Fast and High Efficient Color Image Compression Using Machine Learning

Traditionally, the previous learning based color image compression mechanism involves storing the luma component and some color pixel seeds selected by active learning. The decompression is basically a process of prediction for the rest color using a regression model trained by a semi-supervised learning algorithm. In this paper, we propose an algorithm with an improved objective function such that the regression model is trained by supervised learning to minimize the total prediction error and thus avoid the computation burden of the color seeds selection. Besides, the kernel principal component analysis map adopted to reduce the computation complexity is carried out on the same given set at both the encoder and decoder to ensure that predictions are the same at both ends without signaling extra information. In the end, the luma information and parameters of the regression model are transmitted to the decoder for color prediction and then the color image is reconstructed. The experimental results show the performance superiority of the proposed algorithm to the previous ones.