Text segmentation using linear transforms

Block-based linear transforms have found widespread use in image and video compression. However popular compression algorithms using such transforms, such as JPEG, which are very effective in compressing continuous tone images, do not perform well on mixed-mode images which have a substantial text component. With a growing number of applications where such images occur, e.g., color facsimile, digital libraries and educational videos, there are advantages in being able to classify each block as being text or continuous tone. With such a classification, different compression parameters or even algorithms may be employed for the two kinds of data to obtain high compression with minimal loss in visual quality. In this paper we propose algorithms for text segmentation based on a variety of linear transforms. We analyze the algorithms based on the accuracy and robustness of segmentation. Our results show that any of the popular linear transforms (DCT, DHT, DFT, WHT, DWT) can be used for accurate and robust text segmentation. An important practical implication of our results is that system designers can now use the same transform for both segmentation and compression, thus obtaining substantial savings in computational cost while improving quality.