基于近似存在性查询的高效图像异常检测方法  

An Efficient Image Anomaly Detection Approach Based on Approximate Membership Query

作　　者：伍凌川 史慧芳 邱枫 石义官 WU Lingchuan;SHI Huifang;QIU Feng;SHI Yiguan(Automation Research Institute Co.,Ltd.of China South Industries Group Corporation,Mianyang 621000,China;School of Mechanical Engineering,Beijing Institute of Technology,Beijing 100081,China)

机构地区：[1]中国兵器装备集团自动化研究所有限公司,绵阳621000 [2]北京理工大学机械与车辆学院,北京100081

出　　处：《电子科技大学学报》2024年第3期424-430,共7页Journal of University of Electronic Science and Technology of China

基　　金：国防基础科研项目(JCKY2022209A002);国家自然科学基金重点项目(T2293771)。

摘　　要：对于图像异常检测问题,查询测试样本在正常样本集中的K近邻距离并估计其异常程度,是一类准确率较高、对复杂分布的效果较稳定的方法。此类方法采用近似最近邻搜索(Approximate Nearest Neighbour Search, ANNS)索引进行K近邻搜索。但由于ANNS查询操作较高的计算开销和现实问题中庞大的数据量,此类方法的计算效率难以应对低时延、高吞吐量的应用场景。该文基于局部敏感哈希和布隆过滤器,提出了一种近似存在性查询(Approximate Membership Query,AMQ)方法,用特征近似存在性预测异常样本。相比于ANNS,AMQ具有更低的计算复杂度且更适合单指令多数据并行,可以有效解决基于特征库检索方法的计算性能瓶颈。在MVTec-AD数据集上的实验结果显示,基于AMQ的方法的异常分割准确率仅比ANNS方法降低1%左右,但推理时延、吞吐量和内存开销显著较优,接近端到端深度学习异常检测模型的计算效率。An accurate and stable approach to image anomaly detection is to query the K-nearest neighbours of the image features from normal examples and estimate the anomaly score,relying on Approximate Nearest Neighbour Search(ANNS)indices.ANNS query operation has high computational cost on large datasets,unpractical for low-latency and high-throughput scenarios.Based on locality sensitive Hashing and Bloom filters,an Approximated Membership Query(AMQ)based approach is proposed to predict anomalies by approximate membership of features.AMQ can address the performance bottleneck of search-based methods,given its lower complexity and better compatibility with single-instruction multiple-data parallelism than ANNS.Experimental results on MVTec-AD show that the accuracy of AMQ-based method is just decreased about 1%in comparison with ANNS-based methods,while the inference latency,the throughput and the memory footprint are significantly improved,close to the efficiency of end-to-end deep learning anomaly detection models.

关 键 词：异常检测 无监督学习 近似存在性查询 布隆过滤器 局部敏感哈希 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]