Analysis of the various YOLOv8 models in Detecting Defects in PCB

In the arena of electronics, the Printed Circuit Board (PCB) forms the cornerstone of every component, extending from everyday gadgets to advanced defense equipment. The escalating pace of modern life necessitates accelerated PCB production, calling for swift and precise mechanisms to ensure defect-free output for reliable manufacturing. Though such quality control tasks were traditionally manual, the advent of machine learning and deep learning has ensured automated and reliable solutions. A standout among these is YOLO, a speedy computer vision algorithm renowned for its unmatched speed and accuracy. This paper seeks to conduct an all-encompassing study of the cutting-edge YOLOv8 sub-models, delving into their individual performance attributes. The work encapsulates the implementation of these YOLOv8 models, followed by a rigorous analysis of their results on a range of metrics like $\mathrm{mAP50}, \mathbf{m A P 5 0 - 9 5}$, precision, and recall. An integral part of this study is the preprocessing of the HRIPCB dataset, comprising six distinct defect categories, namely Missing hole, Mouse bite, Spur, Spurious copper, Open circuit, and Short. The ensuing robust methodology involves comprehensive data preprocessing, data cleansing, and standardization. Ensuring a high-level of data quality, image enhancement techniques such as Colour Balance Adjustment, Histogram Equalization, and Image Filtering - Sharpening are employed. The study shines a spotlight on the individual training and testing processes of the YOLOv8 sub-models, namely v8n, v8s, v8m, v8l, and v8x, leveraging various hyperparameters to optimize model performance. By surpassing the ${9 5. 8 \%}$ mAP50 achieved by earlier YOLO versions, this research aspires to revolutionize PCB defect detection. It paves the way for enhanced quality control mechanisms in PCB manufacturing, thus contributing to the advancement of electronics production regimes