Lightweight Image Super-resolution with Local Attention Enhancement

In recent years, methods based on convolutional neural network (CNN) have been the mainstream in single image super-resolution (SISR). Although these methods have achieved excellent performance, the massive amount of model parameters and heavy computation limit their application. On the other hand, channel attention (CA) mechanism, which can enhance network performance, has also been widely used in SR task recently. However, the channel attention mechanism is introduced from high-level vision tasks to the SR task. The original design of this mechanism doesn’t consider the specificity of the SR task. To address these issues, we propose a lightweight expansion and distillation residual network (EDRN) for image super-resolution. Specifically, through the diverse use of different feature channels and different convolution kernel sizes, our network can effectively reduce the amount of parameters while achieving superior performance. To further explore the potential of channel-wise attention in the SR task, we develop a novel plug-and-play local channel attention enhancement strategy (LCAES) to make the network better use the characteristics of local features of the image. Furthermore, comprehensive quantitative and qualitative evaluations demonstrate that the proposed method performs favorably against state-of-the-art SR algorithms in terms of visual quality, reconstruction accuracy, and parameter amount.