Awesome-Anomaly-Image-Generation

Anomaly Image Generation

In industrial scenarios, anomaly data is extremely scarce due to high acquisition costs. The scarcity of anomaly data limits the performance of anomaly detection algorithms. Therefore, anomaly image generation methods can be used to obtain more anomaly data and improve the performance of anomaly detection algorithms. Anomaly image generation is a challenging task as it not only needs to generate a large number of diverse anomaly images with limited training data but also needs to generate corresponding pixel-level labels.

Papers

Generate Anomaly Images and Masks

Title	Publication	Date	Code
DualAnoDiff: Dual-Interrelated Diffusion Model for Few-Shot Anomaly Image Generation	arXiv 2024	2024.08	Github
AnomalyFactory: Regard Anomaly Generation as Unsupervised Anomaly Localization	ECCVW 2024	2024.08	-
CUT: A Controllable, Universal, and Training-Free Visual Anomaly Generation Framework	arxiv 2024	2024.06	-
LogicAL: Towards Logical Anomaly Synthesis for Unsupervised Anomaly Localization	CVPRW 2024	2024.05	-
AnomalyXFusion: Multi-modal Anomaly Synthesis with Diffusion	arxiv 2024	2024.04	-
AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model	AAAI 2024	2023.12	Github
DREAM: Efficient Dataset Distillation by Representative Matching	ICCV 2023	2023.08	Github
Few-Shot Defect Image Generation via Defect-Aware Feature Manipulation	AAAI 2023	2023.03	Github
Prototypical Residual Networks for Anomaly Detection and Localization	CVPR 2023	2022.12	-
Few-Shot Defect Segmentation Leveraging Abundant Defect-Free Training Samples Through Normal Background Regularization And Crop-And-Paste Operation	ICME 2021	2021.07	-
CutPaste: Self-Supervised Learning for Anomaly Detection and Localization	CVPR 2021	2021.04	-

Generate Anomaly Images Only

Title	Publication	Date	Code
Few-shot Image Generation via Cross-domain Correspondence	CVPR 2021	2021.04	-
Defect-GAN: High-Fidelity Defect Synthesis for Automated Defect Inspection	WACV 2021	2021.03	-
Differentiable Augmentation for Data-Efficient GAN Training	NeurIPS 2020	2020.06	Github
Defect Image Sample Generation With GAN for Improving Defect Recognition	T-ASE 2020	2020.02	-

Dataset

• MVTec AD is a dataset for benchmarking anomaly detection methods with a focus on industrial inspection. It contains over 5000 high-resolution images divided into fifteen different object and texture categories. Each category comprises a set of defect-free training images and a test set of images with various kinds of defects as well as images without defects. URL
• Visa dataset contains 12 subsets corresponding to 12 different objects. It contains 10,821 images with 9,621 normal and 1,200 anomalous samples. Four subsets are different types of printed circuit boards (PCB) with relatively complex structures containing transistors, capacitors, chips, etc. URL

Performance

Since the vast majority of papers have only conducted experiments on the Mvtec dataset, we only list the experimental results on Mvtec.

Generation quality

Mvtec AD

Motheds	IS	LPIPS
DiffAug	1.58	0.09
CropPaste	1.51	0.14
CDC	1.65	0.07
SDGAN	1.71	0.13
DGAN	1.69	0.15
Defect-GAN	1.69	0.15
DFMGAN	1.72	0.20
AnomalyDiffusion	1.80	0.32
AnomalyXDiffusion	1.82	0.33
DualAnoDiff	1.90	0.37

Performance of Anomaly Localization

Mvtec AD

Pixel-level Anomaly Localization

Motheds	AUC	AP	F1
Dream	92.2	54.1	53.1
PRN	96.9	66.2	64.7
DFMGAN	90.0	62.7	62.1
AnomalyDiffusion	99.1	81.4	76.3
AnomalyXDiffusion	99.3	86.1	80.6
DualAnoDiff	99.1	84.5	78.8

Image-level Anomaly Localization

Motheds	AUC	AP	F1
Dream	94.6	97.0	94.4
PRN	91.6	96.6	92.4
DFMGAN	87.2	94.8	94.7
AnomalyDiffusion	99.2	99.7	98.7
AnomalyXDiffusion	99.2	99.8	98.7
DualAnoDiff	98.9	99.7	98.6