基于深度卷积神经网络的手机玻璃盖板表面缺陷分类检测研究

Research on Classified Detection of Surface Defects of Mobile Phone Glass Cover based on Deep Convolutional Neural Network

为了实现手机玻璃盖板表面点缺陷、线缺陷和块状缺陷的分类检测,主要研究可以自动提取特征的深度卷积神经网络模型。首先针对数据缺乏问题,提出子图像划分和随机缺陷合成算法,构建了MPGC-DET数据集。为了提高模型的泛化性,以现代成熟的深度卷积神经网络模型为基础,并结合迁移学习和SE模块,搭建手机玻璃盖板表面缺陷分类检测模型。实验表明,最终分类准确率达96.40%,并且使用了Grad-CAM技术进行可视化分析,结果显示模型是根据缺陷所在区域特征进行预测的,说明没有出现过拟合现象。

In order to realize the classified detection of point defects,line defects and block defects on the glass cover surface of mobile phone,this paper proposes to study deep convolutional neural network(DCNN)model that can automatically extract features.First,to address the problem of data lack,this paper proposes to use sub-image division and random defect synthesis algorithms,and construct the MPGC-DET(Mobile Phone Glass Cover-Detection)data set.In order to improve the generalization of the model,this research takes modern and mature deep convolutional neural network model as the basis and combines transfer learning and SE modules,to build a classified detection model of mobile phone glass cover surface.Experiments show that final classification accuracy rate is 96.40%.Grad-CAM(Gradient-weighted Class Activation Mapping)technology is used for visual analysis.Results show that the proposed model performs prediction based on the characteristics of the area where the defect is located,which indicates that there is no sign of over-fitting.