Perceptually Optimized Loss Function for Image Super-Resolution

Most of the learning based single image super-resolution networks employ intensity loss which measures pixel-wise difference between the estimated high resolution image and the ground truth. Since image components are different with respect to their saliency for HVS, it is desired to weight their impact on the loss functions accordingly. In this paper, a simple perceptual loss function is introduced based on the JPEG compression algorithm. In fact, the two compared images are transformed into DCT domain and then divided by the weighted quantization matrix. The difference between the resultant DCT coefficients shows the most effective components for HVS and can be considered as a perceptual loss function. The experimental results illustrate that employing the proposed loss promotes the convergence speed, and also, provides better outputs in terms of qualitative and quantitative measures.

I. Introduction

Single image super-resolution (SISR) aims to construct a high-resolution (HR) image using a low-resolution (LR) input. With the advent of technology, super-resolution is considered as a low-cost and practical way to increase the image resolution by improving the details of a signal. The results of super-resolution can be applied in a wide variety of applications, including surveillance [1], remote sensing [2], medical diagnosis [3], [4] and text image [5] or face super-resolution [6]. This has led to the increase in research for better SISR methods.