超音波フェーズドアレイ像のYOLOによる欠陥検出

None

神田 侑奈*; 太田 陽生*; 鍋島 邦彦  ; 矢田 浩基 ; 古谷 正祐*

Kanda, Yuna*; Ota, Yosei*; Nabeshima, Kunihiko; Yada, Hiroki; Furuya, Masahiro*

This study aims to directly identify weld defects in ultrasonic testing (UT) images using the YOLO object detection algorithm. We evaluated model accuracy for various annotation methods. A model trained only on "defect" labels results in poor generalization on test data due to overfitting, which misidentifies normal structural echoes as defects. Introducing a "wall" label to define non-defects explicitly significantly reduced these false positives and improved performance. This demonstrates that a multi-class annotation strategy, which distinguishes defects from non-defect features, is essential for developing a reliable automated inspection system.