LE-YOLO for Lightweight Tiny Defect Detection on Printed Circuit Boards

Abstract

It is challenging to detect tiny surface defects on printed circuit boards (PCBs) due to their small size, elongated features, and the complex textured background. This paper develops Lightweight Edge-guided YOLO(LE-YOLO), a lightweight detector based on YOLOv8n. First, a lightweight cross-layer output reconstruction strategy removes the P5 stage, retains P2, P3, and P4 detection scales, and introduces a shallow backbone feature into the P3 fusion stage to preserve details while compressing the model. Second, an edge-guided lightweight attention module combines a local high-frequency edge branch with an efficient channel branch to enhance P2 features. On the public PCB dataset, the five-run mean precision, recall, mAP50, and mAP50-95 are 95.52 percent, 90.20 percent, 93.36 percent, and 50.02 percent, respectively, and the sample standard deviation of mAP50-95 is 0.44 percentage points. The model contains only 0.989 million parameters and requires 10.50 GFLOPs. Compared with YOLOv8n, it improves mAP50-95 by 1.82 percentage points while reducing the parameter count by 67.11 percent.