Image Super-Resolution Reconstruction Method Based on Hybrid Convolution and Transformer

Image super-resolution (SR), a fundamental task in low-level vision, refers to the process of reconstructing high-resolution images from low-resolution images and its applications are extensive thus widely applied in medical imaging, satellite remote sensing and video enhancement among others. Given the challenges posed by previous convolutional or transformer-based SR methods in modeling local details, and long-distance dependencies, in this paper we present a new hybrid model, HCTrans-SR (Hybrid Convolutional and Transformer-based Super-Resolution Network). Based on the dominant SwinIR framework, we designed a two-branch feature extraction module with one branch based on a multi-scale residual-dense convolutional network to extract fine textures, while the other branch includes a window-based Shifted Transformer block that fuses the axial attention mechanism to model global context. To leverage the multi-level features from the two branches effectively, we also propose a cross-scale feature interaction module that dynamically weights and aggregates different levels of representation using adaptive gating and channel attention. We further propose a multi-stage reconstruction decoder to progressively optimize the spatial accuracy and structural sharpness of the image, and develop a novel multi-domain loss that combines pixel-level, perceptual domain, and frequency domain constraints in order to achieve better reconstruction quality. Results from our experiments show that the proposed method demonstrates more robust reconstruction ability in the reconstruct high-frequency texture region.