Building an Object Detection Model for Defects in Printed Circuit Boards using Convolutional Neural Network

The rapid advancement of electronic technology has led to significant developments in Printed Circuit Boards (PCBs), which are fundamental components in almost all modern electronic devices. Defects in PCBs can occur both during production and while the device is in operation, often resulting in device malfunctions. Traditional manual inspection methods, which rely on human expertise to examine each PCB component, are no longer efficient. Therefore, a more efficient and automated solution is required for the inspection process. This study aims to develop the model for PCB defect detection using Convolutional Neural Network (CNN) methods. The two models compared are Faster R-CNN and Mask R-CNN applied to a PCB dataset. The results indicate that the Faster R-CNN model, using hyperparameters of 40 epochs, a batch size of 1, and a learning rate of 1 × 10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−4</sup>, achieves an average Intersection over Union (avgIOU) of 79.70%, a mean Average Precision (mAP) of 81.39%, and an accuracy of 97.45%. The Mask R-CNN model, with hyperparameters of 15 epochs, a batch size of 1, and a learning rate of 1e-04, demonstrates superior performance with an avgIOU of 81.70%, an mAP of 82.02%, and an accuracy of 97.96%. Additionally, the quantized Faster R-CNN model with a resolution of 800 pixels exhibits improved performance compared to its non-quantized counterpart. The model size is reduced from 158 MB to 118 MB, and the inference time decreases from 1.1250 seconds to 1.1234 seconds. Consequently, this model will be integrated into a website for PCB defect detection. Overall, the Mask R-CNN model outperforms the Faster R-CNN model in detecting PCB defects, achieving higher avgIOU, mAP, and accuracy.