VECTOR QUANTIZATION OF IMAGES BASED ON A COMPOSITE SOURCE MODEL (CODING, DATA COMPRESSION, PICTURE)

Compression of digitized pictures is finding increasing application in teleconferencing, archiving, and remote sensing. In this study, we have investigated the potential capability of Vector Quantization, a relatively new source coding technique, for the compression of still, monochromatic images. A vector quantizer operates on blocks (vectors) of contiguous samples of the input signal--in our case, an image. As the block size increases, the performance of vector quantization approaches the best that is theoretically possible, but at the cost of an exponential growth in computational complexity.

We have developed a coder based on vector quantization, with moderate block sizes, which performs better than existing coders of comparable complexity. Since blocks of samples, rather than individual ones, are treated as atomic entities in vector quantization, we have developed a new vector model for images called the Composite Source Model. Each block is viewed as being the output of one of a bank of subsources selected by a switch. Each subsource generates blocks of a distinct perceptual type, e.g., blocks with an edge at a particular orientation.

Based on the new model, we have proposed a new coding method called Classified Vector Quantization. In this method, each block of samples in an image is classified to determine which subsource or class it belongs to. A different vector quantizer, tailored to code blocks of the appropriate perceptual type, is employed for each class. Important perceptual features of images, such as edges, are emphasized in this manner. Good visual quality has been achieved with compression ratios in the range of 10:1. We have also developed a post-processing algorithm which improves image quality without increasing the bitrate.

We have derived a new result for the Rate-Distortion function of the composite source model. We have proposed a new statistical composite source model and demonstrated that it generates blocks of samples very similar to blocks from real images. Finally, we have shown that the performance of Classified Vector Quantization is within 0.3 bits per sample of the Rate-Distortion function for the new model.