IMAGE COMPRESSION USING LEARNED VECTOR QUANTIZATION

This paper presents a study and implementation of still image compression using learned vector quantization. Grey scale, still images are compressed by 16:1 and transmitted at 0.5 bits per pixel, while maintaining a peak signal-to-noise ratio of 30 dB. The vector quantization is learned using Kohonen’s self organizing feature map (SOFM). While not only being representative of the training set, the prototype vectors also serve as a basis for other histogram-similar images. Hence, these codebooks quantize other images not in the training set. Various optimization techniques are investigated. The effects of the uniform, linear, and cubic nested learning rate and neighborhood functions on rate of convergence are studied. Simulated annealing is applied to the SOFM network. By inserting impulses of high temperature at increasing time intervals, codebooks learn more quickly. Competitive learning, frequency sensitive competitive learning, and Kohonen’s neighborhood learning are studied. An XView interface on the SUN SPARC Station 2 is built to facilitate a user interface, and to graphically illustrate the dynamic learning of the codebooks and vector-by-vector quantization and reconstruction of images.