Enhancing Visual Anomaly Detection with Auxiliary Information

• Author / Creator

  Zhang, Zhaoxiang

• This thesis delves into the advancements in visual anomaly detection (AD), a challenging task in identifying outliers in images such as defects and lesions, which is crucial in many applications including medical diagnosis and industrial manufacturing. This thesis addresses two main challenges: increasing the detection accuracy in unsupervised medical tumor detection and enhancing the performance of zero-shot anomaly detection (ZSAD) models, both with the assistance of auxiliary data.

  In the first part, the thesis focuses on unsupervised AD in medical imaging. It introduces a novel pseudo-anomaly synthesis module designed to generate diverse anomalies in shape and intensity for pseudo-supervised learning. This approach leads to a two-stage training strategy aimed at fostering a well-generalized model that significantly improves tumor segmentation performance.

  In the second part, the thesis presents the Dual-Image Enhanced CLIP ZSAD model. This innovative approach merges visual and semantic data to refine anomaly classification and localization. By leveraging unlabeled visual references and implementing test-time adaptation with pseudo anomalies, the model achieves a notable improvement in detection accuracy, surpassing current leading methods.

  These contributions significantly enhance both unsupervised medical tumor segmentation and ZSAD accuracy through auxiliary data. The introduction of random-shape synthesized anomalies and two-stage training strategy, serves as a foundational framework for refining the pseudo anomaly generation and training methodology. Furthermore, by exploring a vision-language model framework in anomaly detection, this research lays the groundwork for future advancements in the field. These findings underscore the demand for robust, adaptable solutions and set a promising trajectory for ongoing research in AD systems.

• Subjects / Keywords

  • Anomaly Detection
  • Unsupervised Learning
  • Zero-Shot Learning

• Graduation date

  Fall 2024

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/r3-321p-w187

• License

  This thesis is made available by the University of Alberta Library with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.