基于知识蒸馏的PCB板缺陷检测方法  

PCB Board Defect Detection Method Based on Knowledge Distillation

作　　者：陈政宇 张峰[1] 张士文[1] Chen Zhengyu;Zhang Feng;Zhang Shiwen(School of Electronic Information and Electrical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海交通大学电子信息与电气工程学院,上海200240

出　　处：《电气自动化》2024年第6期113-116,共4页Electrical Automation

摘　　要：为了提升PCB板表面缺陷检测的效率,提出了一种基于知识蒸馏的PCB板缺陷检测方法。搭建了一大一小两个网络模型,大模型检测精度高,检测效率低;小模型检测效率高,但检测精度低。使用知识蒸馏方法将大模型的“知识”蒸馏至小模型中,蒸馏不改变小模型的检测效率,但会使小模型的检测精度获得巨大提升。试验结果表明:大模型在交并比为0.5时的平均精度高达0.938,经知识蒸馏,模型的检测效率在RTX4090上帧率可达454.5。所提方法成功实现了对印刷电路板缺陷的高精度高效检测,具有重要意义。In order to enhance the efficiency of surface defect detection on PCB boards,a PCB board defect detection method based on knowledge distillation was proposed.Two network models were built:a large one with high accuracy but low efficiency,and a small one with high efficiency but lower accuracy.Using knowledge distillation method to distill the“knowledge”of the large model into the small model did not change the detection efficiency of the small model,but it would greatly improve the detection accuracy of the small model.The experimental results show that the average accuracy of the large model is as high as 0.938 while the intersection to union ratio is 0.5.After knowledge distillation.The detection efficiency of the small model on RTX4090 can reach a frame rate of 454.5.The proposed method has successfully achieved high-precision and efficient detection of defects in PCBs,which is of great significance.

关 键 词：PCB板检测 缺陷检测 模型压缩 知识蒸馏 

分 类 号：TP274.2[自动化与计算机技术—检测技术与自动化装置]