基于孪生网络的工业缺陷弱监督视觉检测算法  

Weakly supervised visual detection algorithm for industrial defects based on Siamese network

作　　者：张海刚 鲁伽祎 匡国文 陈志彬[2] ZHANG Haigang;LU Jiayi;KUANG Guowen;CHEN Zhibin(Institute of Applied Artificial Intelligence of the Guangdong-Hong Kong-Macao Greater Bay Area,Shenzhen Polytechnic University,Shenzhen 518055,Guangdong Province,P.R.China;School of Electronic and Information Engineering,University of Science and Technology Liaoning,Anshan 114051,Liaoning Province,P.R.China)

机构地区：[1]深圳职业技术大学粤港澳大湾区人工智能应用技术研究院,广东深圳518055 [2]辽宁科技大学电子与信息工程学院,辽宁鞍山114051

出　　处：《深圳大学学报（理工版）》2023年第6期657-664,共8页Journal of Shenzhen University(Science and Engineering)

基　　金：深圳职业技术学院深圳市高端人才科研启动资助项目(6021310030K);深圳职业技术学院科研基金资助项目(6022310006K);广东省教育厅重点领域专项资助项目(2020ZDZX3082,2023ZDZX1081,2023KCXD077)。

摘　　要：工业品缺陷存在目标多尺度和随机性等特点,易导致现有的目标检测算法出现误检或漏检现象.基于卷积神经网络架构,提出一种高性能的工业缺陷视觉检测模型,记为SNDec(Siamese network detection).采用并行权重共享的孪生网络(Siamese network, SN)将工业缺陷转化为视觉差异特征.孪生网络由并行特征提取通道构成,并以两通道的差异特征作为输出,能够在抑制同类属性的前提下,最大程度地凸显缺陷特征.结合弱监督定位算法,所提模型能够在实现高精度识别的同时,获取工业缺陷发生位置.通过引入卷积块注意力机制(convolutional block attention module, CBAM),进一步提升了模型检测精度.在真实工况采集的注塑瓶盖数据集以及公开的MVTec数据集上,将当前主流的工业品缺陷视觉检测算法(ResNet 50、1-NN、U-Student和GANomaly)与SNDec模型进行比较.结果表明,SNDec模型取得了89.2%的分类准确率和90.1%的召回率,比ResNet50模型分别提高了5.7%和3.2%.仿真结果验证了基于差异特征实现工业缺陷检测的有效性,结合弱监督定位算法Grad-CAM,SNDec模型能够以热力图形式实现准确的工业缺陷定位,为工业缺陷特征刻画提供更为丰富的信息.The characteristics of industrial product defects,such as multi-scale targets and randomness,can easily lead to false or missed detection in target detection algorithms.Based on the convolutional neural network architecture,we propose a high-performance visual detection framework for industrial defects.A parallel weight sharing Siamese network is used to transform industrial defects into visual difference features.Siamese network is composed of parallel feature extraction channels,with difference features of two channels as output.On the premise of suppressing homogeneous properties,Siamese network can highlight defect features to the greatest extent.Combined with weakly supervised localization algorithms,the proposed model can achieve high-precision recognition while obtaining the location of industrial defects.In addition,the model introduces the convolutional block attention module,which further improves the detection accuracy of the proposed model.In the experimental stage,the mainstream industrial defect detection algorithms(ResNet50,1-NN,U-Student and GANomaly)were compared with the proposed model by using the injection bottle cap dataset collected from real operating conditions and publicly available MVTec dataset.The model achieves a classification accuracy of 89.2%and a recall rate of 90.1%,which are 5.7%and 3.2%higher than the ResNet50 model,respectively.The simulation results have verified the effectiveness of industrial defect detection based on differential features.The weakly supervised positioning algorithm can effectively determine the location of target defects and provide richer information for characterizing features of industrial defects.

关 键 词：人工智能 模式识别 计算机神经网络 计算机图象处理 工业视觉 孪生网络 卷积神经网络 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] TP183[自动化与计算机技术—控制科学与工程] TP391