一种基于归一化流的地铁道床异常检测方法

Metro ballast anomaly detection method based on normalizing flow

通过检测和定位道床上的异常情况,可以有效地确保地铁车辆的安全。基于无监督的异常检测方法由于只需要通过正常图像进行训练,不需要太多难以采集的异常图像,因此得到了广泛的应用。综上所述,文中提出一种基于归一化流的无监督地铁道床图像异常检测和定位方法。将多层特征图交叉融合,以提升模型对图像特征的学习能力。建立地铁道床数据集,利用该数据集训练并验证模型实用性。在MVTec AD数据集上的实验结果表明,文中方法性能优于其他同类算法,与DifferNet和CS⁃flow相比,所提方法的AUC提高了0.1093和0.0218。在地铁道床数据集上,所提方法达到了95.95%的检出率和0.9083%的误报率。这些结果表明了该模型对地铁道床异常检测的有效性,以及较好的泛化能力。这为人工智能替代人工巡检地铁道床异常提供了一种新的方法。

It is effective to ensure metro vehicle safe by detecting and locating anomalies of the ballast.Non⁃supervision⁃based anomaly detection methods have been widely applied because they only need to be trained by normal images and do not require too much anomaly images which is difficult to collect.Therefore,an unsupervised metro ballast anomaly detection and localization method based on normalizing flow is proposed.Multi⁃layer feature maps are cross⁃fused to enhance the model ability to learn image features.A metro ballast dataset is established to train and validate the practicality of the model.It performs better than other same type algorithms.The experimental results on MVTec AD dataset demonstrate that the proposed method increases AUC of 0.1093 and 0.0218 in comparison with DifferNet and CS⁃flow.On ballast dataset,the proposed method obtains recall rate of 95.95%and false alarm rate of 0.9083%.These results indicate the effectiveness of the model in detecting anomalies in metro ballast and its good generalization ability.This provides a new method for artificial intelligence to replace manual inspection of metro ballast anomalies.