Computer Science Department , MS Thesis Presentation , Lichun Gao " Resisting Pollution: Evaluating Robustness of Image Anomaly Detection Methods on Polluted Data""

Lichun Gao

MS Student 

WPI – Computer Science Department

Tuesday, November 26, 2024

Time: 11:30 a.m. – 12:30 p.m. 

Location: Gordon Library 303 Conference Room 

Committee Members:

Advisor: Prof. Elke Rundensteiner 

Reader: Prof. Roee Shraga

 Abstract: 

Anomaly detection in images is a critical task in computer vision, essential for various applications ranging from industrial inspection to medical diagnosis. This thesis provides a comparative analysis of existing image anomaly detection methods under the challenging scenario of noisy training data, where both normal and anomalous instances are present without labels, aligning with unsupervised learning paradigms.

We evaluate a selection of methods, including generative models for reconstruction-based anomaly detection, including GANomaly, DDPM, Autoencoder and representation-based methods such as PatchCore, PaDiM, STFPM. Using benchmark datasets VisA, MVTec, and MNIST, we assess the robustness and effectiveness of these models. Our evaluation not only includes traditional quantitative metrics but also qualitative analyses, highlighting specific regions of detected anomalies. 

This study is motivated by the reliance of current image anomaly detection methods on clean training data, which is often impractical in real-world scenarios where data cannot be guaranteed to be entirely clean. The findings offer significant insights into the strengths and limitations of each approach, providing valuable guidance for future research and practical implementations in noisy and unsupervised settings.