An environment for content-based image retrieval from large spatial databases

Ž .ISPRS Journal of Photogrammetry & Remote Sensing 54 1999 263–272

An environment for content-based image retrieval from largespatial databases

Peggy Agouris ), James Carswell, Anthony StefanidisDepartment of Spatial Information Science and Engineering and the National Center for Geographic Information and Analysis, UniÕersity

of Maine, Boardman Hall no. 348, Orono, ME 04469-5711, USA

Received 30 September 1998; accepted 29 April 1999

Abstract

In this paper, we address the problem of content-based image retrieval using queries on shape and topology. We focus onthe particularities of image databases encountered in typical topographic applications and present the development of aspatial data management system that enables such queries. The query requires user-provided sketches of the shape and

Ž .spatial configuration of the object or objects which should appear in the images to be retrieved. The objective of the searchis to retrieve images that contain a configuration of objects sufficiently similar to the one specified in the query. Ourapproach combines the design of an integrated database with the development of a feature library and the necessarymatching tools. In this paper, we present our overall scheme, introduce some individual database components, and providesome implementation results. q 1999 Elsevier Science B.V. All rights reserved.

Keywords: image retrieval; feature libraries; shape similarity; topology similarity; matching; spatial databases

1. Introduction

The intelligent retrieval of images from largedatabases is the focus of substantial research efforts

Žwithin the computer vision community Jain, 1992;Gudivada and Raghavan, 1995; Ogle and Stone-braker, 1995; Pickard and Minka, 1995; Cohen and

.Guibas, 1996; Cox et al., 1997; Gupta et al., 1998 .The objective is to retrieve specific images from alarge database by querying the properties of theseimages. As a result of pioneering research efforts,certain prototype systems have been reported, with

Ž .some notable examples being Virage Gupta, 1995 ,Ž .Chabot Ogle and Stonebraker, 1995 , IBM’s QBIC

) Corresponding author. E-mail: [email protected]

Ž . ŽFlickner et al., 1995 , VisualSeek Smith and Chang,. Ž .1995 , ImageRover Sclaroff et al., 1997 and

Ž .PicHunter Cox et al., 1997 . Most of these effortsaddress the problem within the context of multime-dia applications and therefore, focus on general-use,multimedia-type image databases. In such applica-

Žtions, low-level image properties e.g., colour and.dominant pattern are often adequate for information

retrieval since the image members of the databasedisplay substantial differences in these properties andcan be distinguished by them alone.

Advancements in remote sensor technology haveresulted in increasingly larger and more compleximage databases depicting characteristics of the

Ž .earth’s surface e.g., topography . Thus, the commonpractice of using general metadata structures to ac-

0924-2716r99r$ - see front matter q 1999 Elsevier Science B.V. All rights reserved.Ž .PII: S0924-2716 99 00025-8

( )P. Agouris et al.r ISPRS Journal of Photogrammetry & Remote Sensing 54 1999 263–272264

cess specific images is becoming less effective andintelligent image query techniques are needed. Topo-graphic image databases, however, are unique in thesense that they contain large numbers of imagesŽ .typically aerial andror satellite which are verysimilar in terms of general low-level image proper-ties. General-purpose image retrieval approaches suchas those mentioned above are not as effective forinformation retrieval in topographic applications.What distinguishes images in a topographic databaseis their actual content: the shape and configurationof the objects they contain. Accordingly, shape andtopology are more suitable primitives for use intopographic image retrieval systems.

Common approaches to shape-based retrieval sys-tems often employ complex shape descriptors toestablish indexing structures. Using these descrip-tors, shape similarity becomes a nearest neighbour-type search in multidimensional spaces, often sup-

Žported by tree structures like k–d trees Stein andMedioni, 1992a,b; Califano and Mohan, 1994; Del

.Bimbo et al., 1994; Beis and Lowe, 1997 . Whilethese techniques tend to be reasonably successful, acommon pitfall is the relatively low interpretationpotential of the retrieval failures. The similarity of aspecific shape to another is often hidden behindcomplex descriptor metrics.

