Super-Resolution Reconstruction Integrating Hybrid Attention Network and Diffusion Model

Benefiting from deep learning, image reconstruction techniques have rapidly advanced. Diffusion models, as a class of emerging deep generative models, have been applied in various fields. For the task of image super-resolution, more pixel features are required, whereas diffusion models only utilize U-Net networks for image denoising and restoration, failing to fully exploit image features. In this paper, we propose a diffusion model super-resolution reconstruction algorithm, HAT-SRDM, based on a hybrid attention encoding network. Firstly, a hybrid attention mechanism is employed to capture both global and local information of the image, activating more useful pixel features. Subsequently, the obtained image features are used as conditions to control the iterative refinement of the diffusion model, along with the utilization of residual image prediction model outputs. Finally, the training loss function of the diffusion model is optimized to better reconstruct image details. Experimental results are compared with several methods on four test sets, demonstrating good performance in terms of PSNR, SSIM, LPIPS, and LR-PSNR metrics.