Anomaly Detection in Images Using One-class and Multi-class Learning Approaches

This thesis focuses on anomaly detection, a classification task that involves identifying whether an image is normal/inlier or anomalous/outlier. Learning schemes typically train on data solely from the inlier class, but some recent methods have proposed semi-supervised extensions that also leverage a small amount of training data from outlier classes. We propose insights into our four novel semi-supervised variational frameworks for classification of images. Results on many real-world industrial and medical image sets show the benefits of our methods over existing methods. We also present the results for the ablation studies to show the benefits of individual novel components in each method.