Large Kernel Frequency-enhanced Network for Efficient Single Image Super-Resolution

In recent years, there has been significant progress in efficient and lightweight image super-resolution, due in part to the design of several powerful and lightweight attention mechanisms that enhance model representation ability. However, the attention maps of most methods are obtained directly from the spatial domain, limiting their upper bound due to the locality of spatial convolutions and limited receptive fields. In this paper, we shift focus to the frequency domain, since the natural global properties of the frequency domain can address this issue. To explore attention maps from the frequency domain perspective, we investigate and correct some misconceptions in existing frequency domain feature processing methods and propose a new frequency domain attention mechanism called frequency-enhanced pixel attention (FPA). Additionally, we use large kernel convolutions and partial convolutions to improve the ability to extract deep features while maintaining a lightweight design. On the basis of these improvements, we propose a large kernel frequency-enhanced network (LKFN) with smaller model size and higher computational efficiency. It can effectively capture long-range dependencies between pixels in a whole image and achieve state-of-the-art performance in existing efficient super-resolution methods.

1. Introduction

As a low-level computer vision task, single-image super-resolution (SISR) aims to reconstruct a high resolution (HR) image from its low resolution (LR) counterpart. Since SR-CNN [9] introduced deep learning to super-resolution for the first time, there has been a significant surge in the development of deep-learning-based SR models. By leveraging large amounts of data and powerful computing resources, deep learning has enabled researchers to develop increasingly sophisticated SR models that can generate high quality image from low-resolution inputs. Despite their impressive results, due to their high complexity and computational cost, traditional super-resolution networks are often difficult to use in practical applications. In this context, efficient super-resolution (ESR) networks with greatly reduced parameters and less computational complexity are gradually being introduced and developed.