Fine-Grained Dynamic Loss for Accurate Single-Image Super-Resolution

Abstract

With a wide range of applications, single-image super-resolution (SISR) is a very important computer vision task. There are many SISR strategies which are based on the CNNs improvement, such as residual connections, deeper networks, and attention mechanisms. Although such kind of strategies often improve the SISR performance, they increase CNNs’ computational complexity and complex image texture areas still remain hard to be reconstructed. In particular, all the existing loss functions are minimized in the whole dynamic range, e.g. 0-255 in case of 8 bit image, which causes difficulty of CNNs’ learning in texture image regions above all. Developing new loss function provides a promising SISR solution, i.e. one should be beyond the existing regression loss functions, which encounter problem in reconstructing the image texture details. For such goal, this paper proposes a dynamic fine-grained loss function. It consists of both an existing regression loss function and a new pixel classification loss together with a dynamic regression range loss. Extensive experiments conducted on the benchmark SISR show that the proposed method can achieve better results and without extra computational cost.