In this paper, we address the development of amore direct and organised approach that uses rastershapes in support of shape-based image queries. InSection 2, we present system architecture considera-

Ž .tions behind Image Query IQ , our prototype sys-tem for image retrieval. We provide an overview of a

Ž .matching tool to support such queries Section 3and emphasise the role and organisation of feature

Ž .libraries for image retrieval Section 4 . In Section 5,we address the use of topology in queries. Someexperimental results and our future research plansare presented in Section 6. It should be mentionedthat while our research originates from topographicapplications, the developed methodology could beapplied to any type of imagery.

2. System architecture issues

In our view, a searchable topographic databasecomprises images, outlinerobject information for

Fig. 1. Operation environment for IQ.

physical entities depicted in these images and meta-data. Metadata of interest include the typical infor-mation that describes and enhances the contentandror properties of common topographic data filesŽe.g., sensor characteristics, date of capture and

.scale . A schematic description of the operation cy-cle in our system is shown in Fig. 1.

Ž .Our goal is to use as query input: 1 the outlinesŽ 1.i.e., a sketch of an object or a configuration of

Ž .objects; and 2 metadata. The operator provides thisinput information and the database is searched toyield images which satisfy the given metadata pa-

Žrameters and in which appear objects and, if appli-.cable, spatial configurations similar to the given

sketch. This allows us to take advantage of theintuitive method humans use to represent or expressspatial scenes, namely sketching. In order to supportsuch query scenario, our searchable database com-

1 The term sketch is used here in a broad sense. It can refer toan on-screen drawing of an outline using software tools. It can

Žalso refer to a pre-existing outline e.g., from a digital image or a.scanned map selected by the user.

( )P. Agouris et al.r ISPRS Journal of Photogrammetry & Remote Sensing 54 1999 263–272 265

prises three components: image, metadata and fea-Ž .ture libraries Fig. 2 .

Ø Image library: contains one entry for everyimage contained in the database and provides alinkrpointer to the corresponding filename.

Ø Metadata library: contains a listing of potentialvalues for a set of attributes that describes generalproperties of the images and links to members of theimage library. These attributes include date and timeof acquisition, date and time of introduction in thedatabase, scalerresolution and location of the imageexpressed in hierarchically arranged geographic enti-ties such as state, county and city. For more complexdatabases, the attributes may be extended to incorpo-rate information on the sensor and imagery typeŽ .e.g., BrW, colour or hyperspectral .

Ø Feature library: is an organised arrangementŽ .of distinct feature outlines i.e., object shapes and

Fig. 2. Database design.

links to image files where such features appear. Therole of the feature library is to provide the cruciallink that allows us to reduce the search space of aquery from a large image database to an abridgedgroup of features.

In addition to these three libraries, a semanticlibrary may exist that contains semantic object infor-mation such as use and ownership. Using the abovedatabase design during the on-line part of a query,the input query outlines are matched to elements ofthe feature library using an on-line matching tool.Acceptable matches then give links to specific imagefiles and locations within those images. A compari-son to the input metadata values renders certainimages invalid candidates. If a semantic library isalso used, the matching results will have to passthrough another check whereby the semantic proper-ties of the detected objects will be examined. Thequery results are returned to the user who then hasthe option to edit hisrher query.

In order to support the above process, anothersequence of actions has to be performed off-lineevery time an image is introduced into the database.The user manually inputs the appropriate metadatainformation for this image and the metadata libraryis updated to add the new entry. Subsequently, ob-jectsrfeatures are extracted from the input imageusing any digital image analysis tool and these newfeatures are compared to the existing feature libraryusing our off-line matching tool. Library entries areupdated accordingly to include links to the objectsexisting within the newly introduced image. Linksbetween metadata and feature libraries are also up-dated to connect the metadata values of the newimage to the detected features.

The rationale behind our database design becomesapparent when analysing the meaning of the meta-data, the feature libraries and their connection. As-suming that we have n distinct metadata values, apoint within the n-dimensional metadata space corre-sponds to all images of the same area, captured at thesame scale, at the same date and with a similarsensor. When one or more of these parameters canaccept less specific values, we move from points to‘blobs’ within the n-dimensional metadata space.For example, photos of various scales of a specificarea taken on a specific date form a blob in themetadata space. This blob represents the scale space


of the area at the time of data capture. By definingŽ .points or blobs of the metadata space, we actually

define a set of representations of a specific geo-graphic area and of the features within it.

