Comparative Analysis of Deep Learning Models for PCB Defects Detection and Classification

Abstract

In the modern-day, just about any electronic device uses Printed Circuit Boards (or PCBs). Nevertheless, these are still prey to manufacturing defects, which are extremely hard to detect manually. Other researchers have come up with a slew of ways to improve PCB inspection and problem categorization. As a result, fewer innovative approaches in this sector are being developed as a result of the researchers' failure to disclose their datasets before. The Weibo Huang and Peng Wei study titled 'A PCB Dataset for Defects Detection and Classification'[1] was used to produce a synthetic PCB dataset. The dataset consists of 1386 photos with six types of faults for the detection, classification, and registration of tasks. As part of the evaluation, it looks at how well various convolutional neural network designs diagnose errors using the reference-based technique (proposed in the same research that generated this dataset). When compared to traditional techniques that need pixel-by-pixel processing, a novel method first locates the faults and then classifies them using the selected neural network, which performs better on the dataset than previous methods. This method also has a proposed neural network [1] whose performance is tested against popular models like Inception and VGG using transfer learning.