• Type of project: Research Project
  • Contact: sayako.kodera@tu-ilmenau.de

Motivation

The motivation behind this project stems from the need to reliably identify defects in mechanical parts using thermographic measurements. Anomalies in these parts manifest as disruptive thermal distributions, making anomaly detection crucial for ensuring the safety and reliability of mechanical systems.

Project Overview

Computer vision techniques have been immensely benefitted from recent development of AI technologies. This research project aims to evaluate the applicability of such enriched computer vision anomaly detection methods for thermography data. Specifically, this research project focuses on the comparison and evaluation of existing anomaly detection methods for computer vision tasks. The primary objective is to assess these methods not only in terms of performance, but also some additional key aspects, such as interpretability, noise robustness, sample efficiency, and cost-effectiveness.

Research Steps

  1. Method Selection: Identify the most promising and widely-used anomaly detection methods developed for computer vision tasks (not limited to thermography data).
  2. Analytical Evaluation: Rigorous analysis of these methods, evaluating them analytically and/or conceptually in terms of suitability for thermography data and the considered application, specifically accounting for the following aspects; interpretability, noise robustness, sample efficiency, and cost-effectiveness.
  3. Empirical Testing: Conduct empirical tests using the provided thermography dataset to evaluate the performance of the selected methods.
  4. Result Analysis: Discuss the obtained results, and extensive analysis of the selected methods in terms of the suitability for our dataset accounting for all key aspects.

Expected Outcomes

  • An extensive literature review of existing anomaly detection methods.
  • A rigorous, mathematical discussion of the existing methods
  • Experimental evaluation of the selected methods
  • Extensive discussion on the strengths and weaknesses of each method in the context of thermographic data analysis.

Prerequisites

  • Solid foundations in mathematics and statistics.
  • Proficiency in Python programming.
  • Fundamental knowledge of signal processing is appreciated but not mandatory.

References

  1. “A Unifying Review of Deep and Shallow Anomaly Detection” by L. Ruff et al (2021)
  2. “A Contrario multi-scale anomaly detection method for industrial quality inspection” by M. Tailanian et al (2022)