CA-ESRGAN: Super-Resolution Image Synthesis Using Channel Attention-Based ESRGAN

Abstract

Advancements in deep learning algorithms and high-performance GPU computation have spurred extensive research in image processing. This study focuses on the synthesis of super-resolution images, a technique aimed at generating high-resolution images from low-resolution images that contain various types of degradation and noise. In recent years, the emergence of convolutional neural networks and adversarial generative networks has made it possible to generate higher quality images. In this research, we propose two approaches to enhance the image quality based on a mainstream algorithm ESRGAN, a GAN-based generator and discriminator model architecture. The first approach applies two types of channel attention (SENet and ECA-Net) to the generator. SENet explicitly models the interdependence of convolutional features across channels, thereby enhancing the quality of the representation generated by the network. ECA-Net decomposes the channel attention of SENet to reduce model complexity while preserving performance through appropriate cross-channel interactions. The second approach applies LPIPS to the image evaluation methods of the discriminator. LPIPS serving as an image quality assessment metric, enhances perceptual evaluation by combining feature extraction from a pre-trained neural network using human evaluation data. To assess the effectiveness of our proposed methods, we employed four benchmark datasets for synthesizing super-resolution images. We used two image quality evaluation metrics: NIQE, which evaluates the naturalness of images, and LPIPS, which provides a human-like perceptual evaluation results. Experimental results demonstrate a significant enhancement in image naturalness and perceptual evaluation values compared to previous studies highlighting the effectiveness of the proposed methods.

Keywords

• Super-resolution
• generative adversarial networks
• channel attention