Rate-Distortion-Perception Tradeoff Based on the Conditional Perception Measure

In this paper, we study the rate-distortion-perception tradeoff generalizing the classical rate-distortion theory by adding a perception constraint to generate visually pleasing reconstructions. The perception metric measures the divergence between the distributions of the input and the reconstruction when both distributions are conditioned on the encoder's output. This metric, originally introduced by Mentzer et al. for the video compression setting, is called as conditional perception measure. We characterize the rate-distortion-perception tradeoff for a general source. In the Gaussian setting, we show that jointly Gaussian reconstructions are indeed optimal. Interestingly, to achieve a perceptually perfect reconstruction, comparing to the minimum mean square error (MMSE) reconstruction, we only need extra 0.5 bits/sample for the compression rate.