A vision-based two-stage framework for defect detection of tiny wafer probe marks

Defect detection of tiny wafer probe marks is crucial for semiconductor manufacturing. Current defect detection in industry relies on manual inspection, which is laborious and time-consuming. Although artificial intelligence-based vision methods have been widely applied to surface defect inspection across diverse scenarios, little effort has been made on wafer probe mark defect detection, which remains challenging due to the tiny size of probe marks and the complex backgrounds. To solve the problem, we first propose a two-stage framework for defect detection of tiny wafer probe marks. In the first stage, a wafer probe mark detection network (WPMNet) is proposed, which integrates a spatial–frequency dual-domain mamba module for enhanced global context modeling and background suppression, as well as a dual-scale modulated gating unit to strengthen multiscale texture and geometric representations, so that the tiny wafer probe marks can be accurately detected. In the second stage, the center point of each detected probe mark is extracted, and used as a point prompt to guide the segment anything model to segment the corresponding pad. Based on the morphology of the segmented probe marks and pads, tiny wafer probe mark defects can be detected. Experiments on a dataset of wafer probe marks demonstrate that the proposed WPMNet achieves the best overall detection performance among 11 comparative methods. For probe marks defect detection, the proposed framework achieves an accuracy of 0.95 with an average inference time of 0.46 s per image (1280 × 960), indicating its effectiveness and potential for practical wafer defect detection.