OCR, MRC & JPEG2000

One of the novel aspects to JPEG2000 is that, although an algorithm for segmenting color images is not specified, the JPEG2000 spec does support segmentation-based coding. In fact, the most effective rates for color compression are obtained by analyzing, understanding, and reversing the page layout process. It becomes important in JP2 segmentation-based coding to be able to separate foreground from background, and more specifically, text regions from non-text regions.

Unlike many OCR systems, which can "tolerate" missing text regions that are not aligned horizontally or vertically, color compression using Mixed Raster Content (MRC) coding or based on JPEG2000 part 6, is much less forgiving. The basis of MRC coding is separating out the high frequency signal information from the low frequency information. Usually, the high frequency information in an image is text-related. Of course, there is also line art, edge structures, and other objects that may degrade when kept at low resolution. But text regions must be recognized and lifted for MRC-based compression to be non-degrading. This necessitates finding all text regions, regardless of skew, rotation, etc.

Accurate detection of all text regions is very helpful in improving OCR accuracy. In this way, a system that supports MRC-based or JPEG2000 compression coding (like CVISION) will benefit with respect to improved OCR as well. The detected text regions can be fed directly to the OCR engine to be sure that no image text is left unsearchable. Accurate OCR and perceptually lossless compression for color images both rely on robust, precise page segmentation of foreground text from background image. This segmentation leads to the best color compression rates and the most reliable OCR.

Click here to read next topic: Reverse Video & OCR

Return to Table of Content