Saliency guided adaptive residue pre-processing for perceptually based video compression

In this paper we investigate a method for selectively modifying a video stream using a contrast sensitivity model based on the human visual system. This paper builds upon earlier work (Shaw, et.al., 2012) by expanding the algorithm to the Chroma channels and incorporating a saliency model to dynamically select between the different perceptual quantization levels. Our new algorithm selectively chooses the perceptual compression level based on the visual importance of the region as determined by the saliency model. The method has been implemented in the H.264 JM 18 reference encoder, and has been shown to achieve an average of 37% improvement in data compression over the baseline method without perceptible degradation of the video quality.