Printed Circuit Board Defect Detection Using Deep Learning via A Skip-Connected Convolutional Autoencoder.

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Increasing complexity of printed circuit boards (PCBs) necessitates robust quality control.
• Automated PCB surface inspection is crucial due to the significant impact of small defects.
• Traditional machine vision struggles with setting pass/fail criteria for imbalanced defect datasets.

Purpose of the Study:

• To develop an advanced automated PCB inspection system.
• To address challenges in defect detection with small and imbalanced datasets.
• To improve the accuracy and reliability of PCB surface inspection.

Main Methods:

• Implementation of a skip-connected convolutional autoencoder for defect detection.
• Training the autoencoder to reconstruct non-defect images from defect images.
• Utilizing image augmentation techniques to enhance model performance on limited data.

Main Results:

• The unsupervised autoencoder achieved a high detection rate of up to 98%.
• A low false pass rate below 1.7% was recorded on test data.
• The system demonstrated promising performance on a dataset of 3900 images.

Conclusions:

• The proposed convolutional autoencoder-based system offers an effective solution for automated PCB inspection.
• The method successfully overcomes limitations associated with small and imbalanced datasets.
• This approach significantly enhances the quality control process in PCB manufacturing.