BACKGROUND LEARNING AND LETTER DETECTION USING TEXTURE WITH PRINCIPAL COMPONENT ANALYSIS (PCA)

1
BACKGROUND LEARNING AND LETTER DETECTION USING
TEXTURE WITH PRINCIPAL COMPONENT ANALYSIS (PCA)

• CIS 601 PROJECT
• SUMIT BASU
• FALL 2004

2
OUTLINE

• Introduction
• Background Learning and Letter Detection
• What is Texture?
• PCA based texture representation
• Texture detection approach
• Results
• Analysis and Conclusions
• References

3
INTRODUCTION

• Document image understanding involves
• Layout segmentation
• Logical labeling of blocks at different levels.

• Simplest Text/Non-text separation

• Knowledge of further information such as type
  style
• should be useful in many applications
  e.g.logical
• layout recognition, document image indexing
  and
• retrieval.

• Proposed generic method based on visual
  perception
• TEXTURE

4
What is Texture?

• An Important approach for describing a region is
  to quantify its texture content

• Texture can be defined as that where there is a
  significant variation in intensity levels between
  nearby pixels that is, at the limit of
  resolution, there is non-homogeneity.

5
APPLICATIONS OF TEXTURE

• Simplest use
• Physical segmentation by classification
  of blocks using 2 or 3 classes (text/non-text,
  text/image/line drawing) by simple features like
  black/white transitions

• Further analysis of document structuring
  Characterizing of fonts (using geometrical
  properties, statistical features or generic
  techniques like Feature Based Interaction Maps
  (FBIM)), skew-detection.

6
APPLICATIONS OF TEXTURE ( contd)

• Application specific to this project
• Background learning and Letter detection using
  Texture with Principal Component Analysis
  (PCA).
• This document analysis is a necessary
  pre-processing stage for many document-processing
  systems such as
• Optical Character Recognition. (OCR)
• Document Retrieval.
• Document Compression.

7
PRINCIPAL COMPONENT ANALYSIS (PCA)

• Technique capable of deriving low dimensional
  representation which is applied extensively to
  identify texture of images.

• Involves a mathematical procedure that transforms
  a number of possibly correlated variables into a
  smaller number of uncorrelated variables called
  principal components.

8
PRINCIPAL COMPONENT ANALYSIS (PCA) (contd)

• The first principal component accounts for as
  much of the variability in the data as possible,
  and each succeeding component accounts for as
  much of the remaining variability as possible.

• Since images are array of data points with each
  point representing color, PCA can be used for
  reducing the image data (extracting features) to
  smaller dimension to represent the image
  qualities. remaining variability as possible.

• The reduced feature represents the spatial
  distribution of the pixel gray values.

9
PRINCIPAL COMPONENT ANALYSIS (PCA) (contd)

• PCA (Principal Component Analysis)
• Project the samples (points) perpendicularly onto
  the axis of ellipsoid
• Rotates the ellipsoid to be parallel to the
  coordinate axes
• Use the fewer and more important coordinates to
  represent the original samples

• Transforms of PCA
• The first few eigenvectors of the covariance
  matrix

10
BACKGROUND LEARNING FOR LETTER-DETECTION

• Given a document image we first convert it to a
  gray level image. Since we are working with local
  texture representation only, this is not going to
  effect the processing of the image.

• Then we divide the document image into sub-images
  where all sub-images are non-overlapping blocks
  of a specific size (We intend to use height 32
  and width 32 pixels)

• Normalize each sub-image independent of the other
  sub-images by subtracting the mean of the
  sub-image from each pixel.

11
BACKGROUND LEARNING FOR LETTER-DETECTION(contd)

• Normalizing would help in getting rid of any
  deviation that a specific sub-image might have
  from the other sub-images, for instance
  difference in brightness.

• We then wish to use the sub-images to compute the
  principal components using PCA.

