Multi-scale Attention Residual Convolution Neural Network for Single Image Super Resolution (MSARCNN)

Abstract

Traditional super-resolution methods often struggle to capture fine details and extract features, especially at higher frequency which leads to poor reconstruction of images. Further some SR methods neglect the significance of complexity while designing deeper networks. Deeper networks are challenging to train and have greater computational load which limits the performance of SR method making it less compatible for other devices. To address this problem, we propose a novel Multi-Scale Attention Residual Convolutional Neural Network(MSARCNN). The model combines eight multi-scale attention residual convolution and a Dilated Convolution Block(DCB). Each MSARCB comprises of a squeeze and excitation block which recalibrates feature maps by emphasizing informative channels and a Pixel Attention Block(PAB) which utilizes attention-based weighting to enhance local feature representation. The MSARCB employs multi-scale hierarchical feature extraction with the help of parallel convolution layers with varying channels and DCB with dilation rates of 1,3,5 and 7 which helps in capturing both spatial features and fine details by enlarging the effective receptive field without increasing the number of learnable parameters. Experiments on four benchmark dataset demonstrate that the proposed model significantly outperforms other state-of-the-art lightweight SR methods, providing a exceptional balance of reconstruction performance, model complexity and parameter count.