Lightweight Multi-Stage Holistic Attention-Based Network for Image Super-Resolution

Abstract

High-resolution images are crucial for many applications, but factors such as environmental conditions can reduce image quality. Super-resolution (SR) techniques address this by generating high-resolution images from low-resolution inputs. While deep learning SR models have made significant progress, they can be computationally expensive and struggle with differentiating between various image scales. Lightweight SR methods, suitable for resource-constrained devices, often compromise image quality. This study introduces a multi-stage holistic attention-based network, using Gaussian Laplacian pyramids to decompose images and apply holistic attention modules at each level. This approach reduces parameters and computational costs while maintaining image quality, achieving a PSNR score of 28 and SSIM of 0.91 with only 29,000 parameters. The model demonstrates the potential for efficient and high-quality image reconstruction. Future work will focus on improving quality while minimizing costs and exploring other advanced techniques. The code will be made available upon request