Enhancing weld defect detection and classification with MDCBNet: A Multi-Scale Dense Cross Block Network for Improved explainability

Abstract

MDCBNet, a Multi-Scale Dense Cross Block Network designed explicitly for enhancing the detection and classification of weld defects, is presented in this paper. The primary emphasis of this work is attaining systematic weld defect detection and classification using 2-class and 5-class scenarios. The network's explainability is also analyzed to understand its classification decisions better. From extensive experimentation, it is observed that the proposed technique can accurately differentiate between acceptable and defective welds and find the type of defect from the images captured during welding. Moreover, a novel method for comparing networks with tightly grouped accuracies is proposed for robust evaluation of the performance of networks. The proposed approach surpasses pre-trained deep learning techniques, exhibiting an increased accuracy of 2 %–10 % across various metrics, and demonstrates superiority in weld defect classification compared to existing methods.