Enhancing Industrial PCB and PCBA Defect Detection: An Efficient and Accurate SEConv-YOLO Approach

Abstract

Real-time accurate defect detection presents a critical bottleneck in high-throughput Printed Circuit Board Assembly (PCBA) manufacturing. Although contemporary state-of-the-art (SOTA) object detection models excel on diverse general-purpose datasets, their inherent architectural complexity and over-parameterization often render them ill-suited for the highly standardized visual conditions and stringent real-time performance requirements of industrial Printed Circuit Board (PCB) defect inspection. This study addresses this significant performance gap by introducing SEConv-YOLO, a streamlined and powerful object detection model that is meticulously tailored for PCBA Wire defect detection (broken, sweeping, and missing wires). Our approach pioneers a trifecta of architectural innovations: (1) a lightweight Squeeze Excite Convolution (SEConv) feature extraction module, designed to efficiently capture salient defect characteristics while minimizing computational overhead; (2) an advanced Weighted Residual Spatial Pyramid Pooling (WRSPP) feature fusion neck, engineered to enrich contextual information without introducing significant computational bloat; and (3) the novel application of the Normalized Complete Intersection over Union (N-CIoU) loss function, strategically implemented to accelerate model convergence and significantly enhance localization accuracy, particularly for small defects. Rigorous experimentation demonstrates that SEConv-YOLO not only substantially surpasses the baseline YOLOv8n in both speed and accuracy but also consistently outperforms other leading SOTA detectors. This study culminates in a robust, production-ready solution that establishes a new benchmark for automated quality control in PCBA manufacturing, thereby enabling enhanced product reliability and increased operational efficiency.