抄録
Shrinkage cavities in solder joints are often misidentified as cracks in automatic inspection systems, resulting in false positive detections. This study proposes a shrinkage cavity detection method that combines PatchCore and image classification using images acquired before heat shock testing. First, solder regions were extracted using SegGPT, and several preprocessing conditions were compared. PatchCore was then applied for primary anomaly detection, followed by secondary classification using ResNet50 to remove false positives. Experimental results showed that cropping the central solder region and subtracting the mean anomaly map improved image-level detection performance. Furthermore, the addition of secondary classification reduced the number of false positives from 27 to 19, improving Precision from 0.799 to 0.846 and F1-score from 0.881 to 0.897. The proposed method is expected to improve crack inspection accuracy by excluding shrinkage cavities detected before heat shock testing.
https://ror.org/0151bmh98 
