Image coding via bintree segmentation and texture VQ

Abstract

Image compression is often approached from an angle of statistical image classification. For instance, VQ-based image coding methods compress image data by classifying image blocks into representative two-dimensional patterns (codewords) that statistically approximate the original data. Another image compression approach that naturally relates to image classification is segmentation-based image coding (SIC). In SIC, we classify pixels into segments of certain uniformity or similarity, and then encode the segmentation geometry and the attributes of the segments. Image segmentation in SIC has to meet some more stringent requirements than in other applications such as computer vision and pattern recognition. An efficient SIC coder has to strike a good balance between accurate semantics and succinct syntax of the segmentation. From a pure classification point of view, free form segmentation by relaxation, region-growing, or split-and-merge techniques offers an accurate boundary representation. But the resulting segmentation geometry is often too complex to have a compact description, defeating the purpose of image compression. Instead, we adopt a bintree-structured segmentation scheme. The bintree is a binary tree created by recursive rectilinear bipartition of an image.