基于轻量级网络的PCB元器件检测  

PCB Component Detection Based on Lightweight Network

作　　者：产世兵 刘宁钟[1] 沈家全 CHAN Shi-bing;LIU Ning-zhong;SHEN Jia-quan(School of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China)

机构地区：[1]南京航空航天大学计算机科学与技术学院,江苏南京211106

出　　处：《计算机技术与发展》2020年第10期14-20,共7页Computer Technology and Development

基　　金：国家自然科学基金(61375021)。

摘　　要：随着电子工业的迅速发展,电路板元器件的缺陷检测愈加重要。传统的人工检测方法效率很低,而且容易因为视觉疲劳造成错误检测,可靠性低,速度慢。目前广泛应用的自动光学检测设备,缺点明显,速率低,对直插元器件的检测精度低,无法适应电路板元器件的多样性检测。随着对卷积神经网络的深度研究,神经网络在目标检测方面已经达到了优秀的效果,但是常见的网络对PCB元器件中的小目标以及实时检测并不理想。对基于Faster RCNN和PeleeNet网络的研究,实现了轻量级小目标检测模型;通过先验知识修改了RPN网络的包围框大小;针对PCB元器件样本的小目标样本少的问题,利用了小目标样本增广技术,提高了整体的检测速度以及精度。通过消融实验体现了改进部分对PCB元器件实时检测的重要性;通过对比实验,该方法在保证检测精确度降低很小的同时,缩小了模型的大小,在数据集上具有0.858的mAP,检测时间为0.034 s,相比Faster RCNN(基础网络为VGG16或ResNet50)的检测速度有了不错的提高。With the rapid development of electronic industry,defect detection of circuit board components has become increasingly important.Traditional manual detection method is inefficient and easy to cause error detection due to visual fatigue,low reliability and slow speed.At present,the widely used automatic optical detecting equipment has obvious disadvantages,low speed,low detection precision for the components directly inserted,which cannot be adapted to the diversity detection of the circuit board components.With the in-depth study of convolutional neural networks,the neural network has achieved excellent results in target detection,but the common network is not ideal for small targets and real-time detection in PCB components.Based on the research of Faster RCNN and PeleeNet network,the lightweight small target detection model is realized.The bounding box size of the RPN network is modified by prior knowledge.Aiming at the problem of small target samples for PCB component,the small target sample augmentation technology is used to improve the overall detection speed and accuracy.Through the ablation experiment,the importance of the improved part to the real-time detection of PCB components is reflected.Through the comparison experiment,the proposed method greatly reduces the size of the model under the premise of ensuring that the detection accuracy does not change much.In the data set,the mAP is 0.858,and the detection time is 0.034 s,which is a much higher rate than the Faster RCNN(VGG16 or ResNet50).

关 键 词：PCB元器件 卷积神经网络 轻量级网络 小目标检测 实时检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]