Goal of Text-Image Segmentation

• input is the raw image (grayscale, binary)
• eventually...
  
• want a segmentation into text blocks, non-text blocks, and background blocks
• however...
  
• there isn't enough information available at this early processing stage to make the correct decision
• approach
• low-level text/image segmentation outputs, for each pixel, the probability whether it is text, non-text, or background
• this information is then passed on to the layout analysis engine

Design Criteria

• what units do we classify?
• connected components, individual pixels, patches, ...
• what features do we extract?
• texture features, geometric features, ...
  
• what classifier do we use?
• MLP, decision tree, ...
• how do we integrate the local decisions?
• leave it up to subsequent processing stages, already perform some integration, ...

Current Implementations

• Leptonica's code
• appears to be rule based, extensively tuned
• OCRopus block classifier
• use whitespace-based segmentation of the image into blocks
• for each block, extract a number of features
• classify using logistic regression
  

ITextImageSegmentation

image = bytearray()
binarized = bytearray()
result = intarray()

binarizer = make_BinarizeBySauvola()
ticlass = make_TextImageSegByLogReg() -- make_TextImageSegByLeptonica()

read_image_gray(image,arg[1])
binarizer:binarize(binarized,image)
ticlass:textImageProbabilities(result,binarized)
write_image_packed(arg[2],result)

Text/Image Segmentation

Text-Image Segmentation in OCRopus

• Note:
• the methods don't work very well yet
• the methods are resolution dependent
• integration of the output into the rest of the layout analysis method is not complete yet
• fortunately... they aren't all that important because layout analysis implicitly gives good text/image segmentation
• To be done:
• we're implementing several additional methods
• consider contributing your own methods