Project DetailsThe project focuses on image denoising using diffusion models, with an emphasis on improving existing performance and achieving high quality results. The goal is to process images containing various types of noise, such as Gaussian noise, Poisson noise and real noise, and restore them to a state as close as possible to the original image quality. This approach is essential in many fields, such as medical imaging and image processing, where precise noise removal directly impacts the quality of analyses, and the decisions made based on the images.
The chosen model for the project is NCSNv2 - a convolutional neural network originally designed for image generation tasks. During the project, the model was adapted for denoising purposes and modified to handle different types of noise. The experiments in the project relied on a diffusion process that included iterative sampling and progressive signal enhancement, using the Annealed Langevin Dynamics method, which provides a stepwise mechanism for noise removal and data reconstruction. Another approach in our work involved incorporating the Anscombe transform into the denoising process to handle Poisson noise effectively.
The results were high-quality, both in terms of numerical metrics and visual appearance. This project opens new opportunities for future research, including further improvements, adaptations for more complex noise types and the integration of additional techniques to enhance image quality.
