Boundary equilibrium SR: effective loss functions for single image super-resolution

Recently, single image super-resolution (SISR) has made great progress due to the rapid development of deep convolutional neural networks (CNN), and the application of Generative Adversarial Networks (GAN ) has made super-resolution networks even more effective. However, GAN-based methods have many problems such as lengthy and unstable convergence. To solve these problems, this paper presents a mechanism that employs boundary equilibrium in the image super-resolution network to balance the convergence of the generator and the discriminator and improve the visual quality of the generated synthetic images. Furthermore, current methods often use perceptual loss based on the VGG network. However, experiments show that the visual quality improvement brought by this perceptual loss is very limited, so we propose an improved perceptual loss based on Learned Perceptual Image Patch Similarity (LPIPS) to acquire better human visual effects rather than adopting the traditional perceptual loss based on VGG. The experimental results clearly show that using our proposed method can considerably improve the performance of image super-resolution and obtain clearer details than state-of-the-arts.

Automated summary

Recently, significant progress has been made in single image super-resolution (SISR) with the use of deep convolutional neural networks (CNN) and Generative Adversarial Networks (GAN). However, GAN-based methods suffer from lengthy and unstable convergence. To address these issues, this paper proposes a mechanism that incorporates boundary equilibrium in the image super-resolution network, allowing for balanced convergence of the generator and discriminator and improved visual quality of generated images. Additionally, the paper introduces an improved perceptual loss based on Learned Perceptual Image Patch Similarity (LPIPS), which outperforms traditional VGG-based perceptual loss in terms of acquiring better human visual effects. Experimental results demonstrate that the proposed method significantly enhances image super-resolution performance and achieves clearer details compared to state-of-the-art methods.