【摘要】 由于缺少实例级标签,使得深度神经网络在工业表面检测领域的应用受到了限制.为解决这一问题,本文面向实际的热轧钢板表面缺陷检测任务,提出基于弱监督学习的缺陷检测网络,该网络引入类激活映射模型,使用容易获取的图像级标签进行模型训练,进行钢板表面的缺陷检测.为了进一步提升检测精度和克服类激活映射模型原有的缺点,本文采用性能更优的残差网络作为主干网络进行特征提取,并提出了多层级特征融合网络进行类激活图的生成,来获取更多的细节信息和更准确的目标激活区域.通过在公开缺陷数据集NEU-CLS上进行实验,结果表明本文提出的方法能够在标签不完备的情况下进行缺陷检测任务,并取得0.68%分类错误率和17.75%定位错误率,胜过其他同类的方法.更多还原

【Abstract】 Due to the lack of instance-level annotations,the application of deep neural network in the field of industrial surface detection is limited. In order to solve this problem,a weaklysupervised-learning-based defect detection network is proposed for the practical defect detection task of hot rolled steel plate surfaces. This network introduces the class activation mapping model,which can be used to train the model with the image-level annotations that can be obtained relatively easily,and performs the defect detection on the surface of steel plates. In order to further improve the detection accuracy and overcome the shortcomings of the class activation mapping model,the residual network with better performances is used as the backbone network for feature extraction,and the multi-level feature integration network is proposed to generate the class activation maps. In this way,the network can obtain more detail information and activate target areas more accurately. Extensive experiments have been carried out on the NEU-CLS dataset,and the results show that the proposed method can detect defects with incomplete labels,and obtain the classification error rate of 0. 68% and the localization error rate of 17. 75%,which are better than the other related methods.更多还原