Study of anomaly detection methods in an industrial setting

Details

Supervisors

De Vleeschouwer, Christophe ; Collin, Anne-Sophie

Faculty

Ecole polytechnique de Louvain

Degree label

Master [60] en sciences informatiques

Abstract

Anomaly detection is a critical aspect of the industrial process which allows to ensure the integrity and reliability of the product at hand. Defective samples are rare and hard to acquire, making it challenging to use traditional supervised methods. In contrast unsupervised learning techniques offer an alternative by leveraging the inherent structure within the data to detect anomalies without the need for labelled samples. This master thesis explores the applied integration of the Variational Quantized Variational Autoencoder (VQVAE) and PixelSnail models by leveraging their respective strengths in order to capture both data structure and spatial dependencies in industrial imagery. The target implementation will be applied over the MVTec Anomaly Detection (MVTec AD) dataset for experimental proposes.