一种基于特征正则约束的异常检测方法

An anomaly detection method based on feature regular constraints

摘要:

异常检测是计算机视觉的一个经典问题.针对异常样本稀少在真实场景中异常很难被捕捉，且标签难以获取，提出一种仅用正常样本进行训练的端到端异常检测模型.首先，通过自动编码器对输入图像进行编码，得到它的低维特征；然后，用一个自回归概率密度估计器对低维特征的概率分布进行正则约束，解码器再将其恢复至原始输入大小；最后，使用一个分类器来判断生成图片的真假.编解码器之间使用了跳线连接，能够最大限度地提高该模型对正常样本的记忆能力.本文在CIFAR10和UCSD Ped2数据集上进行了实验，测试结果显示，IFAR10总共10个类别的平均曲线下面积（AUC）达到73.5%，UCSD Ped2的平均曲线下面积（AUC）达到95.7%.结果证明，该模型能够明显提高异常检测的效果.

Abstract:

Anomaly detection is a classic problem in computer vision. It is difficult to capture the anomalies in the real scene and is difficult to obtain the labels as well, an endtoend anomaly detection model trained only with normal samples is proposed. First, the input image is encoded by an automatic encoder to obtain its lowdimensional features, and then an autoregressive probability density estimator is used to constrain the probability distribution of low dimensional features. The decoder restores it to the original input size. Finally, a classifier determines the authenticity of the generated picture. A jumper connection is used between the codecs to maximize the memory of the model for normal samples. In this paper, the experiments were conducted on the CIFAR10 and UCSD Ped2 datasets. The results showed that the average area under the curve (AUC) of the 10 categories of CIFAR10 reached 73.5%, and the area under the average curve (AUC) of UCSDPed2 reached 95.7%. This model can effectively improve the effect of anomaly detection.