To tackle the challenges of high computational complexity and inadequate multi-scale feature fusion in single-image super-resolution, this paper proposes the Multi-Scale Attention Network with Dynamic Offset Sampling (MADN). MADN leverages depthwise separable convolution and 1 \(\times \) 1 pointwise convolution for efficient local feature extraction, significantly reducing the number of parameters while maintaining robust feature representation. To address limitations in capturing multi-scale information, a grouped multi-scale hybrid attention module is designed to integrate multi-scale convolution kernels with cross-channel and cross-scale spatial attention, thereby improving the extraction of high-frequency details. Additionally, dynamic upsampling is introduced to replace the traditional upsampling method, enabling content-aware feature reconstruction through learnable sampling offsets. Experimental results demonstrate that MADN achieves competitive performance and efficiency. Compared to SwinIR, the light version improves PSNR by 0.03 dB on Manga109 (39.15 vs. 39.12) with 21% fewer parameters and comparable complexity, while the large version reduces parameters by 26% and computation by 4.4% on Set5, maintaining a performance gap of less than 0.1 dB. The code is publicly available at: https://github.com/Ryuusei123/MADN.
