UASR: An Unified-Attention Mixer Network for Efficient Image Super-Resolution

Read the full article See related articles

Listed in

This article is not in any list yet, why not save it to one of your lists.
Log in to save this article

Abstract

Recent works in single-image super-resolution (SISR) have brought notable improvements to the field. Transformer-based methods enhance reconstruction quality by capturing long-range dependencies. However, the quadratic computational complexity of multi-head self-attention (MHSA) introduces efficiency bottlenecks in HR image processing, and insufficient local feature extraction limits the recovery of fine texture details and edge sharpness. In contrast, convolutional neural network (CNN)-based methods suffer from limited receptive fields, leading to inadequate high-frequency detail recovery and blurring artifacts. Generally, Transformer-based and CNN-based methods fail to simultaneously address the challenges of computational efficiency, global dependency modeling, and local feature extraction. To integrate the strengths of both paradigms, we propose Unified-Attention Super-Resolution(UASR) network, a lightweight architecture based on the Convolutional Transformer(ConvFormer) layer. Specifically, UASR replaces MHSA with the Unified-Attention Mixer (UA-M) that efficiently captures global dependencies at a low computational cost. Additionally, the Reparameterized Edge-Extraction FeedForward Network (REFN) supplements UA-M by focusing on extracting texture and edge features. Furthermore, we introduce a Spectral Unified-Attention Block (SUAB) that extends the capabilities of UA-M into the frequency domain, thus improving detail reconstruction and accelerating the computation process. Compared to current CNN-based and Transformer-based SISR models, experimental results demonstrate that our method strikes an effective balance between accuracy and efficiency, enhancing texture fidelity and super-resolution performance.

Article activity feed