B-BACN: Bayesian Boundary-Aware Convolutional Network for Defect Characterization

Detecting accurate crack boundaries is important for condition monitoring, prognostics, and maintenance scheduling. In this work, we propose a Bayesian Boundary-Aware Convolutional Network (B-BACN) to tackle this problem, that emphasizes the importance of both uncertainty quantification and boundary refinement for producing accurate and trustworthy detections of defect boundaries. We formulate the inspection model using multi-task learning. The epistemic uncertainty is learned using Monte Carlo Dropout, and the model also learns to predict each samples aleatoric uncertainty. A boundary refinement loss is added to improve the determination of defect boundaries. Experimental results demonstrate the effectiveness of the proposed method in accurately identifying crack boundaries, reducing misclassification and enhancing model calibration.