CRUN: a super lightweight and efficient network for single-image super resolution

In recent years, single-image super-resolution (SISR) has made significant progress using convolutional neural networks. However, to achieve better performance, many methods deepen the network depth and stack a large number of parameters. Unfortunately, this leads to high computational complexity and is difficult to apply on ordinary intelligent devices used by the public. To address this issue, we propose a lightweight SISR network that uses thin convolutional groups to significantly reduce network size. We also optimize the network structure to compensate for the loss in performance caused by the reduction in the number of parameters. Our proposed CRUN(Cascade Compound Residual UNet Network) model utilizes cascade residuals structure as the backbone to deepen the network and increase its receptive field. We then use skip connections and dense connection structures to design deep feature extraction blocks that reduce the network’s dependence on feature dimensions. Moreover, we introduce a Multi-scale Feature Enhancement Block (MFEB) to fuse features at different levels, compensating for the weakness of thin convolutional groups in obtaining features. Our experimental results show that the CRUN model achieves superior performance compared to state-of-the-art lightweight models while requiring relatively fewer resources. This improvement makes our proposed model more practical and suitable for usage on everyday devices.