Context-Aware Detail Recovery Network for single image super-resolution

Abstract

With the advancement of deep learning, single-image super-resolution (SISR) has achieved significant breakthroughs. Although existing methods have substantially improved performance, challenges remain in accurately recovering fine visual details. To address this issue, we propose a Context-Aware Detail Recovery Network (CADRN), which leverages the complementary strengths of Convolutional Neural Networks (CNNs) and Transformers through three cascaded modules. Firstly, the Detail Extraction Engine (DEE) incorporates a Selective Feature Enhancer (SFE) that utilizes hand-crafted convolutional kernels for feature displacement, effectively enhancing local context while preserving structural details. Secondly, The Global Multi-Scale Context Module (GMCM) employs adaptive attention mechanisms to effectively capture long-range dependencies for global context modeling. Finally, the Feature Refinement Module (FRM) employs a carefully designed bidirectional processing pipeline to further capture and restore fine details. Experimental results demonstrate that CADRN achieves state-of-the-art performance, offering enhanced detail reconstruction while maintaining computational efficiency.