Improving PCB Defect Detection with KAN-YOLO: A Knowledge-Enhanced Feature Fusion Approach

This paper introduces an advanced iteration of the YOLOv8 framework tailored for identifying defects in printed circuit boards (PCBs). We present KAN-YOLO model, which incorporates a series of novel components, including C2FKAN, BottleKAN, Zoomcat, SSE, and KAGNConv2DLayer to tackle the complexities associated with precise and efficient defect detection in PCB imagery. These components bolster the model's capacity to capture pertinent features, boost detection efficacy, and strengthen resilience against intricate defect categories such as missing holes, open circuits, and extraneous copper. Empirical assessments on a dataset comprising 1,725 training images and 414 validation images reveal that the KAN-YOLO model surpasses conventional object detection techniques in both accuracy and processing speed. Additionally, the model's proficiency in managing real-time PCB inspection tasks has been confirmed, positioning it as a viable option for automated quality assurance in manufacturing settings. The paper also explores future research directions, focusing on enhancing model scalability, refining data augmentation methods, and exploring cross-domain adaptation for wider industrial use cases.