Scholarlink

<p>The detection of subtle Printed Circuit Board (PCBs) defects, particularly Spur (spurious copper) and Mouse bite (micron-scale notches), remains challenging due to their minimal dimensions and interference from complex textures.Existing methods suffer from low sensitivity to micro-defects and degraded robustness under data variations like random noise injection, color shifts,and scaling.To address these issues, this paper proposes SNF-YOLOv8(Scale-aware Neural Fusion You Only Look Once Version8),an enhanced framework built upon YOLOv8n,integrating a lightweight Scale-aware Multi-level Feature Fusion (SMFF) strategy and an Attention-based Scale Sequence Fusion (ASF) mechanism.The SMFF optimizes bidirectional cross-layer connections and hierarchical feature pyramids to retain fine-grained details of micro-defects,while the ASF employs LSTM networks and dynamic channel-spatial attention to adaptively refine multi-scale feature weights.Experiments on a high-quality PCB defect dataset (2,772 images) demonstrate that SNF-YOLOv8 achieves a 90.0% mAP50,surpassing YOLOv8n by 3.1%,with a 17% reduction in parameters (2.49M) and real-time inference at 250 FPS. Cross-dataset validation on DeepPCB further confirms its robustness,achieving 98.4% mAP.The framework significantly improves detection accuracy for micron-scale defects like mouse bites (7.2% AP gain) and suppresses false alarms in complex backgrounds.This work provides a computationally efficient and reliable solution for industrial PCB inspection, advancing lightweight deep learning applications in automated quality control.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SNF-YOLOv8: A Lightweight PCB Defect Detection Algorithm base on Multiscale Feature Fusion and Attention Scale Sequence Fusion

  • 期刊论文

  • 发表日期：

    2025年8月23日

查看全文
  • Song Xudong,
  • Song Xiumin

摘要

The detection of subtle Printed Circuit Board (PCBs) defects, particularly Spur (spurious copper) and Mouse bite (micron-scale notches), remains challenging due to their minimal dimensions and interference from complex textures.Existing methods suffer from low sensitivity to micro-defects and degraded robustness under data variations like random noise injection, color shifts,and scaling.To address these issues, this paper proposes SNF-YOLOv8(Scale-aware Neural Fusion You Only Look Once Version8),an enhanced framework built upon YOLOv8n,integrating a lightweight Scale-aware Multi-level Feature Fusion (SMFF) strategy and an Attention-based Scale Sequence Fusion (ASF) mechanism.The SMFF optimizes bidirectional cross-layer connections and hierarchical feature pyramids to retain fine-grained details of micro-defects,while the ASF employs LSTM networks and dynamic channel-spatial attention to adaptively refine multi-scale feature weights.Experiments on a high-quality PCB defect dataset (2,772 images) demonstrate that SNF-YOLOv8 achieves a 90.0% mAP50,surpassing YOLOv8n by 3.1%,with a 17% reduction in parameters (2.49M) and real-time inference at 250 FPS. Cross-dataset validation on DeepPCB further confirms its robustness,achieving 98.4% mAP.The framework significantly improves detection accuracy for micron-scale defects like mouse bites (7.2% AP gain) and suppresses false alarms in complex backgrounds.This work provides a computationally efficient and reliable solution for industrial PCB inspection, advancing lightweight deep learning applications in automated quality control.