基于改进DenseNet网络的手机屏幕缺陷检测研究  

Research on mobile phone screen defect detection based on the improved DenseNet network

作　　者：崔杰 杨凯 CUI Jie;YANG Kai(College of Optical and Electronic Technology,China Jiliang University,Hangzhou 310018)

机构地区：[1]中国计量大学光学与电子科技学院,浙江杭州310018

出　　处：《中国计量大学学报》2023年第2期208-215,共8页Journal of China University of Metrology

摘　　要：目的:手机屏幕表面缺陷的传统检测方式主要由人工辨别,这严重影响生产质量把控和效率提升;所以寻求一种高效率、高准确性、高稳定性的检测方法是十分迫切的。方法:提出一种改进DenseNet网络的缺陷检测方法。选取DenseNet网络为基础网络,在模型中添加MobileNet v2网络中的倒置残差结构,以减少推理期间所需的内存占用,提高特征检测效率。同时,使用linear bottleneck可以防止非线性破坏太多信息,保证信息的完整性。结果:本文提出的算法对手机屏幕缺陷检测相较于ResNet网络和传统DenseNet有着显著的效果,准确率到达91.42%,平均精度均值到达93.48%,检测的平均时间为29.8 ms。结论:本文提出的检测方法能够很好地完成对手机屏幕缺陷的检测任务,并为工业检测提供出新的检测思路和方法。Aims:The traditional detection method of mobile phone screen surface defects is mainly manual identification,which seriously affects the production quality control and efficiency improvement.Therefore,seeking for a high efficiency,high accuracy,high stability detection method is very urgent.Methods:A defect detection method based on the improved DenseNet network was proposed.The DenseNet network was selected as the basic network;and the inverted residual structure of MobileNet v2 network was added to the model to reduce the memory occupation required during reasoning and to improve the efficiency of feature detection.At the same time,linear bottleneck prevented nonlinearity from damaging too much information,thus ensuring information integrity.Results:Compared with the ResNet network and the traditional DenseNet,the proposed algorithm had a remarkable effect on the detection of mobile phone screen defects.The accuracy rate reached 91.42%.The mean average precision(mAP)reached 93.48%;and the average detection time was 29.8 ms.Conclusions:The detection method proposed in this paper can complete the detection task of mobile phone screen defects and provide new detection ideas and methods for industrial detection.

关 键 词：缺陷检测 DenseNet网络 倒置残差结构 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]