The role of involution in lightweight super resolution

In recent years, single-image super resolution (SISR) methods using deep convolutional neural networks (CNN) have achieved impressive results. Deep networks learn a complex nonlinear mapping between low-resolution (LR) image patches and high-resolution (HR) versions in a network. However, the enhanced performance of many CNN-based models often comes at the expense of increased memory consumption, making implementation challenging on resource-constrained devices. To minimize the number of parameters, an effective and efficient operator called involution has been employed, which provides better performance at a lower cost compared to its convolution-based counterparts. While involution has previously found success in Image Classification, Object Detection, and Image Segmentation, its application in single-image super resolution is less explored. Information multi-distillation network (IMDN) is identified as a lightweight model providing remarkable performance in SISR. In this work, we propose a strategy that combines involution and convolution in the feature extraction blocks of IMDN. The proposed approach improves network performance with fewer parameters. More specifically, we proposed to use involution more often than convolution as the network gets deeper. Unlike many networks, the blocks of ours do not have the same structure. Our proposed lightweight model matches the PSNR performance of the IMDN model in fewer epochs and achieves this with a 200k reduction in parameters. Additionally, the SSIM metric stands at 0.9675, showcasing an improvement over the baseline model. The code is available at https://github.com/akramkhatami/IMDB_DeepInvo.