Enhancing feature information mining network for image super-resolution

Convolutional neural networks (CNN) have been widely used in image super-resolution tasks in recent years, with remarkable results. Most existing CNN-based image super-resolution methods, on the other hand, deepen the network structure to increase the receptive field and do not fully utilize the intermediate features, resulting in limited extracted information and loss of important information. To address these issues, we propose an enhancing feature information mining network (EFMNet) that aims to enhance feature capture and mining. Specifically, a calibrated multi-scale module (CMS) is proposed that powerfully extracts feature information from different scales by accessing different ranges of pixels in the spatial domain and adaptively adjusts feature information. Furthermore, to effectively retain high-frequency information, a dual-branch attention block (DAB) is developed that captures dependencies between intermediate features, and learns the confidence of each pixel location in the feature map to capture more informative feature. Qualitative and quantitative evaluations from extensive experiments on benchmark datasets demonstrate that our network achieves advanced performance.