3. Matching tool

We have developed a matching tool to handleŽboth the off-line population and update of the fea-

. Žture library and on-line comparison of query input.to the feature library entries matching parts of this

project. Conceptually, our matching algorithm is aŽ .variation of least squares matching LSM , modified

to function with edge files. It is well-known thatLSM offers a robust method for establishing corre-spondences among image windows. Its mathematicalfoundation, based on least-squares principles, per-mits its successful extension for application in multi-

Ž .ple image matching Agouris and Schenk, 1996 ,Ž .feature-wise global solutions Gruen et al., 1995 or

even to establish correspondences in sets of 3-DŽ .images Maas et al., 1994 .

Our matching tool is based on the analysis ofŽdissimilarities between a template user-provided

.sketch and an outline file window. When a templateis compared to an edge file, the template is dividedinto four quadrants. Each quadrant is matched sepa-rately to a corresponding image area, and template

Ž .pixels vote to remain in place or move according toŽ .their similarity or dissimilarity to corresponding

file pixels. Moves can be performed in the "x and ydirections, in one of five options: left, right, up,down or remain in place. This resembles the compar-ison of grey values in LSM and the use of imagegradients to identify shifts, rotations and scaling. Thesum of the votes within each quadrant is recorded,together with recommendations for the direction andmagnitude of move. An example of template reposi-tioning situations may be found in Fig. 3. An analy-sis of the votes of all quadrants allows us to solvesequentially for shift, rotation and scale parametersin order to reposition the template and better matchthe file object. The final solution is obtained after aset of iterations. Furthermore, by analysing votingpatterns within different quadrants, quadrants may beturned ‘off’ to accommodate occlusions.

Fig. 3. Examples of template repositioning.

During the on-line matching process, a featuresketch template is matched against the feature libraryelements using the same matching tool. Once thetemplate has settled onto a match, the matchingaccuracy is determined by how many of the templatepixels continues to vote to move compared to thosethat vote to stay put. The comparison of a templateto various feature library entries will result in match-ing accuracy estimates. These matching estimatescan be used to determine ranking of differentmatches. During the off-line matching process, fea-ture library templates are compared to outline filesŽe.g., edge files or GIS layers co-referenced to digital

.image files to establish matches and links to thecorresponding image files.

4. Feature library: hierarchy, structure and main-tenance

The feature library permits us to narrow the searchspace from a large database of images to a limitedgroup of feature outlines. In order for the query to beefficient, the library needs to be organised in anoptimal way. The ideal optimality criteria are two:the members of the library should be exhaustiÕeŽthus being able to describe all possible query input

.features , and the members of the library should beŽ .independent minimising unnecessary duplications .

The two properties, when satisfied, are equivalent to


an ideal library, which is approaching a base span-ning the space of shapes.

We have developed a feature library that uses aŽ . Ž .hierarchical tree-like structure Fig. 4 to approxi-

mate the abovementioned ideal optimality criteria.The organisation of data within the feature library isintended to enable it to act as a progressive filteringmechanism. Matching results at a specific tree levelare used to eliminate complete branches and, alterna-tively, to consider only few branches. This reducesthe search space from the complete tree to a progres-sively refined limited subset of branches. The libraryorganisation employs two parameters: a matchingpercentage above which objects are considered to besame, and a percentage range within which objectsare assumed to be similar. The similarity range isobviously upper bound by the lower rate of thesameness range. Combined, the two parameters de-fine the degree to which the feature library ap-proaches the ideal library outlined previously. Highvalues for same ensure the independence of elementsat the same level of the library tree. In Fig. 4, the80–100% range was used to define same and the50–79% range to define similar. During a queryloop, a query sketch is first matched against thefeatures at the parent level in the tree. The processthen progresses to child, grandchild and other levelsas necessary. Using these intervals for illustration,the hierarchical levels of the library are organised inthe following manner.

