Finite state vector quantisation with neural network classification of states

The results of employing neural network classification of states (NNCS) in finite state vector quantisation (FSVQ) are presented. In addition to intrablock correlation, already exploited by vector quantisation (VQ), the new design takes advantage of the interblock spatial correlation in typical grey-level images. The main achievement of FSVQ techniques is to assure access to a large master codebook for quantising purposes, thus achieving high image quality, while utilising a small state codebook for the purpose of specifying the block label, thus utilising low bit rates. Typically, FSVQ techniques require a very large memory space for the storage of the numerous state codebooks. However, with NNCS the memory space requirements can be reduced by a large factor (about 102–103) to manageable size, with little or no impairment of image quality, in comparison to FSVQ. This is accomplished by a neural network classification of finite states into representation states, whose associated states all share the same codebook. This codebook is populated by the most frequently occurring codevectors in the representation state. Numerical and pictorial results of simulation experiments are presented for the image LENA. They show that by using NNCS the required bit rate is about 0.25 bits/pixel at 30 dB peak SNR resulting in high quality reconstructed imagery, while the memory requirement is reduced by a factor of 256.