Detection of PCB Socket Defects: A Comparative Analysis of Transfer Learning, Semantic Segmentation and Feature Optimization Techniques

The reliability and functionality of electronic components are critically important for efficiency in industrial production. As one of the fundamental components, printed circuit boards (PCBs) playa crucial role in the operational performance of electronic devices. Any defects in PCBs can lead to operational disruptions and increased costs. Specifically, sockets where sensors are connected are among the most susceptible areas for defects, and accurate and effective detection of these defects is vital for maintaining product quality and operational efficiency. Traditional defect detection methods typically rely on manual inspections, which are time-consuming and prone to human error, limiting their reliability. Advances in machine learning and image processing offer promising alternatives for automated defect detection. In this study, the performances of transfer learning (Googlenet) and semantic segmentation (U-Net) were compared for the classification of socket defects on a PCB. The dataset used in the study was specifically created for this research. The performances of three different models Googlenet, Googlenet+Relieff, and U-Net were evaluated. The Googlenet+Relieff model was identified as the most successful with an impressive accuracy rate of 98.6% and a classification time of 0.274 seconds. This high success rate demonstrates the effectiveness of combining feature selection algorithms with transfer learning. The results indicate that transfer learning has significant potential for improving industrial quality control processes.