• Use first few principal components to obtain a
  projection matrix to project each sub-image to an
  n-dimensional vector that constitutes its texture
  representation. The number of principal
  components to be used would be decided on an
  image-to-image basis.

12
BACKGROUND LEARNING FOR LETTER-DETECTION(contd)

• We now project all sub-images to their texture
  representation as n-dimensional vectors.

• Now we use this background learning to exclude
  background sub-images from further image
  processing.

• The remaining sub-images are the informative
  ones. We can now use the remaining sub-images for
  letter detection.

• In order to exclude background sub-images, we
  use k-means clustering on the n-dimensional
  vectors corresponding to the sub-images.

13
BACKGROUND LEARNING FOR LETTER-DETECTION(contd)

• We approximate k by observing the resultant image
  from PCA and vary k by trial and error method.

• The cluster corresponding to the maximum number
  of sub-images represents the background. By
  removing these sub-images, we would be reducing
  the background and thus reduce total scan area
  for OCR software.

• We developed MATLAB programs to do the
  above-mentioned processing and then use it on
  several document images to compare the
  performance of this procedure and try to further
  improve it.

14
RESULTS
?Original Document
?After PCA with sub-image size 16 and 25
first PCA components
15
RESULTS
?Sub-image size 16 using all principal
components
?Sub-image size 32 using all principal
components
16
RESULTS
?k-means with k 4
?k-means with k 8
17
RESULTS
?k-means with k 12
?k-means with k 15
18
RESULTS
Text Image after removing background
19
RESULTS
?Original Document
?After PCA with sub-image size 32 and 80
first PCA components
20
RESULTS
?Sub-image size 16 using all principal
components
?Sub-image size 32 using all principal
components
21
RESULTS
?k-means with k 6
?k-means with k 12
22
RESULTS
?k-means with k 18
?k-means with k 30
23
RESULTS
Text Image after removing background
24
RESULTS
?Original Document
?After PCA with sub-image size 16
25
RESULTS
?k-means with k 6
?k-means with k 30
26
RESULTS
Text Image after removing background
27
RESULTS
?Original Document
?After PCA with sub-image size 16
28
RESULTS
?k-means with k 6
?k-means with k 30
29
RESULTS
Text Image after removing background
30
RESULTS
?Original Document
?After PCA with sub-image size 16
31
RESULTS
Text Image after removing background
32
ANALYSIS CONCLUSION

• We tried with sub-images of different sizes 16,
  32, 64 etc. Initially we were under the
  impression that smaller sub-images would perform
  better but take more time to execute.
• We figured out that is not necessarily true and
  that depends on the image.

33
ANALYSIS CONCLUSION..

• PCA seemed to be pretty successful in identifying
  the text blocks in the images. In most of the
  images we used, we got a pretty good success rate
  using the sub-images as the training set.
• The cluster corresponding to the maximum number
  of sub-images was the background in all cases.

34
ANALYSIS CONCLUSION

• More number of clusters doesnt necessarily
  produces more text. Some text which was visible
  with less number of clusters, wasnt visible with
  more.
• However, more number of clusters reduced removed
  background.
• There seems to be a trade off and an optimal
  cluster size specific to each image.

35
ANALYSIS CONCLUSION

• Removing the background by replacing the cluster
  corresponding to the maximum number of sub-images
  seems to be a pretty good method of reducing
  space to be scanned by OCR.
• The number of clusters to be used is very much
  image dependent. The image produced by the PCA
  gives some idea.

36
ANALYSIS CONCLUSION

• In all our images PCA followed by clustering was
  successful in removing some background space.
• It also seem to do a pretty good work of image
  detection.
• We conclude that this method works and could be
  used as a tool to reduce space to be scanned by
  OCR.

37
REFERENCES

• Image Retrieval Using Local PCA Texture
  Representation by Longin Jan Latecki, Venugopal
  Rajagopal, Ari Gross
• Gonzalez, Woods, Eddins. Digital Image Processing
  Using MATLAB
• Web material and Notes.