Ø The primary parent leÕel: The primary parentlevel is the level against which every new queryshape is first tested and its respective matching

Fig. 4. Example of the feature library hierarchy.

percentage is recorded. Features at this level are theroots of the feature library. Different features withinthe parent level are considered dissimilar if theirmutual similarity percentages are lower than 50%.When a query is performed, the matching percentagebetween the query feature and a parent feature willfall in one of three ranges: 0–49, 50–79 and 80–100%. In the latter case, with a matching percentagein the 80–100% range, the query feature is consid-ered to be the same as the parent feature. The linksof the parent feature to files in the image databaseare then returned as the results to the query. If morethan one parent feature matching 80 to 100% to thequery feature exists, the links of the top matching

Ž .parent higher matching percentage are returned first,with the links of the remaining matches following. Inpractice, this case of multiple parent matches isunlikely due to the mutual dissimilarity of the parentfeatures. If the matching percentage is in the 0–49%range, the query feature is considered to be ‘differ-ent’ than the parent feature. If this holds for allparent level features, the query feature is a candidateto be inserted into the feature library at the primaryparent level. Subsequent off-line processing is thenrequired to establish its links to images in thedatabase.

Ø The child leÕel: If the query feature matched inthe 50–79% range to any parent feature, it is consid-ered ‘similar’ to the parent and is then tested againstthis parent’s respective child features. In the case ofmultiple parent candidates, we select the one withthe highest percentage first and continue with othercandidates in order. If there are as yet no childfeatures for the selected parent, the query featuregets inserted into the feature library as a descendentof the best-matched parent. Its links are then addedand prioritised through off-line matching. If childfeatures already reside at this level for any of theselected parents, the query feature is matched againsteach of them. For a matching percentage in the80–100% range, the query feature is considered thesame as the corresponding child feature and the linksof this child feature to the image database are re-turned as the results to the query. If the matchingpercentage is in the 0–49% range, the query featureis inserted into the feature library at the child level ofthe best-matched parent and additional links areadded through off-line matching.


Ø The grandchild leÕel: If after matching in the50–79% range to a parent feature, the query featurematched in the 50–79% range to any child feature,the query feature is tested against the respectivegrandchild features. Obviously, the relationship be-tween child and grandchild follows that of parentand child. The processes already described take placeat this library level to identify the ‘same’ grandchild,to establish a new grandchild or to move furtherdown the tree to the level of great-grandchildren.Combined, these similarity tests have to be satisfiedwithin branches of the tree structure form the inser-tion-testing criterion. The children levels may bealternatively referred to as n-child leÕels, with thechild being 1-level, grandchild being 2-level, etc.

The matching parameters mentioned here are in-tended to serve as example values and are not unique.They can be chosen experimentally or even arbitrar-ily. The percentages may become lower or higherand can be considered as tuning parameters of theapproach, with their variations affecting the structureof the feature library in a predictable and organised

Ž .manner. By narrowing the high same matchingrange — e.g., using the range 90–100% instead of80–100% — we increase the number of distinctentries at each level of the tree structure, making thetree wider and more shallow. This will increase thesearch space and make the processing time longer.However, it would also allow the queries to becomevery specific. For example, the queries might be:‘retrieve images containing features which look ex-actly like the query sketch’, as opposed to ‘retrieveimages containing features that look like the querysketch’. Widening the high range — e.g., using a70–100% matching range — will result in fewerentries in the database and will make the tree deeperand narrower. In turn, this produces faster querieswith less accurate results. Narrowing or widening the

Ž .second ‘similar’ matching range will have compa-rable effects to the structure of the feature library.

The above-described method, whereby new fea-tures are compared to existing ones and progressthrough tree levels until they are eventually matchedto an entry or inserted in the library, might result insome inconsistencies. It is possible, for instance, thattwo sub-children from different primary parentsmatch closer together than to any other feature cur-rently within their respective tree hierarchies. This

and other potential inconsistencies are rectified byapplying general feature ‘housecleaning’. Duringfeature housecleaning, we consider the sequence inwhich features are inserted in the library to examinethe presence of potential inconsistencies. The processis performed off-line at regular intervals, to ensurethe continuous proper organisation. The objective isto ensure that all n-child features are residing undertheir proper parent, and to minimise the theoreticalpossibility that unnecessary duplicate features could

Žexist somewhere in the tree hierarchy i.e., differentn-child features that are 80 to 100% similar but

.reside under different tree branches .Feature housecleaning utilises a self-organising

Ž .table called the ‘Temporal Feature Index’ TFI . TheTFI is a two-dimensional cross-referencing table offeature filenames together with their respectivematching percentages. Features inserted into the fea-ture library are simply appended to this table in theorder that they are introduced in the library and alltheir matching percentages recorded as they occur.Thus, ordering within TFI reflects relative insertiontime. Feature housecleaning searches the TFI top–down and compares features to other features in-serted after them to detect inconsistencies. A detailedanalysis of feature housecleaning operations may be

Ž .found in Agouris et al. 1999 .

