A Fuzzy-Soft Competitive Learning Approach for Grayscale Image Compression

In this chapter we develop a fuzzy-set-based vector quantization algorithm for the efficient compression of grayscale still images. In general, vector quantization can be carried out by using crisp-based and fuzzy-based methods. The motivation of the current work is to provide a systematic framework upon which the above two general methodologies can effectively cooperate. The proposed algorithm accomplishes this task through the utilization of two main steps. First, it introduces a specially designed fuzzy neighborhood function to quantify the lateral neuron interaction phenomenon and the degree of the neuron excitation of the standard self-organizing map. Second, it involves a codeword migration strategy, according to which codewords that correspond to small and underutilized clusters are moved to areas that appear high probability to contain large number of training vectors. The proposed methodology is rigorously compared to other relative approaches that exist in the literature. An interesting outcome of the simulation study is that although the proposed algorithm constitutes a fuzzy-based learning mechanism, it finally obtains computational costs that are comparable to crisp-based vector quantization schemes, an issue that can hardly be maintained by the standard fuzzy vector quantizers.