基于球面正则化的支持向量描述视觉异常检测  

Spherical regularized support vector description for visual anomaly detection

作　　者：邓诗卓 滕达 李晓红 陈佳祺 陈东岳[1,2] Deng Shizhuo;Teng Da;Li Xiaohong;Chen Jiaqi;Chen Dongyue(College of Information Science and Engineering,Northeastern University,Shenyang 110819,China;Foshan Graduate School of Innovation,Northeastern University,Foshan 528311,China)

机构地区：[1]东北大学信息科学与工程学院,沈阳110819 [2]东北大学佛山研究生创新学院,佛山528311

出　　处：《仪器仪表学报》2024年第3期315-325,共11页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(62202087);广东省基础与应用基础研究基金(2024A1515010244,2021B1515120064)项目资助

摘　　要：异常检测作为视觉领域中一项独特而关键的任务,在医疗、安保等领域具有广泛的前景。异常检测目前受限于大规模异常数据标注,因此现有方法集中在单类分类和弱监督学习,深度支持向量描述(Deep SVDD)是实现单类分类的常见方法。然而,传统Deep SVDD在开展异常检测时往往面临球体崩塌。针对这一问题,提出了基于球面正则化的SVDD异常检测算法,通过引入软间隔损失与支持向量的思想,优化模型学习流程。进一步地,面向可标注样本,提出了基于SVDD的弱监督异常检测方法。在公开数据集MNIST和CIFAR-10上进行消融和对比实验,实验证明,相比于有监督算法,在MNIST数据集上,SR-WSVDD的性能提高了3.7%,而在CIFAR-10数据集上则提高了16.7%。此外,与其他弱监督算法相比,SR-WSVDD在CIFAR-10数据集上提升了1.8%。所提出的SR-SVDD异常检测算法,弥补Deep SVDD容易发生球体崩塌的缺陷,使模型异常检测结果更加准确。Anomaly detection is an important task in the computer vision,such as medical,security.One of the challenges in anomaly detection is not easy to obtain large-scale annotated anomalous data.Existing methods focus on one-class classification and weakly supervised learning.Deep support vector data Description(Deep SVDD)is an important method to realize one-class anomaly detection.However,previous Deep SVDD often encounter the hypersphere collapse when constructing the model of the hypersphere.To solve this problem,support vector data description based on spherical regularization(SR-SVDD)is proposed in this paper.SR-SVDD applies the idea of support vectors to optimize the learning process by introducing slack terms.Furthermore,this paper proposes weakly supervised support vector data description based on spherical regularization(SR-WSVDD),which utilizes small amounts of labeled data.Ablation experiments and comparison experiments are carried out on MNIST and CIFAR-10.Experimental results show that,compared with supervised algorithms,the performance of SR-WSVDD is improved by 3.7%on the MNIST,and 16.7%on the CIFAR-10.In addition,compared with other weakly supervised algorithms,SR-WSVDD improves by 1.8%on CIFAR-10 dataset.The proposed SR-SVDD solves the spherical collapse of previous Deep SVDD,and makes the anomaly detection results more accurate.

关 键 词：计算机视觉 单类分类 弱监督学习 异常检测 自编码器 支持向量 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TH701[自动化与计算机技术—计算机科学与技术]