Enhanced image super-resolution via two-stage latent probabilistic diffusion model

Abstract In recent years, with the development of technologies in the field of image processing, the demand for high-quality images has been increasing. Although existing image super-resolution (SR) reconstruction methods can achieve good reconstruction results using diffusion models, the increased training costs and reconstruction cycles caused by the pursuit of improved image quality have limited their applications. To accelerate inference while achieving high-quality images, we propose a two-stage latent space efficient diffusion model for image SR reconstruction. Specifically, we transform input features into a latent space for extraction using autoencoders. This approach reduces computational resource consumption while fully capturing high-frequency information. In addition, we design an improved U-Net-based feature extraction strategy, which achieves the extraction and fusion of deep features through a residual compensation structure and cross-attention. Finally, we accelerate the SR reconstruction by iteratively optimizing the denoising process, which significantly shortens the inference cycle. We conducted experiments on mainstream datasets, and the results show that the proposed method achieves excellent performance in both visual effects and quantitative evaluation, which proves the effectiveness of this method in image resolution enhancement.