[論文レビュー] CutPaste: Self-Supervised Learning for Anomaly Detection and Localization

Chunliang Li, Kihyuk Sohn+2 more|arXiv (Cornell University)|Apr 8, 2021

Anomaly Detection Techniques and Applications被引用数 86

ひとこと要約

A self-supervised CutPaste-based approach learns representations from normal data to detect and localize unknown image defects, achieving state-of-the-art results on MVTec AD without anomalous training data. It also enables patch-level localization.

ABSTRACT

We aim at constructing a high performance model for defect detection that detects unknown anomalous patterns of an image without anomalous data. To this end, we propose a two-stage framework for building anomaly detectors using normal training data only. We first learn self-supervised deep representations and then build a generative one-class classifier on learned representations. We learn representations by classifying normal data from the CutPaste, a simple data augmentation strategy that cuts an image patch and pastes at a random location of a large image. Our empirical study on MVTec anomaly detection dataset demonstrates the proposed algorithm is general to be able to detect various types of real-world defects. We bring the improvement upon previous arts by 3.1 AUCs when learning representations from scratch. By transfer learning on pretrained representations on ImageNet, we achieve a new state-of-theart 96.6 AUC. Lastly, we extend the framework to learn and extract representations from patches to allow localizing defective areas without annotations during training.

研究の動機と目的

• Develop a defect-detection method that learns from normal data only.
• Introduce a self-supervised proxy task (CutPaste) to learn representations sensitive to local irregularities.
• Show that learned representations enable effective one-class anomaly detection.
• Extend to patch-based representations for localization without anomaly training data.
• Evaluate robustness across diverse defect types and compare to prior methods.

提案手法

• Train a CNN-based encoder with a binary/3-way classifier to distinguish normal versus CutPaste-augmented normal images.
• Propose CutPaste augmentation by cutting a patch, possibly rotating/jittering it, and pasting it at a random location.
• Optionally use CutPaste-Scar as a long-thin patch variant and train a 3-way classifier (Normal, CutPaste, CutPaste-Scar).
• Build a generative one-class detector on learned representations using Gaussian density estimation (GDE) over top features.
• Optionally extend to patch-level representations by applying CutPaste to cropped patches and producing a pixel-wise anomaly heatmap via dense patch scoring and receptive-field upsampling.
• Evaluate image-level anomaly detection via AUC on MVTec AD, and pixel-level localization via GradCAM and patch-based scores.

Figure 1 : An overview of our method for anomaly detection and localization. (a) A deep network (CNN) is trained to distinguish images from normal (blue) and augmented (green) data distributions by CutPaste (orange dotted box), which cuts a small rectangular region (yellow dotted box) from normal da

実験結果

リサーチクエスチョン

• RQ1 Can CutPaste-based self-supervised learning learn representations that generalize to unseen real defects without anomalous training data?
• RQ2 How do CutPaste and its variants compare to other augmentations and self-supervised tasks for defect detection?
• RQ3 Can patch-level representations enable accurate localization of defects without anomaly annotations during training?
• RQ4 Does transfer learning from ImageNet-pretrained features further improve defect detection performance?
• RQ5 How robust is the approach across texture and object categories on MVTec AD?

