Benchmark Study on YOLOv8 Variants in Localized Multiclass Fault Detection in PCBs

Detecting surface faults on PCBs is critical for ensuring device dependability and performance. Manual PCB inspection methods like X-ray imaging struggle to ensure quality control due to limitations in speed and defect characterization. However, advanced automated technologies, such as object detection models like YOLOv8, have emerged as a more efficient solution. This study focuses on evaluating the performance of different YOLOv8 variants in identifying and classifying various defects on PCBs. A benchmark of how each modification of the model performs detection and the factors behind were investigated here. It was determined that, YOLOv8 is sensitive to the amount of data it is being trained on, as the larger models tend to overlearn on smaller datasets. YOLOv8-m had the best performance, 97.1% precision, concerning the time taken and epoch count, over others. Large YOLOv8 variants, such as YOLOv8x, are also computationally expensive. The models were trained for 200 epochs with a patience value set at 50. Both increasing the number of data points and the resources will improve the performance and facilitate the overall training procedures. A significant gap in automated electronics production is addressed by this study: precise PCB fault localization for higher reliability of the product.