Uncertainty Quantification Based on Gaussian Processes for Image Segmentation Tasks

Over the past several years, deep neural networks have permeated many fields of science research and have become an essential part of real-world applications. However, when the model encounters a test sample with uncertainty, compared with the distribution of the training set data, it is crucial for users to determine which of the model's predicted outputs for those test data are trustworthy and which are not, rather than being forced to give an untrustworthy prediction output. To this end, it is necessary to evaluate the model uncertainty, which enables to improve utilization of the model prediction results, especially some safety-critical systems. For image segmentation tasks of safety-critical systems, this paper proposes an uncertainty quantification method based on Gaussian processes to evaluate the trustworthiness of the output predicted by neural network models for the given input, so as to facilitate the selection and optimization of segmentation models and provide the model interpretability. The simulation results indicate that for the pixel error rate of each image, this method can give the confidence interval of the predicted output and the stability of the predicted output, achieving the objective of quantifying and understanding the confidence level of deep learning models.