GAMSRN: Global Attention Multi-Scale Residual Network for Single-Image Super-Resolution and Low-Light Enhancement

The deep convolutional neural network (CNN) has achieved significant success in image restoration (IR). However, most deep neural networks use a single-sized convolution kernel to capture spatial information, limiting local feature extraction. Additionally, since convolution is designed for local operations, it struggles to capture the inherent global characteristics of an image. This paper addresses these issues by proposing a novel global attention multi-scale residual network (GAMSRN) based on global attention and multi-scale feature extraction. The network employs global attention multi-scale residual blocks (GAMSRB) that use convolution kernels with uniform sizes but varying dilation rates to extract features at different scales without increasing model complexity. Furthermore, the attention mechanism effectively models global features and enhances the network's expressiveness. These components together enhance performance. Our proposed network is generalizable for two applications: super-resolution and low-light enhancement. Extensive qualitative and quantitative comparison with six synthetic benchmark datasets demonstrates the efficacy of our proposed GAMSRN over reported works.