Noise-Robust Diffusion Based Semantic Segmentation

Many semantic segmentation methods that are either fast or having high accuracy based on Jaccard index(IoU) have been proposed in the literature. Therefore, diffusion models, which have shown very successful results recently, have also been used in semantic segmentation studies. Although diffusion based segmentation methods work faster than the current state-of-the-art methods, diffusion models did not provide sufficient performance in terms of accuracy. However, it is important observation that noisy and ordinary images can be used together since the diffusion models have a structure that is robust for learning. In this study, the NRSeg architecture in which diffusion models can use both noisy and ordinary images together was created. The new model performance was measured in terms of IoU and these results were compared with the performances of the state-of-the-art methods in the literature.