Vector Quantization with codebook and index compression

This paper presents a hybrid (lossless and lossy) technique for image vector quantization. The codebook is generated in two steps. 1. The training set is sorted based on the magnitudes of the training vectors. 2. From the sorted list, training vector from every n th position is selected to form the codevectors. Followed by that, centroid computation with clustering is done by repeated iterations to improve the optimality of the codebook. The codebook thus generated is compressed (lossy) to reduce the memory needed to store the codebook with the slight degradation in the quality of the reconstructed image. The bits needed to store the indices of the codevectors are further reduced to improve the compression rate without introducing any extra coding distortion (lossless) using the concept of Search Order Coding (SOC). The technique is tried with different codebooks of sizes 16, 32, 64, 128, 256, 512, and 1024 and the results are compared in terms of bits taken per pixel and PSNR (quality of the reconstructed image) values. The computational complexity of the method is quite low and can be easily implemented in the hardware.