Lightweight Real-World Image Super Resolution via Channel Redundancy for Edge IoT Devices

In real world scenarios, low-resolution images often suffer from complex and unknown distortions during acquisition by edge Internet of Things (IoT) devices. To achieve a balance between super-resolution and distortion suppression, Real World Super Resolution (RWSR) models need to possess both global and local feature modeling capabilities. Moreover, a lightweight design is necessary for edge deployment without compromising performance. Therefore, this paper proposes a lightweight Channel Group Rearrange Transformer (CGRFormer), which exploits channel redundancy to reduce computational burden while enhancing local feature modeling capabilities. CGRFormer comprises two key components: the Channel Group and Rearrange Operation (CGRO) and the Multi-Scale Local Enhancement Feed Forward Network (MSLE-FFN). To address the issue of a heavy computational burden, CGRO divides the feature maps into active and inert groups. Heavy operations are applied only to the active group, significantly reducing overall complexity. A subsequent channel rearrange operation enhances information interaction between the two groups. To enhance local feature modeling, MSLE-FFN incorporates a Multi-Scale Convolutional Group (MSCG) and a Channel Excitation Module (CEM), enabling effective extraction and enhancement of local features at different receptive fields. Overall, CGRFormer effectively balances super-resolution and distortion suppression while maintaining a lightweight framework. Experimental results on the commonly used test datasets show that our model is still competitive with the SOTA model.