Edge priors guided deep unrolling network for single image super-resolution

Abstract

The field of single image super-resolution (SISR) has garnered significant consideration over the past few decades. The primary challenge of SISR lies in restoring the high-frequency details in low-resolution (LR) images, which are crucial for human perception. Current deep learning-based SISR methods either increase the depth of the network or indiscriminately incorporate emerging technologies, such as attention mechanisms, to address this challenge. However, these methods treat the deep networks as a black-box, achieving performance at the cost of efficiency and network redundancy, without carefully considering how internal components interact to enhance the reconstruction quality. To address this limitation, we incorporate edge priors into a classical image restoration model to design network framework and propose an edge priors guided deep unrolling network (EPGDUN), which consists of three components: Edge Feature Extraction Module (EFEM), Intermediate Variable Updating Module (IVUM), and Variable-Guided Reconstruction Module (VGRM). Specifically, we unroll the image restoration model with edge priors via a half-quadratic splitting and proximal gradient descent method to gain three subproblems, whose solving process corresponds to the iterative stages of the three submodules of EPGDUN. The combination of edge priors constrains the network output, strengthening the image boundaries, enhancing interpretability, and improving the network’s understanding of image structure. Extensive experiments illustrate that EPGDUN achieves performance on par with or exceeding that of the state-of-the-arts, uninterpretable black-box neural models and interpretable deep unrolling networks. These findings underscore EPGDUN’s potential to advance low-level vision applications and other domains requiring mathematical interpretability. The source code for EPGDUN will be available at https://github.com/songhp/EPGDUN.