Automated Detection and Classification of Defects in Pcb Using Deep Learning Techniques: A Comparison of Resnet50v1 and Efficient Det D1

To keep pace with the fast-growing Printed Circuit Board (PCB) industry, automated processes are necessary for producing flawless products. Automated visual PCB inspections have become indispensable for addressing the challenges posed by conventional manual inspection systems, as they offer faster, more quantitative, and objective assessments. The aim of this research is to determine an appropriate and effective Convolutional Neural Network (CNN) algorithm for detecting and classifying different PCB defects. Two CNN algorithms, namely Single Shot Detector-Residual Neural Network 50 V1 Model (SSD ResNet50v1) and Efficient Object Detector Model D1 (EfficientDet D1), are evaluated for their classification performance, and their results are compared. The training of these models is conducted on Google Colaboratory, a cloud-based service that provides a free online Jupyter notebook environment for deep learning model training on NVIDIA's Graphical Processing Unit (GPU). The PCB dataset comprises 10,668 images of size 600 × 600 with six common PCB defects frequently found in the industry. The study calculates the classification performance of both algorithms and presents their results. The accuracy achieved by SSD ResNet50v1 is 81.68%, while EfficientDet D1 reaches 71.75%. A comparison of both architectures' accuracies shows that SSD ResNet50v1 performs better than EfficientDet D1.