Partial convolutional reparameterization network for lightweight image super-resolution

In recent years, convolutional neural networks (CNNs) have made significant strides in single image super-resolution (SISR). However, redundancy persists in network models concerning both channels and network structures, constituting a challenge in designing lightweight super-resolution (SR) networks. Consequently, finding a balance between efficiency and performance has emerged as the focus in SR research. In response to these challenges, we propose the Partial Convolutional Reparameterization Network (PCRN) for lightweight SR. Specifically, we initially employ partial convolution to reduce channel redundancy. Subsequently, we employ a complex network structure during model training, while in the inference stage, we utilize reparameterization techniques to compress the model, thus reducing redundancy in the network structure. Moreover, we have introduced enhanced spatial attention (ESA) and efficient channel attention (ECA) modules into our approach to enhance the model’s capability to extract key information. In comparative experiments, the proposed PCRN demonstrates superior performance over other efficient SR methods.