<p>Defects in printed circuit boards (PCBs) can degrade the performance and reliability of electronic devices. Although YOLOv5-based algorithms are commonly used to detect PCB defects, their complex parameters slow down detection speeds on industrial platforms. This paper presents a lightweight, high-performance model for PCB defect detection, called Align Soft-Target Knowledge Distillation PCB Lightweight Defect Detection (ASTKD-PCB-LDD). The model uses the k-means++ algorithm for optimal anchor box selection and the SCYLLA-IoU (SIoU) loss function to improve accuracy in detecting small defects. The Faster-Ghost backbone network and slim-neck architecture reduce computational load and improve inference speed. Additionally, Align Soft-Target Knowledge Distillation (ASTKD) is applied, with the PCB-LDD model as the teacher and a pruned model-created using Layer-Adaptive Magnitude-based Pruning (LAMP)-as the student. This strategy helps to maintain detection accuracy while reducing model size. Experimental results show that the model size is reduced from 14.5 to 4 MB (a 27.6<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11227_2025_7045_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> reduction), achieving 98<InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11227_2025_7045_Article_IEq2.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> mean average precision (mAP), and the detection speed increases from 73.2 frames/s to 112.3 frames/s, improving by 153.4<InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11227_2025_7045_Article_IEq3.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation>. Moreover, the model demonstrates strong applicability and scalability. This approach effectively combines performance and lightweight design, significantly enhancing PCB defect detection efficiency.</p>

ASTKD-PCB-LDD: high-performance PCB defect detection model with align soft-target knowledge distillation and lightweight network design

期刊论文
发表日期：

2025年2月24日

查看全文

Zhelun Hu,
Zhao Zhang,
Shenbo Liu,
Dongxue Zhao,
Longhao Zheng,
Lijun Tang

摘要

Defects in printed circuit boards (PCBs) can degrade the performance and reliability of electronic devices. Although YOLOv5-based algorithms are commonly used to detect PCB defects, their complex parameters slow down detection speeds on industrial platforms. This paper presents a lightweight, high-performance model for PCB defect detection, called Align Soft-Target Knowledge Distillation PCB Lightweight Defect Detection (ASTKD-PCB-LDD). The model uses the k-means++ algorithm for optimal anchor box selection and the SCYLLA-IoU (SIoU) loss function to improve accuracy in detecting small defects. The Faster-Ghost backbone network and slim-neck architecture reduce computational load and improve inference speed. Additionally, Align Soft-Target Knowledge Distillation (ASTKD) is applied, with the PCB-LDD model as the teacher and a pruned model-created using Layer-Adaptive Magnitude-based Pruning (LAMP)-as the student. This strategy helps to maintain detection accuracy while reducing model size. Experimental results show that the model size is reduced from 14.5 to 4 MB (a 27.6 $\%$ $%$ reduction), achieving 98 $\%$ $%$ mean average precision (mAP), and the detection speed increases from 73.2 frames/s to 112.3 frames/s, improving by 153.4 $\%$ $%$ . Moreover, the model demonstrates strong applicability and scalability. This approach effectively combines performance and lightweight design, significantly enhancing PCB defect detection efficiency.