Kernel-Guided Texture-Aware Blind Super-Resolution via Uncertainty Learning

The accuracy of kernel estimation is critical to the performance of blind super-resolution (SR). Traditional kernel estimation methods usually use L1 or L2 loss function to minimize the difference between the estimated kernel and the ground-truth kernel. This tends to make the estimated kernel converge towards the average of all possible kernels, deviating from the ground-truth kernel. To improve the performance of kernel estimation, this paper proposes an uncertainty loss for training a kernel estimation network, focusing on regions with high uncertainty (variance) in the kernel. In addition, a texture-aware SR network is proposed that utilizes the Gumbel Softmax trick to pay more attention to the complex regions of the image texture, thus improving the SR performance. Extensive experiments on synthetic datasets show that our approach achieves promising performance.

I. Introduction

Super-resolution (SR) constitutes a pivotal challenge within the realm of computer vision, aiming to reconstruct high-resolution (HR) images from their corresponding low-resolution (LR) counterparts. In recent years, with the rapid development of deep learning, convolutional neural network (CNN) has shown strong learning ability, and solving SR problems through CNN has become popular. Classical SR models assume that the LR image is obtained by down-sampling the corresponding HR image with a predetermined blur kernel, such as the Bicubic downsampling kernel [1]–[3]. However, real-world degradation processes tend to be excessively intricate to be accurately encapsulated by a single fixed degradation model. As a result, these approaches often exhibit diminished efficacy in SR tasks characterized by an unknown degradation type. To address this intricate predicament, blind SR methods have emerged.