Thursday, August 31, 2023 - 14:00 to 16:00
The School of Computer Science is pleased to present…
MSc Thesis Defense Announcement
Many-to-One: Transformer-based Unsupervised Anomaly Detection and Localization on Industrial Images
MSc Thesis Defense by: Naga Jyothirmayee Dodda
Date: August 31st, 2023
Time: 2:00PM – 3:00 PM
Location: Essex Hall, Room #122
Abstract:
Anomaly detection is of utmost importance in the realm of industrial defect identification, particularly when
employing computer vision-based inspection mechanisms within quality control systems. This research introduces the
Many-to-One (M2O) framework, which relies on a multi-level transformer encoder combined with a single
transformer decoder, which forms many-to-one relation in the framework for detecting and localizing anomalies. The
rise of Industry 4.0 and electric vehicles has increased interest in this area. Although previous research has made
significant contributions, challenges still exist in this area. It is crucial to develop models that can generalize well and
overcome time complexity problems that affect model performance. The proposed M2O framework aims to address
these challenges and improve the robustness and efficiency of anomaly detection and localization in this domain.
M2O is a reconstruction framework that utilizes transformer-based architecture and employs a novel module called
Multi-Level Feature Fuse to address these challenges. In order to establish a benchmark for industrial electrical
connectors, we have introduced a novel dataset named ECAD, which contains real-world anomalies. This dataset has
the potential to inspire further research in this field. Through evaluation against MVTec AD, BTAD, and ECAD, we have
demonstrated that M2O outperforms existing methods. Extensive comparisons have established M2O's ability to
overcome previous limitations, making it a robust solution for detecting anomalies in industrial environments.
Keywords: Unsupervised learning, anomaly detection, anomaly localization, transformer, optimization, dataset
Thesis Committee:
Internal Reader: Dr. Boubakeur Boufama
External Reader: Dr. Jill Urbanic
Advisor: Dr. Ziad Kobti
Chair: Dr. Aznam Yacoub
