Abstract
Image restoration algorithms are typically evaluated by some distortionmeasure (e.g. PSNR, SSIM) or by human opinion scores that directly quantifyperceived perceptual quality. In this paper, we prove mathematically thatdistortion and perceptual quality are at odds with each other. Specifically, westudy the optimal probability for discriminating the outputs of an imagerestoration algorithm from real images. We show that as the mean distortiondecreases, this probability must increase (indicating lower perceptualquality). Surprisingly, this result holds true for any distortion measure(including advanced criteria). However, as we show experimentally, for somemeasures it is less severe (e.g. distances between VGG features). We also showthat generative-adversarial-nets (GANs) provide a principled way to approachthe perception-distortion bound. This constitutes theoretical support to theirobserved success in low-level vision tasks. Based on our analysis, we propose anew methodology for evaluating image restoration methods, and use it to performan extensive comparison between recent super-resolution algorithms.