Improving Image Decomposition Method of the 3-MRC Coding of Scanned Compound Document Images

This paper discusses image decomposition problem of the 3-layer MRC model based coding of scanned (noisy) document images. A widely-used approach for document decomposition is to divide the document image into blocks and split the pixel histogram of each block into two halves by minimizing the sum of variance of its pixels with the mean of the halves. We propose to split a block by minimizing the variance of one half with its minimum pixel and the variance of the other half with its maximum pixel. Our goal is to increase the gap between the two halves by avoiding splitting of any cluster of pixels into both halves. It should help reduce complexity of the generated mask. Moreover, we do not decompose a block if it has no edge points, again to reduce the mask complexity. We also implement a noise reduction heuristic in the mask layer to correct placement of transition pixels. We provide simple analysis and evaluate block energy in terms of the DCT coefficients of the resulting FG/BG layer blocks. Experimental results show that code size of the mask layer of our test images, obtained using proposed processing is reduced to nearly half of the mask obtained by a straight-forward 3-MRC implementation.