Real-Time Terahertz Characterization of Minor Defects by the YOLOX-MSA Network
Xingyu Wang, Zhen Zhang, Yafei Xu, Liuyang Zhang, Ruqiang Yan, Xuefeng Chen
Abstract
Terahertz (THz) imaging has been widely used in industrial non-destructive testing (NDT) of nonpolar materials owing to its unique property. Minor defect detection via THz NDT at high accuracy and fast speed is essential for industrial on-line detection systems. However, traditional defect detection algorithms can’t meet the demand of real-time high-precision detection of minor defects. Therefore, based on the YOLOX algorithm and multi scale attention (MSA) mechanism, the modified YOLOX network called YOLOX-MSA is proposed as a one-stage minor defect detection framework to improve the detection accuracy while supporting the real-time operation. The proposed YOLOX-MSA network improves the mean average precision (mAP) by at least 11.65% on the PCBs dataset with THz characteristics when the Intersection over Union (IoU) is 0.75. Additionally, the proposed algorithm can reach to the detection speed as 24~25 Frame Per Second (FPS). Overall, our proposed method can be beneficial to generalize the THz NDT in the frequency domain on the minor defects of nonpolar material, which will fulfill the impending requirements of real-time defect detection for industrial applications.