New intra prediction methods based on channel correlation and their combination for color image compression

This paper proposes two new intra prediction methods for color image compression, which exploit the correlation between the base and non-base colors. The first method predicts a pixel in non-base color as a weighted sum of neighboring pixels, where the weights are determined as a function of photometric distances from the pixel to the neighboring ones in the base color. The second exploits the linear relationship between the channels, where the parameters for the model are found from the already reconstructed base color, so that no side information is needed. In the implementation, it is also found that the weighted combination of new predictors gives better coding gain, where one of predictors is chosen based on the RD optimization. The experimental results show that the proposed method provides higher coding gain than the conventional H.264/AVC, up to 1.2dB for luminance and 2.3dB for the chrominance image.