Study of Image Denoising Methods through U-net Based Machine Learning Design
Joseph Quan · 2025
There have been significant advances in the field of image denoising using machine learning techniques.In state-of-the-art methods, system complexity and system performance have been steadily improving over the years.Still, there is additional work to simplify the networks while keeping high performance.In this paper, we introduce several image denoising techniques, including those state-of-the-art methods with system complexity and the simplified ones.We did experiments using NAFNet, a deep learning denoising model that made the simplification through avoiding nonlinear activation functions.We found that by refining the dataset and introducing new training images, the quality of the results could be substantially improved.Furthermore, we experimented with different model structures and found that we can reduce model complexity substantially, while model effectiveness does not diminish much.