样本限制场景下金属板带材表面缺陷分类研究  

Classification of surface defect of sheet metal strip in sample restricted scene

作　　者：刘梦婕 王剑平[1] 张果[1] 王海云 罗付华 LIU Mengjie;WANG Jianping;ZHANG Guo;WANG Haiyun;LUO Fuhua(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming,Yunnan 650500,China;Panzhihua Steel and Vanadium Co.,Ltd.Hot-rolled Plate Plant,Panzhihua Iron and Steel Group,Panzhihua,Sichuan 617023,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650500 [2]攀钢集团攀枝花钢钒有限公司热轧板厂,四川攀枝花617023

出　　处：《光电子．激光》2021年第7期711-718,共8页Journal of Optoelectronics·Laser

基　　金：国家重点研发计划资助(2017YFB0306405);国家自然基金资助(61364008);昆明理工大学复杂工业控制学科方向团队建设计划资助和云南省重点研发计划(2018BA070)资助项目。

摘　　要：深度学习在金属板带材表面缺陷检测中取得良好的检测效果,但随着网络层数的增加,针对较小样本的金属板带材表面缺陷数据集训练数据容易出现过拟合现象的问题,为此将残差网络与迁移学习结合提出了一种融合多层次缺陷特征的图像分类算法。该算法采用残差网络模块逐层提取金属表面缺陷特征,获得丰富的位置信息和语义信息缺陷特征的特征图,后续利用分类网络基于该融合特征图得到最终分类结果,同时对特征提取网络进行迁移学习,增加网络泛化能力,优化分类精度。利用钢带表面缺陷检测数据集评估本文算法性能,实验结果表明,提出的算法具有较好的分类效果,优于其他缺陷分类算法,分类准确率可达到99.07%,同时本文所提算法具有良好的抗噪性和泛化性,在金属板带材表面缺陷智能检测中具有较好的应用价值。Deep learning has achieved good detection results in the detection that surface defect detection of Sheet Metal Strip.However,as the number of network layers increases,the training data of strip surface defect data set for smaller samples is prone to overfitting.For this reason,combining the residual network and transfer learning,an image classification algorithm combining multi-level defect features is proposed.The algorithm uses the residual network module to extract the metal surface defect features layer by layer,obtains a feature map of rich location information and semantic information defect features,and then uses the classification network to obtain the final classification result based on the fusion feature map,at the same time,migration learning is performed on the feature extraction network to increase the generalization ability of the network and optimize the classification accuracy.Using the steel strip surface defect detection data set to evaluate the performance of the algorithm in this paper,the experimental results show that the proposed algorithm has a better classification effect than other defect classification algorithms,and the classification accuracy rate can reach 99.07%.Simultaneously,the algorithm has good noise resistance and generalization,and has good application value in the intelligent detection of surface defects of metal plates and strips.

关 键 词：缺陷分类 小样本 残差网络 迁移学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]