Representative code from my Master Thesis are accessible. No raw or preprocessed images are available, but snippets can be seen in some of the Jupyter Notebooks. Hence, it is not possible for the viewer to reproduce results.
Novelty detection based on training either a convolutional autoencoder (CAE) network or convolutional variational autoencoder (CVAE) network using only "normal" samples. Subsquently, computing statistical probability distributions based upon an anomaly score, an optimal threshold can then be automated by utilizing the G-mean. The threshold is used to distinguish between "normal" and "anomalous" samples.
If further interested in the approach and pipeline, please reach out to me. Cheers!