System for Troubleshooting Welded Printed Circuit Boards with Through Hole Technology Using Convolutional Neural Networks and Classic Computer Vision Techniques

Manual inspection in printed circuit board manufacturing is highly susceptible to failure through human error. This opens the way for automated visual inspection. Several methods exist for detection based on images captured by a camera. The objective of this work is to develop a computer vision system using convolutional neural networks and classical computer vision techniques for locating soldering faults on printed circuit boards with through-hole technology. For this purpose, the OpenCV library on Python is used to detect the region of interest within the image prior to the analysis and classification of the convolutional neural network ResNET50. Two types of faults were presented as lack of solder and solder bridge. The results obtained in the experimental classification tests have an accuracy margin higher than 90%. This makes a viable use of automated visual inspection in the testing and inspection processes of errors in the soldering of printed circuit boards. The dataset is available at: https://github.com/asrf001/DatasetPCB.git.