Discrepancy Scaling for Unsupervised Anomaly Detection and Localization

Discrepancy Scaling is a fast, accurate and lightweight deep learning-based method for unsupervised anomaly detection (AD) and anomaly localization (AL) in images.

Discrepancy Scaling uses convolutional neural networks to identify anomalous images and segment them into anomalous and normal regions. It can be applied to both natural images (photographs) and artificial images such as spectrograms. The method is trained using normal images only.

Discrepancy Scaling is implemented in Python 3 using the PyTorch deep learning library.

The method is particularly well suited for quality management in manufacturing but also has applications in construction, maintenance etc.

The code for Discrepancy Scaling is released under the GPL-3.0 license.