主な発見

カテゴリ	DOCC	U-Student	P-SVDD	Rotation	Cutout	Scar	CutPaste	CutPaste (3-way)	Ensemble
テクスチャ	90.6	95.3	92.9	29.7 ± 1.4	35.3 ± 2.3	92.7 ± 0.4	67.9 ± 1.8	94.6 ± 0.6	93.1 ± 1.1	93.9
グリッド	52.4	98.7	94.6	60.5 ± 7.0	57.5 ± 3.0	74.4 ± 2.5	99.9 ± 0.1	95.5 ± 0.3	99.9 ± 0.1	100.0
革	78.3	93.4	90.9	55.2 ± 1.4	67.7 ± 1.5	99.9 ± 0.1	99.7 ± 0.1	100.0 ± 0.0	100.0 ± 0.0	100.0 ± 0.0
タイル	96.5	95.8	97.8	70.1 ± 1.9	71.8 ± 4.0	96.7 ± 0.9	95.9 ± 1.0	89.4 ± 2.8	93.4 ± 1.0	94.6
木材	91.6	95.5	96.5	95.8 ± 1.1	92.0 ± 0.8	98.9 ± 0.2	94.9 ± 0.5	98.7 ± 0.3	98.6 ± 0.5	99.1
平均	81.9	95.7	94.5	62.3 ± 2.6	64.9 ± 2.3	92.5 ± 0.8	91.7 ± 0.7	95.7 ± 0.8	97.0 ± 0.5	97.5
物体	bottle	99.6	96.7	98.6	95.0 ± 0.7	88.7 ± 0.8	98.5 ± 0.2	99.2 ± 0.2	98.0 ± 0.5	98.3 ± 0.5
ケーブル	90.9	82.3	90.3	85.3 ± 0.8	80.2 ± 1.4	78.3 ± 1.7	87.1 ± 0.8	78.8 ± 2.9	80.6 ± 0.5	81.2
カプセル	91.0	92.8	76.7	71.8 ± 1.4	69.5 ± 1.1	82.9 ± 0.7	87.9 ± 0.7	95.3 ± 0.8	96.2 ± 0.5	98.2
ヘーゼルナッツ	95.0	91.4	92.0	83.6 ± 0.8	69.7 ± 1.3	98.9 ± 0.2	91.3 ± 0.6	96.7 ± 0.4	97.3 ± 0.3	98.3
金属ナット	85.2	94.0	94.0	72.7 ± 0.5	84.6 ± 0.7	86.9 ± 1.5	96.8 ± 0.5	97.9 ± 0.2	99.3 ± 0.2	99.9
ピル	80.4	86.7	86.1	79.2 ± 1.4	78.7 ± 0.7	82.2 ± 1.4	93.4 ± 0.9	85.8 ± 1.3	92.4 ± 1.3	94.9
ねじ	86.9	87.4	81.3	35.8 ± 2.9	17.6 ± 4.4	11.3 ± 2.2	54.4 ± 1.7	83.7 ± 0.7	86.3 ± 1.0	88.7
歯ブラシ	96.4	98.6	100.0	99.1 ± 0.2	98.1 ± 0.6	94.8 ± 1.0	99.2 ± 0.2	96.7 ± 0.4	98.3 ± 0.9	99.4
トランジスタ	90.8	83.6	91.5	88.9 ± 0.4	82.5 ± 1.2	92.0 ± 0.7	96.4 ± 0.7	91.1 ± 0.6	95.5 ± 0.5	96.1
ジッパー	92.4	95.8	97.9	74.3 ± 1.6	75.7 ± 1.0	86.8 ± 0.9	99.4 ± 0.1	99.5 ± 0.1	99.4 ± 0.2	99.9

• From scratch, CutPaste achieves 95.2 AUC for image-level detection on MVTec AD, outperforming prior work by at least 3.1 AUC.
• With ImageNet-pretrained backbones, CutPaste yields 96.6 AUC, setting a new state-of-the-art.
• Patch-based representations reach 96.0 pixel-level localization AUC, surpassing prior methods.
• Ensembling 5 CutPaste (3-way) models improves image-level AUC to 96.1.
• CutPaste variants (CutPaste and CutPaste-Scar) outperform rotation, Cutout, and scar baselines for defect detection.
• Transfer learning with CutPaste can further improve pretrained EfficientNet features, achieving 96.6 AUC without and with fine-tuning.

Figure 2 : Visualization of (a, green) normal, (b, red) anomaly, and (c–h, blue) augmented normal samples from bottle, toothbrush, screw, grid, and wood classes of MVTec anomaly detection dataset [ 5 ] . Augmented normal samples are generated by baseline augmentations including (c) Cutout and (d) Sc