A comprehensive review of fast computing approaches for real-time image defects in smart manufacturing of PCB electronics

Abstract The fast growth of Industry 4.0 and Industry 5.0 is causing an increasing need for high-quality electronics and, consequently, defect-free Printed Circuit Boards (PCBs). Traditional computer vision-based inspection systems cannot meet modern manufacturing environment demands since these systems fail to meet the requirements of real-time processing. The present paper provides a review and analysis of existing PCB defect detection technologies, paying attention to four areas: deep learning-based defect detection algorithms, specialised computing platforms, hardware acceleration approaches, and their hybrids. Among the available deep learning approaches, the You Only Look Once (YOLO) framework stands out as a high-performing method, achieving high detection accuracies while enabling real-time defect detection. The specialized computing platforms, particularly Field-Programmable Gate Arrays (FPGAs), have demonstrated considerable potential due to their low-latency processing capabilities and high energy efficiency. However, several challenges remain unresolved, including limited generalization across different defect categories, the scarcity of adequately annotated datasets, and insufficient validation under real-world operational conditions. To overcome these challenges, this paper suggests adopting a comprehensive framework that combines image acquisition, image pre-processing, deep learning-based defect detection, and an adaptive system updating process. Further directions for future studies may include utilizing physics-based modeling for better defect representation, improving model explainability via data-centric methods, and exploring the concept of model-less or self-adapting neural networks. This will enable the development of highly efficient and real-time PCB quality inspection frameworks for use in industry.