Anomaly Detection in Images - Part 1

Anomaly detection problems are ubiquitous in engineering: the prompt detection of anomalies is often a primary concern, since these might provide precious information for understanding the dynamics of a monitored process and for activating suitable countermeasures. In fact, anomalies are typically the most informative regions in an image (e.g., defects in images used for quality control). Not surprisingly, detection problems have been widely investigated in the image processing and pattern recognition communities, and are key in application scenarios ranging from quality inspection to health monitoring.

The tutorial presents a rigorous formulation of the anomaly-detection problem that fits with many imaging scenarios and applications. The tutorial describes, by means of illustrative examples, the most important anomaly-detection approaches in the literature, and their connection with the machine-learning perspective of semi-supervised and unsupervised learning/monitoring. Special emphasis will be given to anomaly-detection methods based on learned models, which are often adopted to handle images and signals. tIn particular, these will be divided into traditional models (including autoencoders, learned projections and dictionaries yielding sparse representations) and deep learning models (including CNNs, deep-one-class classifiers and deep generative models).

The tutorial is accompanied by various examples from our research projects where we applied anomaly-detection algorithms to solve real world problems: visual quality inspection for monitoring chip and nanofiber production.