【作者】 周杰；

【导师】 陈雁翔；

【作者基本信息】 合肥工业大学 ， 信号与信息处理， 2020， 硕士

【摘要】 异常检测,顾名思义,是对与预期模式不相匹配的数据、事件或观测值的识别。作为计算机视觉领域中热门的研究方向,异常检测在许多领域具有良好的应用前景。例如与网络安全相关的信息传输检测、与制造业相关的材料纹理检测、以及在医学领域的医学影像检测等等。然而,受制于数据标记耗时耗力以及数据不均衡等问题,基于监督学习的异常检测模型在某些方面难以达到预定期望。为了解决上述问题,本文将介绍一种新的结合一分类思想的无监督学习模型,即基于生成对抗网络的异常检测模型。该模型通过对单一类数据进行无监督学习来识别相关异常,无需对数据详细标注,避免了数据不均衡的问题。并利用生成对抗网络强大的拟合能力重构数据,进而通过比较重构数据与测试数据间的距离来区分异常的图像或音频。总之,本文主要研究内容如下:1.通过引入Ano GAN模型,详细介绍了基于生成对抗网络的异常检测方法的原理。并指出该模型存在的问题,进而在此基础上介绍了编码器与生成对抗网络相结合的两个模型Efficient-GAN与GANomaly。除此之外,还通过上述模型的实验结果对比,提出了在不同图像复杂度下异常评分设计应该不一致的观点,并在之后的实验中验证了上述观点。2.在上述模型的基础上提出了一种改进模型,即基于联合分布的流水线式模型。该模型相较于Ano GAN解决了推断阶段需要多次优化的问题,减少了时间消耗,并利用联合分布提高了模型精度。此外该模型通过加入自注意力机制还将图像领域的异常检测方法推广至音频领域,拓宽了模型的应用范围。更多还原

【Abstract】 Anomaly detection,as the name suggests,is the identification of data,events,or ob-servations that do not match the expected pattern.As a popular research field in the field of computer vision,anomaly detection has good application prospects in many fields.For example,information transmission detection in network security,material texture detec-tion related to manufacturing,and medical image detection in the medical field,etc.However,subject to the problems of time-consuming and labor-intensive data la-beling and data imbalance,the anomaly detection model based on supervised learning is difficult to meet predetermined expectations in some cases.In order to solve the above problems,this thesis will introduce a new unsupervised learning model that combines the idea of One-Class,that is,an anomaly detection model based on a generative adver-sarial network.The model recognizes anomalies by performing unsupervised learning on a single class of data without the need to label data in detail,avoiding the problem of data imbalance.It also uses the powerful fitting ability of network to generate data,and then distinguishes abnormal images or audio by comparing the distance between the reconstructed data and the test data.In short,the main research contents of this thesis are as follows:1.By introducing the Ano GAN model,the principle of the anomaly detection method based on the generative adversarial network is introduced in detail.And we pointed out the problems of the model,and on this basis,the two models of Efficient-GAN and GANomaly combining encoder and generative adversarial network are introduced.In addition,through the comparison of the experimental results of the above models,a point of view that the evaluation criteria of the images under different complexity should be inconsistent is proposed,and the above points are verified in subsequent experiments.2.Based on the above model,an improved model is proposed,that is,a pipeline model based on joint distribution.Compared with Ano GAN,the model solves the prob-lem of multiple optimizations in the inference stage,reduces time consumption,and uses joint distribution to improve model accuracy.In addition,the model extends the anomaly detection method in the image field to the audio field by adding a self-attention mecha-nism,which broadens the application range of the model.更多还原