Semi-Supervised Deep Learning Approaches for Anomaly Detection in Computer Vision

Abstract

Anomalies are samples which differ significantly from ordinary appearance or behaviours to such a degree that they lay outside what is considered standard in a given task. Deviations may be due to defective or broken regions of a sample, or due to foreign objects present in samples. Detecting such deviations in samples is the task of anomaly detection. In the task of X-Ray Security Scanning or Factory Line Inspection, missing the detection of anomalous instances, especially in the former, can cause catastrophic impact to safety. Missing anomalies within tasks such as Plant Disease Detection or Wind Turbine Blade Fault Detection are likely to cause increased detriment to the assets of these tasks if they are not caught soon enough. The work presented in this thesis aims to push towards automation of the detection of anomalies in such critical tasks. Firstly, an extensive review is conducted into prior approaches and paradigms which have been presented for anomaly detection. As most tasks in visual anomaly detection do not have the luxury of having copious and diverse anomalous samples, if any, methods have since shifted to semi-supervised learning whereby training is conducted solely across non-anomalous samples. An obvious problem with such training is the detection of subtle anomalies (deviations which vary only slightly from normality) in a given task. This was the motivation behind the PANDA architecture, a generative semi-supervised method presented in this thesis. This method, specifically designed to detect subtle and coarse anomalies obtains state-of-the-art results in AUC score across a substantial pool of challenging datasets. Following from this, a trend in anomaly detection has seen denoising approaches obtaining state-of-the-art and robustness to the task, however, such noising approaches are manually defined and random by nature. This thesis presents a method to add optimised, custom noise for any given anomaly detection task. The results of this method show that even a very basic architecture can obtain close to state-of-the-art performance when using this unique noising approach. Finally, an approach to detect faults in wind turbine blades is introduced in the form of a two-stage detection approach which first establishes a more accurate method of blade detection and extraction compared with prior object detection approaches, and then uses off-the-shelf anomaly detection methods to perform successful defect detection of super-pixel sub-regions of the detected blades.