Advancing Visual Anomaly Detection through Datasets, Algorithms, and Evaluation

Abstract

Anomaly detection refers to the process of identifying patterns that deviate from those
in the training data. Visual anomaly detection refers to finding areas of an image
that contains irregularities not seen during training. This is particularly useful in
scenarios where normal data is abundant but defective examples are rare. Rolls-Royce,
the sponsor of this project, faces such a challenge: aircraft engines must be regularly
inspected for cracks and other defects, yet the high quality of components means that
defective examples are scarce.
Current anomaly detection datasets and metrics do not adequately represent such a
problem. Datasets often lack variability in lighting and perspective, while widely used
metrics are abstract and therefore difficult for industry professionals to interpret. In
light of these issues, a number of contributions are made: (1) a novel software package
for anomaly detection with standardised benchmarking; (2) a new evaluation metric
designed for interpretability; (3) a comprehensive benchmarking of existing algorithms;
(4) a novel dataset which contains varied lighting and perspectives; (5, 6) two new
algorithms achieving state-of-the-art performance; and (7) a novel anomaly detection
paradigm with bounding box predictions.
This work has led to three publications (British Machine Vision Conference, Transactions on Machine Learning Research, and Advances in Visual Computing). Additionally, five GitHub repositories and a pip-installable package have been released. Several
directions for future work are proposed to further advance industrial anomaly detection.

Date of Award	24 Jun 2026
Original language	English
Awarding Institution	University of Bath
Sponsors	Rolls-Royce PLC
Supervisor	Neill Campbell (Supervisor), Samuel Bull (Supervisor), Alan Hunter (Supervisor), Phillip Tregidgo (Supervisor), Jon Morrison (Supervisor) & Joseph Flynn (Supervisor)