Defect Detection Based on Improved YOLOx for Ultrasonic Images

Non-destructive testing is a method of detecting defects in materials or electronic components without causing damage to the detected objects. The most commonly used detection technology is ultrasonic detection. However, for images generated by ultrasonic inspection, manual recognition and traditional image processing methods are mostly used for defect identification, which are both inefficient and costly. The detection of defects in printed circuit boards (PCB) is a particularly difficult problem in the field of industrial inspection, which has strict requirements. We adopt a deep learning method to implement intelligent defect detection in our work. To address the lack of training data, we collect PCB surface images using a high-resolution ultrasonic microscope and create a dataset by annotating the defects in the images. The final dataset can be used for defect object detection based on deep learning. Furthermore, we propose an improved object detection method for defect detection that adopts the four-scale Swin Transformer as the multi-scale feature extraction network and uses the decoupled head from YOLOx to output defect categories and locations. To better learn the defect features, we pretrain on datasets of PCB images obtained using other methods, such as charge-coupled devices and CMOS sensors. Subsequently, we transfer to our own created dataset to perform training and testing. Experimental results show that our improved model achieves an average precision of 99.9% on our PCB test dataset, and an average precision of 85.1% on PASCAL VOC 2007 test dataset while extending to the conventional object detection.