Attention Guidance Distillation Network for Efficient Image Super-Resolution

Over the past decade, neural network-based super-resolution techniques have been developed on a large scale with impressive achievements. Many novel solutions have been proposed, among which lightweight solutions based on convolutional neural networks have been designed for applications in edge devices. To better realize this application, we propose a more lightweight attention guidance distillation network (AGDN). We design the attention guidance distillation block (AGDB) with more efficient space, channel and self-attention as the infrastructure of AGDN. Specifically, multi-level variance-aware spatial attention (MVSA) is designed to better capture structurally information-rich regions with new multi-scale convolution and local variance alignment. Reallocated contrast-aware channel attention (RCCA) is designed to enhance the processing of common information in all channels while redistributing weights across channels. Sparse global self-attention (SGSA) is introduced for selecting the most useful similarity values for image reconstruction. Extensive experiments demonstrate that AGDN strikes a better balance between performance and complexity compared to other models, achieving SOTA performance on several benchmark tests. In addition, our AGDN-S ranks first in the FLOPs track and second in the Parameters track of the NTIRE 2024 Efficient SR Challenge. The code is available at https://github.com/daydreamer2024/AGDN.

1. Introduction

Single-image super-resolution (SISR) is a fundamental task in the field of computer vision that aims to generate high-resolution (HR) images from low-resolution (LR) images. As a critical component of low-level vision tasks, SR techniques find widespread use in various real-world applications, such as remote-sensing imaging, medical imaging, and security surveillance. In recent years, propelled by advancements in deep learning technology, SR techniques have made remarkable strides, leading to the emergence of numerous innovative network architectures. Beginning from the early convolutional neural networks [7] and progressing through residual networks [13], to transformer models [4] and diffusion models [26], the field of SR has witnessed a flourishing development. Nevertheless, with the continuous evolution of technology, SR networks have grown increasingly complex, and their network structures have expanded in size. While these sophisticated SR networks enhance the quality of image reconstruction, their deployment on edge devices with limited computational resources presents challenges due to the escalating model capacity and intensive computational demands.