Single Image Super Resolution Based on Dual-Path Large Kernel Learning

To excel in image super-resolution, deep neural network-based models often employ a stacking strategy of network modules. However, this approach leads to parameter explosion and information redundancy, thereby restricting the feasibility of deploying these models on mobile devices. To tackle this challenge, this study introduces a lightweight Dual-path Large Kernel Learning approach, namely DLKL, and applies it to image super-resolution. In DLKL, the first step involves the utilization of multi-scale large kernel decomposition to effectively establish long-range dependencies between pixels and successfully preserve local information. Subsequently, DLKL enhances the feature expression ability and achieves effective feature fusion through its dual-path network architecture. Furthermore, DLKL reduces the number of parameters while maintaining performance, thus striking a balance between network performance and efficiency. The remarkable performance of DLKL has been validated through a multitude of experiments by quantitative metric tests and visual evaluations. Comparative analysis, conducted against other prevailing algorithms, provides evidence of the DLKL method’s consistent proficiency in generating images featuring enhanced texture clarity and more faithful representation of natural structures. These findings serve to reinforce the method’s inherent superiority and robustness.