Improving neural network approach to lossless image coding

In the paper optimization of lossless image coders based on Adaptive Neural Networks (AdNN) is addressed. Firstly, a detailed analysis of the influence of AdNN parameters on coder performance (average bitrate, time complexity) is done. Secondly, an improved technique denoted AdNN₊ is proposed. Its main features are introduction of contexts, variable training window size, and post-processing by NLMS algorithm. Experiments show that indeed, the new method is better than others based on neural networks, and that it can even compete with the best existing image lossless coding algorithms.