Texture Content Based Successive Approximations for Image Compression and Recognition

The representation of an image as a collection of "words", also called "textons" or "primitives", from a predefined vocabulary has become an essential step in developing advanced computer vision algorithms. Such vocabularies are often derived via some form of unsupervised or supervised machine learning technique, and by arbitrarily choosing an optimal set of words, often referred to as a dictionary. Even though such representations have been shown to capture the 2D image content, textons or image descriptors were never used for image compression or for recognizing image content in the compressed domain. In this work we are presenting a novel texture based image representation technique that can be used for both image compression and recognition purposes. The developed algorithm requires no filter, no post processing (for image enhancements) and serves as a viable tool for image compression when compared to standards such as JPEG and JPEG2000. It is also a powerful tool for recognizing content in the compressed domain. Performance of the developed image compression descriptor is bench-marked for object recognition applications in the compressed domain using Caltech-101.