Fuzzy Uncertainty-Based Out-of-Distribution Detection Algorithm for Semantic Segmentation

Deep learning models have achieved high performance in numerous semantic segmentation tasks. However, when the input data at test time do not resemble the training data, deep learning models can not handle them properly and will probably produce poor results. Therefore, it is important to design algorithms for deep learning models to reliably detect out-of-distribution (OOD) data. In this paper, we propose a novel fuzzy-uncertainty-based method to detect OOD samples for semantic segmentation. Firstly, to capture both data and model uncertainties, test-time augmentation and Monte Carlo dropout are applied to a ready-trained image segmentation model for generating multiple predicted instances of a given test image. Then interval fuzzy sets are generated from these multiple predictions to describe the captured uncertainty via distance transform operators. Finally, an image-level uncertainty score, which is calculated from the generated interval fuzzy sets, is used to indicate if it is an OOD sample. Experiments on testing three OOD test sets on a skin lesion segmentation model show that our proposed method achieved significantly higher classification accuracy in detecting OOD samples than three other state-of-the-art uncertainty-based algorithms.