Comparação de métodos para identificação de defeitos em placas de circuito impresso usando redes neurais convolucionais

Abstract

This paper presents a comparative analysis between the YOLOv11 and Detectron2 deep learning models for identifying and classifying defects in printed circuit boards (PCBs). The research addresses the growing demand for Automated Optical Inspection (AOI) systems that are more efficient than traditional methods based on predefined rules, which have limitations in the face of the variability of production processes and the emergence of new types of defects. The study uses computer vision techniques based on convolutional neural networks to detect and classify common defects in PCBs, including short circuits, track interruptions, missing holes, and excess copper. The methodology employed involved training both models using a specific dataset of PCB images with annotated defects, followed by comparative evaluation through performance metrics such as mAP50 and mAP50-95. The results demonstrate that YOLOv11 achieved superior performance with mAP50 of 0.98 and mAP50-95 of 0.57, compared to Detectron2's values of 0.92 and 0.41, respectively. In terms of convergence, YOLOv11 presented a gradual and consistent learning process over 100 epochs with early stopping, while Detectron2 demonstrated faster and more stable convergence, stabilizing in approximately 2000 iterations. Detectron2 stood out for its greater predictability and stability in the training process, advantageous characteristics for industrial applications that require reproducibility. The conclusions indicate that both architectures are technically feasible for the proposed application, with the choice depending on specific factors such as available computational resources, real-time requirements, and integration needs with existing systems. The work contributes to the electronic manufacturing industry by providing data-based support for decision-making on the implementation of defect detection technologies in production lines, aiming at improving quality, reducing operational costs and increasing productivity.