Image noise is, and has always been a significant challenge in digital image processing. In recent years cheaper image capturing solutions and real-time path- and ray-tracing rendering have made noise reduction even more important than ever. The recent growth of Artificial Intelligence and machine learning research have provided an opportunity to target this problem from a different perspective. This research and its small experiments explore the current state of image noise reduction using both traditional and machine learning based methods. We try to explain in a more understandable way the different types of image noise on digital photographs, and ray and path tracing images, how they are generated and what methods can use to remove them. We found that most commercial digital photo editing software use simple methods like non-local means filtering and bilateral filtering, often combined with additional edge detection enhancements, such as high-pass filtering. We explored some deep learning based tools as well that performed comparable or better than traditional methods, but suffered from high computational time. Our experiments also showed that Monte Carlo noise can be removed much more effectively and faster using deep learning than traditional methods, such as non-local means filtering, and that deep learning based noise removal algorithms are now the favored tools for ray and path tracing.
|Uddannelser||Datalogi, (Bachelor/kandidatuddannelse) Kandidat|
|Udgivelsesdato||31 dec. 2019|