基于近似存在性查询的高效图像异常检测方法OA北大核心CSTPCD
An Efficient Image Anomaly Detection Approach Based on Approximate Membership Query
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.
伍凌川;史慧芳;邱枫;石义官
中国兵器装备集团自动化研究所有限公司,绵阳 621000中国兵器装备集团自动化研究所有限公司,绵阳 621000||北京理工大学机械与车辆学院,北京 100081
计算机与自动化
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
anomaly detectionunsupervised learningapproximate membership queryBloom filterlocality sensitive hashing
《电子科技大学学报》 2024 (003)
424-430 / 7
国防基础科研项目(JCKY2022209A002);国家自然科学基金重点项目(T2293771)
评论