Partial convolution residual network for lightweight image super-resolution

Recently, convolutional neural network (CNN) based approaches have shown remarkable achievement for single image super-resolution (SISR). However, CNN-based SR methods often struggle with the trade-off between image reconstruction quality and model complexity. In this paper, we propose a Partial Convolution Residual Network (PCRN), which improves SR performance in terms of image reconstruction quality and model size through two design aspects. Firstly, we revisit the commonly used pixelshuffle upsampling function in SR models, demonstrating that the pixelshuffle operation causes a significant dispersion of the receptive fields of convolutional kernels, leading to a reduction in the model’s ability to extract local texture features. To address this problem, we designed a pointwise convolution residual module to focus on the receptive field, allowing the model to capture local features more precisely, thereby effectively enhancing image restoration performance. Secondly, inspired by FasterNet, we leverage partial convolution to reduce the model’s complexity and introduce the partial convolution residual block as the foundation of PCRN. This design significantly reduces redundant channels and memory access operations, further improving model performance. Experimental results demonstrate that PCRN can outperform other typical SR models.