Sub-Image Anomaly Detection with Deep Pyramid Correspondences

By Niv Cohen, Yedid Hoshen. arXiv.org, 2020

675 citations recorded by the source corpus.

Nearest neighbor (kNN) methods utilizing deep pre-trained features exhibit very strong anomaly detection performance when applied to entire images. A limitation of kNN methods is the lack of segmentation map describing where the anomaly lies inside the image. In this work we present a novel anomaly segmentation approach based on alignment between an anomalous image and a constant number of the similar normal images. Our method, Semantic Pyramid Anomaly Detection (SPADE) uses correspondences based on a multi-resolution feature pyramid. SPADE is shown to achieve state-of-the-art performance on unsupervised anomaly detection and localization while requiring virtually no training time.

Subfields: Artificial Intelligence, Computer Networks and Communications, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/9277dc70c74bcadf80dab11c28ead83fd085deec

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Networks and Communications

In Computer Vision and Pattern Recognition

Citations in the corpus

Towards Total Recall in Industrial Anomaly Detection (2021, methodology): “This issue has already been noted in SPADE [10] for anomaly segmentation, which makes use of both low- and high-level feature maps.”
AnomalyGPT: Detecting Industrial Anomalies using Large Vision-Language Models (2023, methodology): “Few-shot Industial Anomaly Detection We compare our work with prior few-shot IAD methods, selecting SPADE [5], PaDiM [6], PatchCore [23], and WinCLIP [11] as the baselines.”
Surface Defect Detection Methods for Industrial Products: A Review (2021, methodology): “Typical algorithms mainly include SPADE [54], PaDIM [55], PatchCore [56], etc.”
CutPaste: Self-Supervised Learning for Anomaly Detection and Localization (2021, methodology): “We note that some of these methods use ImageNet pretrained model for transfer learning, either by fine-tuning (DOCC) or distillation (uninformed student).”
Deep Industrial Image Anomaly Detection: A Survey (2023, neutral): “PatchCore [68], SPADE [63], PaDim [73] conduct the ablation study on 16 normal training samples.”
Student-Teacher Feature Pyramid Matching for Anomaly Detection (2021, methodology): “When equipped with the same backbone as SPADE [11], our method further boosts the results, i.e. 0.973 and 0.923 in AUC-ROC and PRO, respectively.”
PaDiM: a Patch Distribution Modeling Framework for Anomaly Detection and Localization (2020, positive): “interpolation and we use a Gaussian filter on the anomaly maps with parameter σ = 4 like in [5].”
SPot-the-Difference Self-Supervised Pre-training for Anomaly Detection and Segmentation (2022, background): “Moreover, due to the rarity of anomalous data, there has been a predominant focus on 1-class anomaly detection, which only requires normal images for model training [6,10,14,26,33,44].”
DeSTSeg: Segmentation Guided Denoising Student-Teacher for Anomaly Detection (2022, methodology): “Besides, the memorybased approaches [7,20,25,35] build a memory bank of normal data in training.”
EfficientAD: Accurate Visual Anomaly Detection at Millisecond-Level Latencies (2023, methodology): “Common methods include multivariate Gaussian distributions [15, 28, 45], Gaussian Mixture Models [37, 68], Nor-malizing Flows [21,44,48,49,64], and the k-Nearest Neighbor (kNN) algorithm [14,38,39,47].”
FastFlow: Unsupervised Anomaly Detection and Localization via 2D Normalizing Flows (2021, methodology): “…networks to obtain the features of normal images and model the distribution with some statistical methods, then detect the abnormal samples that have different distributions (Bergman and Hoshen 2020; Rippel, Mertens, and Merhof 2021; Yi and Yoon 2020; Cohen and Hoshen 2020; Defard et al. 2020).”
CFA: Coupled-Hypersphere-Based Feature Adaptation for Target-Oriented Anomaly Localization (2022, background): “However, as seen in Table 1, the complexity of the modeling process or memory bank space in the previous meth-ods [4, 5, 16] tends to increase in proportion to the size of the target dataset, i.e., |X| .”
Anomaly Detection via Reverse Distillation from One-Class Embedding (2022, methodology): “Recent pre-trained model based approaches achieve promising performance by extracting features from anomaly-free samples as a measurement [7, 8].”