Fuzzy neural network based prediction coding for bayer pattern image

In this paper, a sequential lossless compression of raw data from image sensor with Bayer pattern is proposed. Inspired by model of JPEG-LS, the proposed encoder consists of fuzzy neural network predictor, adaptive correction part based on context and adaptive arithmetic coder. As in JPEG-LS, it is a empirically observed that the global statistics of residuals from the ANN fixed predictor in raw data, which effectively exploits structural redundancies between mosaic-like color components, are well-modeled by a TSGD centered at zero. In the meantime, we propose a context determination approach based on causal interpolation that achieves high coding efficiency. Consequently, we can encode mosaic images on the fly at low complexity level. Compared with existing methods of lossless compression for Bayer raw data, the performance of proposed method is apparently the best.