The Perception-Distortion Tradeoff
CVPR 2018 (oral)
Technion - Israel Institute of Technology
Image restoration algorithms can be characterized by their average distortion and by the perceptual quality of the images they produce. We show that there exists a region in the perception-distortion plane which cannot be attained, regardless of the algorithmic scheme. When in proximity of this unattainable region, an algorithm can be potentially improved only in terms of its distortion or in terms of its perceptual quality, one at the expense of the other.
Image restoration algorithms are typically evaluated by some distortion measure (e.g. PSNR, SSIM, IFC, VIF) or by human opinion scores that quantify perceived perceptual quality. In this paper, we prove mathematically that distortion and perceptual quality are at odds with each other. Specifically, we study the optimal probability for correctly discriminating the outputs of an image restoration algorithm from real images. We show that as the mean distortion decreases, this probability must increase (indicating worse perceptual quality). As opposed to the common belief, this result holds true for any distortion measure, and is not only a problem of the PSNR or SSIM criteria. However, as we show experimentally, for some measures it is less severe (e.g. distance between VGG features). We also show that generative-adversarial-nets (GANs) provide a principled way to approach the perception-distortion bound. This constitutes theoretical support to their observed success in low-level vision tasks. Based on our analysis, we propose a new methodology for evaluating image restoration methods, and use it to perform an extensive comparison between recent super-resolution algorithms.
Place your algorithm on the Perception-Distortion plane
out if your algorithm is the closest to the Perception-Distortion bound with
this code, which plots
algorithms on the Perception-Distortion plane. This code can be used to
quantify performance in any image restoration task (e.g. denoising,
deblurring, inpainting, etc.).
For the task of super-resolution 4x, we have pre-computed the image quality scores for 16 recent algorithms and have included them in the code. This allows you to simply add your algorithm to these plots: