On Deep Machine Learning Methods for Anomaly Detection within Computer Vision

Abstract

This thesis concerns deep learning approaches for anomaly detection in images. Anomaly detection addresses how to find any kind of pattern that differs from the regularities found in normal data and is receiving increasingly more attention in deep learning research. This is due in part to its wide set of potential applications ranging from automated CCTV surveillance to quality control across a range of industries. We introduce three original methods for anomaly detection applicable to two specific deployment scenarios. In the first, we detect anomalous activity in potentially crowded scenes through imagery captured via CCTV or other video recording devices. In the second, we segment defects in textures and demonstrate use cases representative of automated quality inspection on industrial production lines. In the context of detecting anomalous activity in scenes, we take an existing state-of-the-art method and introduce several enhancements including the use of a region proposal network for region extraction and a more information-preserving feature preprocessing strategy. This results in a simpler method that is significantly faster and suitable for real-time application. In addition, the increased efficiency facilitates building higher-dimensional models capable of improved anomaly detection performance, which we demonstrate on the pedestrian-based UCSD Ped2 dataset. In the context of texture defect detection, we introduce a method based on the idea of texture restoration that surpasses all state-of-the-art methods on the texture classes of the challenging MVTecAD dataset. In the same context, we additionally introduce a method that utilises transformer networks for future pixel and feature prediction. This novel method is able to perform competitive anomaly detection on most of the challenging MVTecAD dataset texture classes and illustrates both the promise and limitations of state-of-the-art deep learning transformers for the task of texture anomaly detection.