Real-time Minor Defect Recognition of Pseudo-Terahertz Images via the Improved YOLO Network

Terahertz (THz) imaging has been widely used in non-destructive testing (NDT) of nonpolar materials owing to its unique properties of remarkable accuracy. However, THz imaging has been suffered from serious constraints such data deficiency, low spatial resolution, blurred contour and high background noise due to the limitation of THz wavelength and the agonizingly delayed development of THz devices. Here we have proposed a degradation model to generate massive PCB images with THz characteristics (named as PCB Pseudo-THz images) to overcome the shortcoming of the deficient dataset of THz imaging. Then, the modified YOLO V4 network is proposed to precisely identify four different types of defects on the PCB board. Moreover, the concept of transfer learning is also implemented to improve the detection and classification accuracy of various types of defects. The proposed model can not only obtain accurate detection of minor defects in the PCB samples that are inaccessible by human eyes, but also achieve the real-time fault classification and location. 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 the industrial applications.