Natural Synthetic Anomalies for Self-Supervised Anomaly Detection and Localization (2109.15222v3)

Abstract: We introduce a simple and intuitive self-supervision task, Natural Synthetic Anomalies (NSA), for training an end-to-end model for anomaly detection and localization using only normal training data. NSA integrates Poisson image editing to seamlessly blend scaled patches of various sizes from separate images. This creates a wide range of synthetic anomalies which are more similar to natural sub-image irregularities than previous data-augmentation strategies for self-supervised anomaly detection. We evaluate the proposed method using natural and medical images. Our experiments with the MVTec AD dataset show that a model trained to localize NSA anomalies generalizes well to detecting real-world a priori unknown types of manufacturing defects. Our method achieves an overall detection AUROC of 97.2 outperforming all previous methods that learn without the use of additional datasets. Code available at https://github.com/hmsch/natural-synthetic-anomalies.

• The paper presents a novel self-supervised framework that uses Poisson image editing to produce natural synthetic anomalies for precise detection and localization.
• The NSA method trains models solely on normal data, achieving 97.2% AUROC for detection and 96.3% for localization on the MVTec dataset.
• The approach offers promising applications in industrial inspection and medical imaging, enhancing unsupervised anomaly assessment with realistic synthetic anomalies.

Natural Synthetic Anomalies for Self-Supervised Anomaly Detection and Localization

The paper presents a novel methodology for anomaly detection and localization, coined Natural Synthetic Anomalies (NSA). This approach leverages self-supervised learning to train models using only normal data, addressing the challenge of detecting a priori unknown anomalies. The NSA method utilizes Poisson image editing to generate synthetic anomalies with natural sub-image irregularities. This advances beyond the limitations of previous augmentation techniques that often resulted in less natural-looking anomalies due to obvious discontinuities.

Methodology Overview

The authors propose a self-supervised task aimed at enhancing an end-to-end model's capability to detect and localize anomalies at a sub-image level. In the NSA framework, scaled patches from separate images are seamlessly blended using Poisson image editing to create diverse synthetic anomalies. This process incorporates:

• Patch Sampling: A variety of rectangular patches are sampled and resized, maintaining intentional variability in size and aspect ratio.
• Poisson Blending: A critical ingredient in NSA to ensure the anomalies appear more natural, avoiding the unnatural artifacts introduced by simple cut-and-paste operations.
• Label Generation: A unique labeling strategy is implemented, where pixel-wise labels are derived from the local intensity differences between anomalous and normal regions, allowing the model to learn more refined anomaly detection.

NSA stands out for its ability to create numerous and diverse synthetic anomalies that closely resemble potential real-world irregularities without pre-trained models or additional data requirements.

Experimental Evaluation

The NSA methodology was evaluated on the MVTec Anomaly Detection dataset, which contains a suite of classes representing various objects and textures with both normal and defective instances. NSA achieved an impressive 97.2% AUROC for detection and 96.3% AUROC for localization, outperforming other self-supervised and unsupervised methods that do not utilize additional datasets like ImageNet for pre-training.

Moreover, NSA was tested on a curated subset of a public chest X-ray dataset (rCXR). Here too, NSA demonstrated superior performance over baseline self-supervised techniques such as Foreign Patch Interpolation (FPI) and its variant using Poisson blending (PII), confirming NSA's generalizability and potential applicability in medical imaging contexts where annotations are scarce or unavailable.

Discussion and Future Implications

The paper contributes foundational insights into the construction of synthetic anomalies and their implications for model training in anomaly detection tasks. The ability to create realistic synthetic anomalies suggests significant potential for NSA in various domains, including industrial inspection and healthcare, particularly for unsupervised detection scenarios.

Potential future directions include:

• Domain Adaptation: While NSA does not rely on pre-trained models, exploring adaptations for specific domains, including different modalities of medical imaging, could enhance anomaly localization.
• Integration with Other Techniques: Combining NSA with density estimation approaches or integrating with other unsupervised learning strategies might improve understanding in less-structured anomaly detection tasks.
• Anomaly Degree Quantification: Further refinement in labeling strategies might include probabilistic measures for anomaly severity, expanding the scope to risk assessment models.

In conclusion, the NSA framework enriches the toolkit for self-supervised anomaly detection, offering a robust strategy to infer anomalies with minimal supervision, a critical requirement in contemporary data analysis tasks across diverse application domains.