The Effect of Perceptual Loss for Video Super-Resolution

Abstract

Intensity-based loss which measures pixel-wise difference is commonly used for most of the learning-based super-resolution approaches. Since the error of different components has disparate impacts on human visual system, the structural error which calculates the error of the perceptually influential components is proposed for the loss function of the video super-resolution. The proposed loss function is presented based on the JPEG compression algorithm and the effect of using quantization matrix on resultant output. The proposed loss function can be employed instead of the traditional MSE loss function. In this paper, we explored the effect of using this perceptual loss for VESPCN method. The experimental results illustrate better outputs in terms of average PSNR, average SSIM, and VQM.