Defect detection for printed circuit board (PCB) based on deep learning requires a large number of well-labeled datasets for model training. There are indeed some datasets on defect detection for PCB, but most of these datasets only contain surface images collected by optical inspection. Up to now, there is almost no ultrasonic dataset for the defect detection for PCB, which is in high demand for industrial ultrasonic inspection. To solve the problem of insufficient data for ultrasonic defect detection of PCB, we use a high-resolution ultrasonic microscope developed by our laboratory for data acquisition to collect data in C-scan, and then construct a standard PCB ultrasonic dataset named U-PCBD for defect detection.

U-PCBD consists of two versions: version 1 and version 2.

• Version 1: This version of the dataset is obtained by scanning surfaces of 20 PCB samples, which includes 200 original images of 640 * 640 pixels. It is then expanded to 4320 images by data augmentation including flipping, rotating, brightening, blurring and noising. There are 14346 defect annotations in the final dataset. The detailed information is presented in Table 1. The training set includes 3456 images, and test set includes 864 images.
  

  Table 1. Details of Dataset for Version 1

  Type of Defect	Number of Annotations	Number of Images
  Mouse Bite	2214	1800
  Open Circuit	2430	1890
  Short	2712	2280
  Spur	4788	3030
  Spurious Copper	2202	1986

• Verion 2: This version of the dataset is obtained by scanning the surfaces and internal layers of 50 multi-layer PCB samples, which includes 425 original surface images of 640 * 640 pixels and 84 original internal layer images of 640 * 640 pixels. It is then expanded to 15300 surface images and 3024 internal images by data augmentation including flipping, rotating, brightening, blurring and noising. There are 47880 defect annotations in the final dataset. The detailed information is presented in Table 2. The training set includes 14660 images, and test set includes 3664 images.
  

  Table 2. Details of Dataset for Version 2

  Images	Type of Defect	Number of Annotations	Number of Images
  Surface	Mouse Bite	5622	5064
  	Open Circuit	7272	6588
  	Short	7068	6162
  	Spur	10764	7962
  	Spurious Copper	5016	4116
  Internal	Mouse Bite	1944	1272
  	Open Circuit	2598	2130
  	Short	1836	1476
  	Spur	3780	2340
  	Spurious Copper	1980	1938

The above two versions of the dataset are labeled using LabelImg, which is publicly available at https://github.com/heartexlabs/labelImg. There are five types of defects including spur, open circuit, spurious copper, short and mouse bite, as shown in the following figure. Each image contains 3 to 5 defects. Since there are few defects in the natural state, we use manual addition of defects to increase the number of defect samples. Both versions of the above dataset are organized in VOC format.

Example of Defects

We make U-PCBD dataset available for academic research purposes. To access the dataset for research (noncommercial use), please send an email to luk@ucas.ac.cn with a signed AGREEMENT. Please note that any commercial use of the dataset is prohibited.