DMNet: detail modulation network with gradient-texture guidance for lightweight image super-resolution

Abstract Recently, numerous lightweight image super-resolution methods with low computational complexity have been proposed, achieving a promising trade-off between reconstruction quality and efficiency. However, most existing methods are constrained by their compact network design, making it challenging to reconstruct high-quality images, particularly in regions with intricate edges and texture details. To address these issues, we propose an efficient Detail Modulation Network (DMNet), which focuses more on the restoration of high-frequency components. Specifically, we first present a Gradient Awareness Block (GAB) that explicitly calibrates gradient information in low-resolution images by leveraging the intrinsic high-frequency representation properties of gradient domain, thereby providing targeted guidance for edge enhancement. Subsequently, we introduce a Texture Calibration Block (TCB) that employs adaptive texture descriptors to dynamically direct attention toward degraded regions and enable accurate texture recovery. Finally, we develop an iterative Detail Modulation Module (DMM), which orchestrates GAB and TCB in a progressive refinement strategy to effectively restore high-fidelity images with sharp edges and clear textures. Extensive experiments across multiple benchmark datasets demonstrate that our DMNet successfully restores clear textures and fine details, achieving a compelling balance between performance and efficiency.