Self-organizing Finite State Vector Quantization for Image Coding

This paper presents a new method in finite state vector quantization (FSVQ) for image coding by using self-organizing neural networks to automatically design the state codebook and next-state function without the need of matching the local statistics to the global statistics. It dynamically predicts the most probable codevectors for the current input as state codebook based on the distance information of codevectors used to encode the neighboring blocks. Our experiments show that, compared to memoryless VQ, the SNR improvement is up to 4.2 dB at the same rate, while the bit rate reduction is more than two times at the same SNR level. In addition, the proposed method is simple for hardware implementation.