Scoring Bayesian Neural Networks for learning from inconsistent labels in surface defect segmentation

Abstract

This paper focus on surface defect segmentation uncertainty challenges that arise due to human errors and biases in the data annotation process, particularly in ambiguous transition and weak feature areas. Firstly, uncertain areas are defined, where it could not be unambiguously identified as defect or defect-free. Then a scoring Bayesian neural network is proposed, using Bayesian neural computation to solve the segmentation probability and provide an expression for uncertain areas. The variance of segmentation probability is utilized to assess the quality of labels, thereby improving model performance. The approach is validated against prevailing state-of-the-art methods on five datasets, demonstrating superior performance. This study provides a crucial pathway for addressing human errors and biases in defect detection. The code is available at https://github.com/ntongzhi/Scoring-BNNs.