Background
AI-driven defect detection is a specific application in the industry that applies AI and machine learning algorithms to localize the location and identify the classification of the defects in industrial products. During the manufacturing process, AI-driven defect detection has been applied to lots of different products to guarantee the quality of the products, such as solar power photovoltaic cells, fabric, and metal products.
Challenges
Interclass difference in scales and shapes. As demonstrated in Fig. 1 (d), there are multiple-class defects (Crack and Fragment) that exist in a single EL image.
Intraclass variations in scales and shapes. Thick Line usually consist of multiple lighter-colored columns as demonstrated in Fig. 1 (a) and (c).
Interclass similarity. As demonstrated in Fig. 1 (c) and (d), Fragment and Black Core gain high similarity since the difference between them is Fragment has more clear and regular borders. The interclass similarity makes the network harder to classify similar defects precisely.
Bounding box overlapping. As demonstrated in Fig.1 (d), the overlapping exists between the bounding box of Fragment and Crack, which brings challenges for the network to learn efficient and accurate features of these overlapped defects.
Location-sensitive. Horizontal Dislocation and Vertical Dislocation usually present as a linear-like horizontal or vertical-distributed defects. However, the only difference between these two kinds of defects is the location distribution.
Junqi Bao, Xiaochen Yuan, Qingying Wu, Chan-Tong Lam, Wei Ke, Ping Li
IEEE Transactions on Instrumentation and Measurement, vol. 74, 2025, pp. 1-12