Presentation is loading. Please wait.

Presentation is loading. Please wait.

1/20 Document Segmentation for Image Compression 27/10/2005 Emma Jonasson Supervisor: Dr. Peter Tischer.

Similar presentations


Presentation on theme: "1/20 Document Segmentation for Image Compression 27/10/2005 Emma Jonasson Supervisor: Dr. Peter Tischer."— Presentation transcript:

1 1/20 Document Segmentation for Image Compression 27/10/2005 Emma Jonasson Supervisor: Dr. Peter Tischer

2 2/20 The Big Picture Background Aims Suggested Solution Tests and Results Conclusion Future Work Overview

3 3/20 The Big Picture Find structure in data Data = binary images of documents Documents contain: – Text – Pictures – Diagrams – Etc.

4 4/20 Why structure in data? Content-Based Image Retrieval (CBIR) –E.g. find documents with photos Optical Character Recognition (OCR) –Extract text from documents JBIG 2 (Joint Bi-level Image experts Group) and MRC (Mixed Raster Content) –Use of structure can improve compression

5 5/20 What’s out there? JBIG 2 –Encoder tells decoder what the segmentation is –No standard segmentation algorithm DjVu –Intended for colour images –Different layers = different segments

6 6/20 Discover structure in binary images, to: – Enhance compression – Explore the influence of different kinds of segmentation – Explore novel approach to segmentation pre-processing Project Aims

7 7/20 Solution Claim: –Segments are groups of pixels which have similar information content Approach: –Extract information content from image and group according to similarity of content

8 8/20 Pipeline of Solution Whitening transformation BlockingSegmentation Compression

9 9/20 Whitening transformation Operates on “contexts” Reversible Black pixels = Incorrect predictions White pixels = Correct predictions

10 10/20 Blocking Group pixels into n x n blocks Rough estimate of information content Density of black pixels in transformed image indicates different information content.

11 11/20 Segmentation of blocked, whitened image Segmentation techniques: – Manual By hand – Automatic Thresholding Clustering Outcome: – Segment map

12 12/20 Segmentation-Based Coding Encoding of segment information – Compress current pixel based on neighbouring pixels Encoding of binary image given segmentation – Compress current pixel based on neighbouring pixels and type of segment

13 13/20 Segmentation Issues Trade-off between: –Information needed to encode segmentation –Information needed to encode image given segmentation I.e. how accurate should the segmentation be?

14 14/20 Evaluation of Segmentation Subjective measure –Is segmentation optimal from a human being’s point of view? Objective measure –Code length of compressed image

15 15/20 Testing Seven “representative” test images Various whitening contexts Various block sizes Different segmentation techniques Different segmentation granularity

16 16/20 Results No single “best” whitening context for all images –Unwhitened generally worse Optimal block size is image-dependent –8 x 8 and 16 x 16 generally perform well No single “best” granularity for all images

17 17/20 Results Compression ratio

18 18/20 Conclusion No single optimal segmentation Better than JBIG 1 Worse than JBIG 2 –Only compared with 1 image due to patent restrictions –Better textual compression needed to compete with JBIG 2

19 19/20 Future Work Test other segmentation techniques Semi-automatic segmentation MML clustering Implement “connected black regions”- compression used in JBIG 2 Represent segmentation in a different format

20 20/20 Any Questions?


Download ppt "1/20 Document Segmentation for Image Compression 27/10/2005 Emma Jonasson Supervisor: Dr. Peter Tischer."

Similar presentations


Ads by Google