Adaptive image quantization based on learning classifier systems

Summary form only given. The performance of a quantizer depends primarily on the selection of a codebook. Most of the quantization techniques used in the past are based on a static codebook which stays unchanged for the entire input. As already demonstrated successfully in lossless data compression, adaptation can be very beneficial in the compression of typically changing input data. Adaptive quantization has been difficult to accomplish because of its lossy nature. We present a model for distribution-free adaptive image quantization based on learning classifier systems which have been used successfully in machine learning. A basic learning classifier system is a special type of message-processing, rule-based system that produces output according to its input environment. Probabilistic learning mechanisms are used to dynamically direct the behavior of the system to adapt to its environment. The adaptiveness of a learning classifier system seems very appropriate for the quantization problem. A learning classifier system based adaptive quantizer consists of the input data, a codebook, and the output. When an input can not be matched, a new codebook entry is constructed to match the input. Such an algorithm allows us not only to deal with the changing environment, but also to control the quality of the quantized output. The adaptive quantizers presented can be applied to both scalar quantization and vector quantization. Experimental results for each case in image quantization are very promising.