Upload
sydney-conley
View
214
Download
0
Tags:
Embed Size (px)
Citation preview
Neural Network Applications Neural Network Applications in OCRin OCR
Daniel Hentschel
Robert JohnstonCenter for Imaging Science
Rochester Institute of Technology
BackgroundBackground
Recently, many Ancient documents have been discovered. Deciphering these documents is integral to understanding our past and predicting our future as a human race.
BackgroundBackground
Unfortunately, the majority of ancient Hebrew documents found are seriously degraded.
BackgroundBackground
This degradation makes deciphering of the documents difficult. Many techniques have been developed to enhance the text in degraded manuscripts in order to improve readability.
The majority of this work has been carried out by Roger Easton and Robert Johnston at the Center for Imaging Science at RIT.
Note:
Question:Question:
Would it be beneficial to use a neural network to help in the deciphering of degraded ancient Hebrew documents?
Research PurposeResearch PurposeCreate an application to assist in
document restoration.Develop a neural network adept
at Hebrew OCR (optical character recognition).
Analyze the functionality of the network when studying degraded characters
Creating an ApplicationCreating an Application
Developed with Visual BasicDeveloped with Visual Basic
Calculate Input FeaturesCalculate Input FeaturesSegmentation.Horizontal and Vertical section averages.Length of the skeleton as a percentage of
the circumscribed rectangle perimeter.Complexity: Square root of the black area
divided by the length of the skeleton.Number of dead ends and intersections in
the skeleton of the character.
SegmentationSegmentation
Simple segmentation. Not elegant, but functional.
Thinning AlgorithmThinning Algorithm
Adapted from “A Fast Thinning Algorithm For Characters” (Flores, Rezende, Carrijo, Yabu-tti)
Red - pixel being analyzed
Green/Blue - to be deleted
Workings of a Neural NetworkWorkings of a Neural Network
Input LayerHidden Layer
Output Layer
Nodes in a Neural NetworkNodes in a Neural Network
Inputs are multiplied by a weighting factor and a bias is added.
Weighted inputs are summed.Sum is applied to a function:
11
2Output
Input Weighted
e
Internal Function of a NodeInternal Function of a NodeNode Function
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
-10 -8 -6 -4 -2 0 2 4 6 8 10
Input
Ou
tpu
t
Operation of a NodeOperation of a Node
Input to the Neural NetworkInput to the Neural Network9 input nodes
– Horizontal (3) and Vertical (2) histogram.– Length of the skeleton as a percentage of the
circumscribed rectangle perimeter.– Complexity: Square root of the black area divided
by the length of the skeleton.– Number of changes of intersections, and dead
ends encountered in the skeleton.
Output of the Neural NetworkOutput of the Neural Network
4 output nodes
Hidden NodesHidden Nodes
3 hidden nodes works well for 4 characters
ResultsResults
Training set:
ResultsResults
Characters not in training set:.07
.00
.94
.04
.05
.05
.92
.05
.08
.04
.03
.94
.07
.95
.02
.02
.86
.05
.05
.02
.08
.04
.04
.92
.12
.77
.01
.13
ConclusionConclusion
The neural network is quite effective at deciphering non-degraded text.
Not enough degraded characters have been studied yet to determine how well the network will perform
Future work:– More than 4 characters– Optimize inputs– Analyze degraded characters
The EndThe End