The paper presents an image restoration method based on the application of a diffusion model utilizing reference images. The proposed approach aims to enhance the quality of restored image details by integrating information from references. Diffusion models, known for their ability to generate high-resolution images, are employed to address the limitations of convolutional neural networks, such as the insufficient quality of images restored after severe degradation. Furthermore, the quality of images generated by the diffusion model can be further improved by incorporating information from reference images. This study describes the proposed model architecture, training procedure, and comparison results with the no-reference model. The experimental results confirm the effectiveness of the proposed approach, specifically demonstrating an improvement in the quality of the resulting image by more than 9% in terms of the PSNR metric and by 14% in terms of the LPIPS metric.

image restoration, neural networks, diffusion models, super-resolution, deep learning

Mathematical modelling in actual problems of science and technics