Natural Synthetic Anomalies for Self-Supervised Anomaly Detection and Localization

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.

Please see our arXiv preprint for more details: https://arxiv.org/abs/2109.15222.

Data

The NIH chest X-ray data can be downloaded here.

The MVTec AD dataset can be downloaded here.

Training

The method for generating self-supervised examples is defined in self_sup_data/self_sup_tasks.py. This is applied dynamically to the images while loading similar to other data-augmentation transforms (see the corresponding Dataset classes). To train anomaly detection models for chest X-ray or MVTec AD data use train_mvtec.py and train_chest_xray.py.

The training scripts contain dictionaries of settings used for the experiments in the paper and supplement. Here are some example commands for training "NSA (logistic)" from the paper:

python3 train_chest_xray.py -s Shift-Intensity-M -d '/path/to/cxr/images/' -o '/where/to/save/models/' -l 'self_sup_data/chest_xray_lists/norm_MaleAdultPA_train_curated_list.txt'
python3 train_mvtec.py -s Shift-Intensity-923874273 -d '/path/to/mvtec_ad/images/' -o '/where/to/save/models/' -n zipper
python3 train_mvtec.py -s Shift-Intensity-M-923874273 -d '/path/to/mvtec_ad/images/' -o '/where/to/save/models/' -n tile

Note that we use mixed gradients for all chest X-ray data and for MVTec AD texture classes. For MVTec AD object classes, we use source gradients due to artefacts of the cloning method (see also appendix A.1 of the paper).

Evaluation

The evaluation procedures are defined in experiments/mvtec_tasks.py and experiments/chest_xray_tasks.py. The evaluation notebooks (mvtec_evaluation.ipynb and chestxray_evaluation.ipynb) can be used to run the evaluation and generate tables for sample-level and pixel-level AUROC and AU-PRO where applicable.