Cerebellar model articulation controller (CMAC) for sequential images coding

Abstract

CMAC is an artificial neural network that uses a postulate of the cerebellum model as its basic structure. This network has a unique address mapping that provides a condition to learn fast and to store information efficiently. By utilizing the features, this paper implements CMAC for sequential images coding. In the encoding process, the pixel position (row, column, and frame) and the pixel value are used in training as input and output, respectively. The trained weights are then quantized to be the encoded data. In the decoding process, weights, which obtained through de-quantization of the encoded data, are used to reconstruct sequential images. Compression achieved because the bit allocation for weights is smaller than for sequential images. In addition, a frame (or a region of interest in a frame) can be retrieved easily from the encoded data by passing spatio-temporal positions to the output mapping in the decoding stage. This paper also compares the performance between the CMAC-based coding and the block-based coding of MPEG. Results show that the CMAC-based coding increases the performance of the mean square error per frame (factor of 28.1 %), frame rate (factor of 14 %), and perceptual quality (factor of 24.4 %).