5. Handling topology

To extend the queries in configurations of objects,we employ the well-known concept of nine-intersec-tion, describing the major topological relations be-

Žtween areal, linear and point features Egenhofer and.Franzosa, 1995 . According to this model, the topo-

logical relationships between two objects are ex-pressed by a 3=3 matrix whose elements expresswhether the mutual relationships between the inte-rior, exterior and outlines of two features are emptyor non-empty sets.

In order to consider topological relations betweenobjects, we break the query for a spatial scene intotwo tasks. First, we identify where individual ob-jects, similar to the query outlines, exist in thedatabase. In order to collect this information, weperform a separate query for each of the distinctoutlines which comprise the query scene. This results


Žin lists of links to image locations one list for each.query object . The location of an outline in an image

is expressed by the image coordinates of its centroid,as well as the top-left and bottom-right coordinates

Ž .of the feature’s minimum bounding rectangle MBR .The topological relations between combinations fromthese lists are then determined by using the MBRcoordinate information. By comparing viable combi-nations to the topological configuration determinedin the query input, we decide whether the configura-tion constraint is satisfied. For example, when at-tempting to retrieve raster files that depict a hexago-nal building and a cross-like structure separated byan L-shaped object, we identify the locations in thedatabase where each of the three objects exists.Subsequently, we analyse this information to pro-duce the combinations that satisfy the desired config-uration. This formulation of scene queries as a com-bination of a space query and a topological checkwithin our system permits easy and fast query edits.When attempting to edit a query by reconfiguring aset of given objects, we do so by re-examining theirtopological relations only, an operation simpler thanre-searching the shape library.

The use of MBRs instead of the actual outlines isdictated by practical considerations. Establishingtopological relations between complex vector out-lines is a cumbersome and computationally intensiveprocess that is highly unsuitable for query opera-tions. MBR coordinates are determined easily in astraightforward manner by using matching parame-

Ž .ters rotations, shifts and scales to position a featurelibrary template onto an image coordinate system.The use of MBRs speeds up the topological buildingprocess and thus allows for topology to be built andqueried in real-time. If we are interested in consider-ing topological relations which extend beyond singleimage files, we have to substitute image coordinatesby ground coordinates. This could lead to the re-trieval of sets of images, instead of a single imagefile, in response to a single query.

Even though it is beyond the scope of this paper,we should mention that topological queries could beextended to include direction and distance measures.However, the reason we first consider topology isbecause it has been shown that when users interac-tively provide sketches of object configurations for aquery, their topological relations are of foremost

Žimportance and of higher reliability Bruns and.Egenhofer, 1996 .

6. Experiments and comments

A prototype system of IQ, the image query envi-ronment described in this paper, has been imple-mented on a Windows NT workstation. This systemincludes a sample database of 188 features thatprovides links to a database of approximately 70image files. These features comprise a grouping ofobjects which were either manually sketched or ex-tracted from the processed image files. Typical ex-amples of manually sketched features were mostly

Žgeneric shapes e.g., circular, representing cooling.towers; rectangular, representing buildings . Features

extracted from the image files, on the other hand,Žwere rather unique in nature e.g., outlines of air-

.planes, buildings and other visible image objects .As a general rule, generic shapes have links tonumerous locations and often multiple locationswithin a single image. Unique shapes, however, havefewer links that tend to be spatially localised. Forinstance, numerous airplanes may appear in an air-port area but none in other areas.

Search and retrieval times for this feature librarywere tested using a single 180 MHz Pentium work-station. First, we tested the matching algorithm bycomparing one feature against a complete but un-structured feature library that was not organised intoa hierarchical tree structure like the one described inSection 4 of this paper. In this case, the feature hadto be compared sequentially against all features inthe library before determining the best match andreturning its links. For the abovementioned librarysizes, this corresponded to search times averagingaround 6 min 30 s per database. On average, it tookapproximately 2 s to match the query feature againsta template in the feature library. Variations in thistime interval corresponded to the differences in ob-ject complexity, which might require more iterations.

