[论文解读] Set Features for Fine-grained Anomaly Detection

Niv Cohen, Issar Tzachor+1 more|arXiv (Cornell University)|Feb 23, 2023

Anomaly Detection Techniques and Applications被引用 12

一句话总结

SINBAD 将每个样本建模为无序元素集合，并使用带有随机方向的投影直方图以及高斯（或 kNN）密度估计，在不训练的情况下检测细粒度、组合型异常。它在图像级逻辑异常（MVTec-LOCO）和时间序列异常检测上达到最先进的结果，且无需数据增强或训练。

ABSTRACT

Fine-grained anomaly detection has recently been dominated by segmentation based approaches. These approaches first classify each element of the sample (e.g., image patch) as normal or anomalous and then classify the entire sample as anomalous if it contains anomalous elements. However, such approaches do not extend to scenarios where the anomalies are expressed by an unusual combination of normal elements. In this paper, we overcome this limitation by proposing set features that model each sample by the distribution its elements. We compute the anomaly score of each sample using a simple density estimation method. Our simple-to-implement approach outperforms the state-of-the-art in image-level logical anomaly detection (+3.4%) and sequence-level time-series anomaly detection (+2.4%).

研究动机与目标

• Motivate the need to detect anomalies arising from unusual combinations of normal elements, not just anomalous parts.
• Propose a set-based description of samples that ignores element ordering and captures distributional properties.
• Provide a simple, effective density-estimation-based anomaly score using the set descriptor.
• Demonstrate state-of-the-art performance on image-level logical anomalies (MVTec-LOCO) and time-series datasets without training or data augmentation.

提出的方法

• Represent each sample as a set of elements (image patches or time-series windows).
• Extract element features with a pre-trained network or handcrafted features.
• Compute per-set histograms by projecting elements along multiple random directions and binning values into one-dimensional histograms; concatenate across projected directions and feature dimensions.
• Estimate the normal-data distribution with a Gaussian density estimator to score anomalies via negative log-likelihood (Mahalanobis distance or kNN whitening).
• Apply the method at multiple granularity levels (different ResNet blocks for images; window pyramids for time series) and fuse scores.
• Optionally use multiple crops to handle localized anomalies and combine scores across projections and levels.]
• research_questions: ["Can anomalies be detected by modeling the distribution of a sample's elements as an unordered set instead of relying on segment-level or global averages?", "Does random-projection histogram representation of set elements provide discriminative power for fine-grained or logical anomalies across image and time-series data?", "How does a simple density-estimation-based scoring of set features compare to state-of-the-art segmentation-based and reconstruction-based methods on benchmarks like MVTec-LOCO and time-series datasets?"]
• key_findings:[

实验结果

研究问题

• RQ1 Can anomalies be detected by modeling the distribution of a sample's elements as an unordered set instead of relying on segment-level or global averages?
• RQ2 Does random-projection histogram representation of set elements provide discriminative power for fine-grained or logical anomalies across image and time-series data?
• RQ3 How does a simple density-estimation-based scoring of set features compare to state-of-the-art segmentation-based and reconstruction-based methods on benchmarks like MVTec-LOCO and time-series datasets?

主要发现

• Set-based features outperform state-of-the-art on image-level logical anomaly detection on MVTec-LOCO.
• On time-series anomaly detection, the method yields strong results without augmentation or training and compares favorably to deep baselines.
• Random-direction histogram descriptors capture distributional differences between normal and anomalous sets more effectively than simple averages.
• Using kNN whitening or Gaussian modeling with Mahalanobis-style scoring improves robustness over pure Gaussian scoring.
• Multi-granularity (multiple ResNet levels, multiple time-series window scales) and crop-based ensembling improve performance on localized anomalies.