Lossless image coding by cellular neural networks with backward error propagation learning

Abstract

This paper proposes a novel hierarchical lossless image coding scheme using cellular neural network (CNN). The coding architecture of proposed method is composed of three steps: split, predict, and entropy coding. The coding performance of proposed method highly depends on that of CNN predictors. The resulting prediction errors are encoded by the adaptive arithmetic coder. To achieve the high coding efficiency, the type of space-variant CNN templates and their parameters are optimized to minimize the actual coding bits of prediction residuals by the minimum coding rate learning with backward error propagation. Experimental results in 21 kinds of standard grayscale test images show that the average coding rates of the proposed scheme is better than that of the conventional schemes.