Machine Learning Methods
SimpleNet reaches 99.6% anomaly-detection AUROC on the MVTec AD benchmark
In a 2023 paper published in the IEEE Computer Society Conference on Computer Vision and Pattern Recognition proceedings, researchers from the University of Science and Technology of China describe SimpleNet, a network for detecting and localizing anomalies in images. The paper reports an anomaly-detection AUROC of 99.6% on the MVTec AD benchmark. AUROC measures how well the system separates normal images from anomalous ones, with higher scores meaning better separation. The authors report that this result reduces the error by 55.5% compared with the next best performing model. SimpleNet also ran at 77 frames per second on a 3080ti GPU.
What is SimpleNet?
SimpleNet uses four parts. A pretrained feature extractor pulls local features, which are compact numerical descriptions of image regions. A shallow feature adapter shifts those features toward the target domain. A simple anomaly feature generator creates synthetic anomaly features by adding Gaussian noise to normal features. The binary anomaly discriminator decides whether each feature looks normal or anomalous. SimpleNet discards the generator during inference.
Why does SimpleNet add Gaussian noise to normal features?
The authors argue that creating synthetic anomalies in feature space is more effective because defects may not have much commonality in the image space.
Why use a simple discriminator?
The abstract calls a simple discriminator efficient and practical. The discriminator separates anomaly features from normal features.
How well did SimpleNet perform?
The paper reports a 99.6% AUROC for anomaly detection on the MVTec AD benchmark. The authors say SimpleNet outperformed previous methods quantitatively and qualitatively. They also report significant improvements on the One-Class Novelty Detection task. The paper reports a frame rate of 77 FPS on a 3080ti GPU. The authors describe SimpleNet as faster than existing methods.
Limits and what's next
These results come from one 2023 paper reporting benchmark tests. The abstract reports results on the MVTec AD benchmark and on a One-Class Novelty Detection task. The abstract does not say whether these results were tested in a real-world deployment or independently replicated.
Source: Zhikang Liu et al., “SimpleNet: A Simple Network for Image Anomaly Detection and Localization,” IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings, 2023.