

o

0

.4I

0

4)

0

01t

0

0cI

-

*,r

4

4)a

C).4.)

-I2

4)

3E4.L434)0.>1E-4

M -4

4) CO1.-4

UO O0tOZ 4

.-.uI2

4)-4

.0ni,w2.r.)la

to

z

04.

OrC C,4 u0O0 0

O

C4.) 4i1 mo 2

co

w *

X .4X0Q Q

+J a4.4L

0"44&02m W

-4)4 m> _

QCI >

QUe _4

.94

-4i1 .-4

02X

C .u T

4)I-4

I.4.4.ea

0z

I I4) C

S4 e

C

w I=0 4.r4 OC C

4 -

4-4 - .-40 C4) -4 Il

P- L4

C

eO C

C) U ,-

4-.0I

.,4

IC0

>1

.9-4114

I.9-4

1..

E-4

m41)

>4

.9-4

CO0

U4

-j C.4

.4

,4 V= in2

- --I l

-

_ X - I l - -- -

l I | - -II IIII-

---·---------------� ----- ---·---------�------�-��---� --- �

coC4

I

%0

0 4 Ia) S O ^* >,4 0 * .-4 . 4-4 Ca

.4 Q

Q)

V l

U (.,-I

>1

*-(0

(o4)

.p-4 0

0oQ

( U00 4-)

co

4. ) -4 4)- 1.

2 00o *. .4

U01.4 I

j 0- (Ua .

. .4 m (U

Ir.Q

0 440

4)4.)

4. 10

,.4)0 Qm,

1 C0Q)4.44.)

r 4)

-H 0 4 Q)

. 0a Q 04 U U0· 4) 0.'H 0 0=

t 4 ))o-H 4 ( u u

. 0 3 0 S

9-4 �4 (U4 W 4-44) ( 4 0. a

> V 0 2 o .) -HO~G~.~3 .e ~ ~r

(U0)(U. 23ca3

En w

rr;CC% m

II'

'0

E 4< UM wq

0O 00414 (11: 0

oU,

c0

4.J

4.-

03

co

I

C)0

C4)14

C4)4.)

NrmV

MG

CP

P4)1.4mW

4.

044

(U

0.9-44.)0443)0

XW

02UZE4

ua0

Zd0

E0nix0u

En

H

0E4

Wwo

04C)

E

Hm

C)a

:

E

;

U

r'0

Hl a H I a

.,q iC); eU]-H gnlQ r

H(U021

O *H

H OaC)

v04

0

0 0O X

0Ul2

30 0O

04

co

z tH 44 a%Za oZ;

E-4C

di 4ila

0

1..4)(U

4)

0H

0o0

04C.)a

(u4)(UC)

C.)0

0m'1

4)4.)0

..0k9-

zOC)H H

<

E4H

al0.-4

4.

US4

0 e

a0(U3 C

M M

0O4)

I 00 I (

4 04'00 0

41 _,1-" 0:I 0OCOt)

'004.

0 . m

0 H

0 4 aHs V _

0

02

04)0014

la

030

w O

4 Q)

Q

-W Ca mM m 1-4)£= 3

o 04()

00Qa4 M

4i 3 O~= - l0

4.)4)C(1)la'U

la4)w04)

2ea

C Q)r. C (1(a 4i Q Qra a a

O )0(U , )

0(UF

-.. D 04e C t

4) 0'0 - 4)

4)4) C 4

02e 00 -0= tv IC

4-J4)4)04)0441)

4.)w00n

4 A

Q- Q

0 C 0

I- 0

0

Iea(UQ

.4,

3 oC

o o

O m(U

o.,-3

o ..0 00

o 0 m

C .l-C(-, ml .

o0

LnIC4U,

0.4.,9-iV

m MU o

W0 -

1.4 .6,Le XnL 4)

H C '4 C

4) Q CNOz.4 CAJ - c2

0- 3U

I

OU2

0H 3

X C

04 Q-

C)l -

- -

l l

.

S 4- O Q ~ - _ - * 0 L-

Q Le i Q 3 U ·r a

a I s XX , az n a 95 I Ei so -m a Q

(1 pq 0 4 Q Q >4 4) & X@

c I-~ o, a o a n :z: X X e V ,I~~~~~~~~~~~~~~~~` C4 4i ." IV s;-cJ I c; ·c o ;X aa 4 o 0 i

O I a a nx >e o SrlW a~ II J no rd j EP v m c w o3

I o

re $e I X4-J 'd I , - r .J

0 F. .1 .. (JO ~1 O J 0 e a ~, · r ~"I 4 S 4 O _P 0C~ 4,) O .1 0J Q ad :3, . 0 44 W,,,to 0 4 n m : >1 in

o wt 4 o o o a ss~ a 3 ·

,-O I V I *' :: : C C : 1 .0 I2_C Z a C : -- c .-- ,,-- O L rJ c C

a I c o3 · i IO

C tX oe a r4 X

,- I - o0 0 . 1 C O 0 1 e C0 I .~ ·4 .i O 4 G a 0 / ,-- .

.-4 .0 0. ::: .W 4 4J M,-~,-~. > ,pq W<3 > " C to tv3 Q

_~ · ,I 0 a 0- C 0 3 C . V m 0- Z 0Ow 4J Q - P O 0 a 4 O C 3Z 3 C z ·e *-- I a V; cX a

It C :4 44 /,,4 & :> , m C r4(

Q 4i. -4 C~ 4-.. > . d 4)O 0 >~ 0 4I .rfI *H) Xi 0 X-4Z

1Q O , .Q , . ., *H .... I . O . X . .

i Q4 ,. _0 I -f a EI 1 M, 04 O XXU V -3

Io u S V * Q O a O --4V I rl C M 4 4 X -.H q Le w 4

I a a O *Hr · ; a L( ·c ao I a o: sr m a a v u c E; *4o I X 4i u ( 4j C) a o 4 V u

I ) v .* q a 0 0, = 4>: I Q _ _ _ e :4 X X Q X n 4~~~~U W 0 4-

a I aa a ft a er X >Q 3 a ss v :O, I1 4;~- O; C {3 > Q1 o: o 4 a uz o rre · rl ·~~~~~~ a) C1 3 L a) k~~~~~~~(1).0 W 1

k I .<, .^ Q~(1 = 3 -;4 (1 Q - -4 _ -_ X > -WVr4 I" ra a C L Z " M OJ C C 0 toX

*, I * o si 3 o o E n £ o a s, i Q; P Or

m i ct *$^ I 4 £4Z I >i Z

D E-e I ;/ e _ z ; X;

; Z i O X 1 O: : ~ I *n, Q ZE-4 (z I L, 4 X O

X I: s:·~ - 4~ Z? ; i 4 $C W DI=~iZ ~1

_��_�_

HEd

Z

i-4zD,co

nW

EDWa,

CmWE-4

Pr

7(4(nC.)H

w

UP;I=:4

(aEu0DmO

Di

uO

unzw;

I)

Wd

Es4

A;V: