Segmenting for wavelet compression

Summary form only given. We propose a new approach to segmenting mixed raster content images for compression by wavelet encoders and binary compressors. Though wavelets are efficient at modeling edges in natural images, text and sharp edges can be more effectively represented by a binary mask than with wavelets. A common approach to mixed raster content documents is to segment images from text, and separately compress images and a binary mask that defines the segmentation. We propose that the segmentation be done jointly with a projection-onto-convex-sets (POCS) smoothing, with the design goal of segmented images that have a minimal number of nonzero wavelet coefficients. Our goal is a document segmentation that is easy to compress. Knowing that the image encoder will be a lossy wavelet coder, we approximate the above goal as minimizing the number of nonzero wavelet coefficients. We attempt to achieve this goal by jointly choosing the mask and data-filling.