Lightweight Dual-Kernel Information Aggregation Network for Efficient Image Super-Resolution

Efficient single-image super-resolution (ESISR) primarily aims to enhance super-resolution (SR) performance while keeping model complexity low, making it more suitable for deployment on edge devices. However, the limited receptive field caused by conventional convolution's locality restricts the backbone networks and attention modules from effectively capturing nonlocal features, leading to suboptimal SR performance. Additionally, insufficient interaction between high- and low-frequency feature information results in incomplete feature representation. To overcome these limitations, we propose a novel ESISR network called the lightweight dual-kernel information aggregation network (LDIAN). First, we design a dual-kernel convolution (DKC) that combines depth-wise 1-D convolution and dilated convolution to efficiently extract richer image features in an expanded receptive field while minimizing model complexity. Building upon DKC, we further develop a dual-kernel enhanced convolution (DEConv) and a dual-kernel enhanced distillation block (DEDB). Additionally, we propose a lightweight dual-kernel attention (DKA) mechanism to focus on more representative features for SR reconstruction. Second, we design an innovative feature fusion structure named the information aggregation block (IAB) to integrate spatial features and strengthen the interaction between high- and low-frequency information, thereby improving feature representation. Extensive quantitative and qualitative experiments demonstrate that the LDIAN achieves state-of-the-art performance with an optimal balance between model performance and complexity. Notably, compared to SRFormer-light, LDIAN-L delivers superior performance across five standard datasets while requiring only about 50% of the model's FLOPs.