A Lightweight Local Attention Network for Image Super-Resolution

For many years, deep neural networks have been used for Single Image Super-resolution (SISR) tasks. However, more extensive networks require higher computing and storage costs, obstructing their deployment on resource-constrained devices. How to ensure the model lightweight while improving its performance is an important direction of SISR research at present. This paper proposed a lightweight local attention network (LLAN) mainly composed of lightweight residual attention groups (LRAGs). LRAG contains lightweight self-calibrated residual blocks with pixel attention (LSC-PAs) and average local attention blocks (ALABs); it utilizes the advantages of the residual connection and attention mechanism. The LSC-PA with strong expression ability can propagate and fuse features better. The ALAB can combine global features and accelerate the network. Furthermore, we discussed three parts of a general SISR network: feature extraction, feature fusion, and reconstruction. Besides, we carried out extensive experiments using five benchmark datasets. The experimental results demonstrated that our method outperforms other compared state-of-the-art techniques, and a better balance between the complexity and performance of the SISR algorithms is achieved.