The next step was to organise the features in thefeature library into their proper tree hierarchy ac-cording to the rules outlined in Section 4. We usedthe 80–100% range for same, 50–79% for similar,and -50% for different. With these parameters andafter house cleaning, the feature library settles in a


tree structure that is 98 features wide and up to fourŽ .child levels deep after the parent level , with 23

features removed as unnecessary duplicates. Thiscorresponds to a trimming of search time nearly inhalf, as the minimum number of features to compareagainst a query template is reduced from 188 to 98.Considering that the tree structure is up to five levels

Ž . Ždeep parent plus four child levels , the time cost at.2 s per feature matching to move down a tree

branch is rather minimal. It should be noted here thatas the feature library increases in size, its momentumalso increases. In other words, a 10-fold increase inthe number of templates would not correspond tosimilar increases in the size of the feature library tree

Žstructure. Instead, the number of duplications within.the same range would increase substantially with

few new templates added in the tree structure itself.We then proceeded with examining the effect on

the feature library structure of changes in the sameand similar intervals. Table 1 shows the change inthe percentage of features included in the featurelibrary as the lower bound of the ‘same’ percentagechanges between 100 and 51%. During these experi-ments, the ‘similar’ percentage was kept between50% and the lower bound of ‘same’. The tableshows, e.g., that assuming the sameness interval to

Žbe 75–100% and correspondingly, the similarity.interval being 50–74% reduces the size of the fea-

ture library by 11% compared to when a strict 100%is used to determine sameness.

While Table 1 shows the effect of the samenessinterval on the total number of distinct features in thelibrary, Table 2 shows the effects of variations of thesimilarity interval on the structure of the library tree.Specifically, Table 2 shows how the percentage ofparent features changes for varying similarity inter-

Ž .vals for a fixed 80–100% sameness interval . Wecan see how by broadening the similarity interval,features in the library migrate down the tree branchesand the number of parent features is reduced. For

Table 1Effects of varying sameness interval on the percentage of featuresin the library

Percentage of 100 98 97 95 94 89 82 79 74 68 61features

Ž .Same % 100 95 90 85 80 75 70 65 60 55 51

Table 2Effects of varying similarity interval on the percentage of parentfeatures in the library

Percentage of parents 98 82 73 57 48 29Ž .Similar % 79 70 60 50 40 30

example, for a 50–79% similarity interval, we expectapproximately 57% of the features to reside at theparent level. By reducing the lower limit of similar-ity, we reduce the percentage of parent features inthe tree. Accordingly, this can be used to reduce thesearch time. The obvious cost for this would be lessprecise results.

Fig. 5 shows an example of the query results. Thequery element is shown in the upper left window. Itmatches successfully to an element of the featurelibrary shown in the lower left window. The libraryelement was introduced in a previous operation andwas extracted from an actual image.

These early results support the effectiveness ofthe overall strategy and of the hierarchical organisa-tion of raster feature templates for image queries.Regarding image queries specifically, one shouldconsider that the size of the feature library increasesat a much slower rate than the increase of the imagelibrary. The feature library eventually reaches a criti-cal mass after which additions of new features arerare, since objects within new imagery are alreadysufficiently represented in the feature library andthere already exist similar templates within it. Bytransferring the query search space from the imagedatabase to the feature library, we manage to limitthis search space and increase the effectiveness ofour technique.

Our future work plans include the extension ofour experimental feature library and image databaseto include diverse topographic datasets, especiallytopographic maps. Furthermore, we plan to extendour system to function for spatio-temporal queries. Ina manner similar to the approach presented above,spatio-temporal queries can be formed as combina-

Ž .tions of spatial queries where two or more sets ofquery results are forced to satisfy certain temporalconstraints. For example, we can attempt to retrievesets of images depicting configurations of buildingsbefore and after the addition of a new wing. This is


Ž . Ž . Ž .Fig. 5. A query template top left matches a feature in the library lower left and receives its links marked by windows in the images asresponse.

equivalent to searching for the before and the afterconfigurations separately and selecting from the two

pools of candidate images those files that depict thesame region but differ in their temporal metadata.


