Lightweight intelligent detection algorithm for surface defects in printed circuit board

• Published in: Computers & industrial engineering Vol. 203; p. 111030
• Main Authors: Zhao, Xiaolong, Zhang, Haifeng, Song, Chonghui, Li, Hao, Guo, Hongyi
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.05.2025
• Subjects: Circuit board; Defect detection; Gather-distribute mechanism; Multi-branch streaming convolution; Circuit board; Gather-distribute mechanism; Multi-branch streaming convolution; Defect detection
• ISSN: 0360-8352
• DOI: 10.1016/j.cie.2025.111030

Printed circuit boards (PCBs) are the core components of electronic devices, and deep learning-based image recognition technology effectively diagnoses defects, ensuring product quality and reliability. To address the challenges of small defect detection and model lightweighting, this paper introduces a lightweight PCB surface defect detection model (PSDDNet). Firstly, a multi-branch streaming convolution (MSC) is designed to aggregate features through continuous convolution and pooling, capturing rich gradient flow information to improve the receptive field and feature representation capabilities. Secondly, a simplified GDLite architecture is designed for feature fusion, utilizing the gather-distribute mechanism to optimize cross-layer information interaction, thereby avoiding feature information loss and confusion, and reducing model complexity. Additionally, the feature extraction architecture is optimized to better focus on small target information, and a lightweight coordinate attention (CA) module is introduced to enhance feature expression capabilities. Extensive experiments on three PCB datasets demonstrate the superiority of PSDDNet, showing a better balance of detection precision and speed compared to other state-of-the-art algorithms. On the PKU-Market-PCB dataset, PSDDNet achieves an inference speed of 65 FPS, while obtaining 98% mAP and 97.7% recall with only 0.9 million parameters. These experiments prove that PSDDNet is a reliable and competitive model, providing a feasible solution for real-time PCB defect detection in industrial applications. •Proposes lightweight PSDDNet, optimizing feature extraction and fusion structures.•Designs MSC module with rich gradient flow for better image information.•Designs C2fe module for richer image features and accelerated model convergence.•Designs lightweight fusion architecture GDLite with gather-distribute mechanism.•GDLite avoids cross-layer feature information loss and enhances feature fusion.