摘要
印制电路板(Print Circuit Board,PCB)表现形式多样,缺陷特征表征困难。针对印制电路板缺陷类别识别难度较大等问题,提出了一种基于卷积神经网络融合多尺度特征的PCB裸板缺陷识别方法。该方法首先提取PCB裸板缺陷的多尺度灰度共生矩阵特征、多尺度方向投影特征以及多尺度梯度方向直方图特征,构建缺陷浅层图像特征,然后基于迁移学习,利用VGGl6-Net预训练神经网络模型的特征提取网络,提取PCB裸板缺陷图像深度语义特征,将得到的浅层图像特征与深度语义特征进行融合,最后将特征向量以特征序列方式输入给支持向量机进行分类识别。试验结果表明,融合深度特征和多尺度浅层特征的算法相较于传统卷积神经网络算法,对PCB裸板缺陷具有较高的识别率。
Printed circuit board(PCB) has various manifestations,and it is difficult to characterize the defects.To address the problem of difficult identification of printed circuit board defect categories,a method for identifying PCB bare board defects based on multi-scale CNN is proposed.The method first extracts the multi-scale grayscale covariance matrix features,multi-scale directional projection features and multi-scale gradient directional histogram features of PCB bare board defects to construct the shallow image features of defects.Then the feature extraction network based on transfer learning is used to extract the depth semantic features of PCB bare board defect images,and the obtained shallow image features are fused with the deep semantic features.Finally,the feature vectors are put into the SVM for classification and recognition.The experimental results show that the algorithm which fuses deep features and multi-scale shallow features has higher recognition rate for PCB bare board defects than the traditional CNN algorithm.
作者
李任鹏
李云峰
LI Renpeng;LI Yunfeng(School of Mechatronics Engineering,Henan University of Science and Technology,Luoyang Henan 471003,China)
出处
《智能计算机与应用》
2023年第10期65-72,共8页
Intelligent Computer and Applications
关键词
缺陷识别
特征提取
迁移学习
缺陷检测
特征融合
defect recognition
feature extraction
transfer learning
defect detection
feature fusion