Acknowledgements

This work was partially supported by the NationalScience Foundation through grant no. SBR-8810917for the National Center for Geographic Informationand Analysis, and through Career grant no. IRI-9702233.

References

Agouris, P., Schenk, T., 1996. Automated aerotriangulation usingmultiple image multipoint matching. Photogrammetric Engi-

Ž .neering and Remote Sensing 62 6 , 703–710.Agouris, P., Stefanidis, A., Carswell, J., 1999. A feature library

approach to on-line image querying and retrieval for topo-graphic applications. Vision Interface ’99, in press.

Beis, J., Lowe, D., 1997. Shape indexing using approximatenearest-neighbour search in high-dimensional spaces. Proc.CVPR’97, pp. 1000–1006.

Bruns, H., Egenhofer, M., 1996. Similarity of spatial scenes. Proc.of Spatial Data Handling ’96, Vol. 4A, Delft, Netherlands, pp.31–42.

Califano, A., Mohan, R., 1994. Multidimensional indexing forŽ .recognizing visual shapes. IEEE Trans. PAMI 16 4 , 373–392.

Cohen, S.D., Guibas, L.J., 1996. Shape-based indexing and re-trieval: some first steps. Proc. 1996 ARPA Image Understand-ing Workshop 2, 1209–1212.

Cox, I.J., Ghosn, J., Miller, M.L., Papathomas, T., Yianilos, P.,1997. Hidden annotation in content-based image retrieval.Proc. IEEE Workshop on Content-Based Access of Image andVideo Libraries, pp. 76–81.

Del Bimbo, A., Pala, P., Santini, S., 1994. Visual image retrievalby elastic deformation of object sketches. Proc. IEEE Sympo-sium on Visual Languages, St. Louis, MO, pp. 216–223.

Egenhofer, M., Franzosa, R., 1995. On the equivalence of topo-Ž .logical relations. Int. J. Geogr. Info. Syst. 9 2 , 133–152.

Flickner, M., Sawney, H., Niblack, W., Ashley, J., Huang, Q.,Dom, B., Gorkaani, M., Hafner, J., Lee, D., Petkovic, D.,Steele, D., Yanker, P., 1995. Query by image and videocontent: the QBIC system. Computer 23–32.

Gruen, A., Agouris, P., Li, H., 1995. Linear feature extractionwith dynamic programming and globally enforced least squares

Ž .matching. In: Gruen, A., Kuebler, O., Agouris, P. Eds. ,Automatic Extraction of Man-Made Objects from Aerial andSpace Images. Birkhauser, Basel, pp. 83–94.

Gudivada, V., Raghavan, V., 1995. Design and evaluation ofalgorithms for image retrieval by spatial similarity. ACM

Ž .Transactions on Information Systems 13 2 , 115–144.Gupta, A., 1995. Visual information retrieval: a Virage perspec-

tive. TR95-01, Virage, San Mateo, CA.Gupta, A., Santini, S., Jain, R., 1998. In search of information in

Ž .visual media. Communications of the ACM 40 12 , 34–42.Jain, R., 1992. NSF Workshop on Visual Information Manage-

ment Systems. Redwood, CA.Maas, H.-G., Stefanidis, A., Gruen, A., 1994. Feature tracking in

3-D fluid tomography sequences. ICIP-94 I, 530–534.Ogle, V.E., Stonebraker, M., 1995. Chabot: retrieval from a

relational database of images. Computer 23–32.Pickard, R.W., Minka, T.P., 1995. Vision texture for annotation.

Ž .Multimedia Systems 13 1 , 3–14.Sclaroff, S., Taycher, L., LaCascia, M., 1997. ImageRover: a

content-based image browser for the World Wide Web. Proc.IEEE Workshop on Content-Based Access of Image and VideoLibraries, pp. 2–9.

Smith, J.R., Chang, S.-F., 1995. Searching for images and videoŽ .on the World Wide Web. Multimedia Systems 3 1 , 3–14.

Stein, F., Medioni, G., 1992a. Structural indexing: efficient 3-DŽ .object recognition. IEEE Trans. PAMI 14 2 , 125–145.

Stein, F., Medioni, G., 1992b. Structural Indexing: efficient 2-DŽ .object recognition. IEEE Trans. PAMI 14 12 , 1198–1204.

Documents

An environment for content-based image retrieval from large spatial databases