Lightweight cross-domain image super-resolution network based on hybrid attention

Image super-resolution has witnessed great progress with the development of deep learning techniques, where SR performance has been continuously surpassed on benchmark datasets. However, the escalating model parameters and computational complexity accompanying performance improvements have led to escalating deployment and operational costs in real applications. We propose an innovative lightweight Cross-Domain Hybrid Attention Network (CDHAN) for image super-resolution reconstruction in this work. The network employs a Frequency-Spatial Decision Network (FSDN) to dynamically analyze the feature significance of input images in both frequency and spatial domains, generating hybrid weight matrices to guide subsequent feature fusion. We design a dual-branch hybrid attention module that synergistically integrates frequency band attention, spatial attention, and channel attention mechanisms, achieving collaborative optimization between global structural restoration in the frequency domain and local detail enhancement in the spatial domain. The architecture incorporates lightweight design principles with residual connections to enhance training stability while maintaining superior detail reconstruction capability with reduced parameters. The experiment results show that our method perform very well, while requiring fewer parameters compared to mainstream models. This work provides new insights into efficient dynamic fusion strategies for image super-resolution tasks.