This research is exploring the development of a very different approach to image coding based on representing an image by a large collection of statistics. The approach is motivated by the recognition that state-of-the-art wavelet image coders owe their superior coding performance to more accurate statistical modeling of coefficients. The accuracy of these models is limited by the cost in bitrate of transmitting the statistics needed to support the modeling. Our approach makes more explicit use of model statistics to allow direct representation of the image with those statistics. The statistics serve both to maximize the accuracy of the statistical modeling and to represent the image itself. Consequently, many more statistics can afford to be transmitted and modeling accuracy is greatly improved. The approach has shown very promising preliminary results, and promises to substantially advance the state-of-the-art in image coding. The approach also suggests new opportunities for managing the perceptual quality of coding algorithms through perceptually optimized quantization of the new representation. These perceptual issues will also be investigated.
