卷积自编码器融合核近似技术的异常检测模型

Anomaly detection model of convolutional autoencoder combined with kernel approximation technology

摘要:

图像中的异常检测是计算机视觉中非常重要的研究主题， 它可以定义为单分类问题；针对图像数据集的规模大，维度高等特性，一种新的深度卷积自编码器(Convolutional Autoencoder, CAE)与核近似单分类支持向量机(One Class Support Vector Machine, OCSVM)相结合的异常检测模型CAE-OCSVM被提出；模型中的深度卷积自编码器负责学习图像的本质特征表示，然后使用随机傅里叶特征对卷积自编码器学习本质特征进行核近似，核近似后输入线性单类支持向量机进行图像异常检测。核近似技术克服了核学习技术时间复杂度高的问题；同时深度卷积自编码器与核近似单类支持向量机通过梯度下降法实现了端到端的学习；模型的AUC性能在四个公开的图像基准数据集上进行了实验验证，同时模型与其它常用的异常检测模型在不同的异常率的情况下进行了性能对比；实验结果证实CAE-OCSVM模型在四个公开图像数据集上的性能都优于其它异常检测模型，表明了CAE-OCSVM模型更适合大规模高维数据集的异常检测

Abstract:

Anomaly detection in images is a very important research topic in computer vision. It can be defined as a one classification problem; for the large-scale and high-dimensional characteristics of image data sets, a novel anomaly detection model is proposed CAE-OCSVM, which is a combination of deep convolutional autoencoder (CAE) and a kernel approximate one-class support vector machine (OCSVM). The deep CAE in the model is responsible for learning the essential feature representation of the image, and then uses random Fourier features to perform kernel approximation to the essential features learned by the CAE. After the kernel approximation, the linear OCSVM performs anomaly detection on the image, and the kernel approximation technology overcomes the problem of high time complexity of kernel learning technology. Meanwhile the CAE and the kernel approximation OCSVM achieve end-to-end learning through the gradient descent method. The AUC performance of the model was tested and verified on four public image benchmark data sets. At the same time, AUC performance was compared with other commonly used anomaly detection models under different anomaly rates. Experimental result confirm that the performance of the CAE-OCSVM model on the four public image data sets is better than other anomaly detection models, indicating that the CAE-OCSVM model is more suitable for large-scale high-dimensional data set anomaly detection.