Supervisor:
  • Sanjukta Ghosh (sanjukta.ghosh@siemens.com)
  • Dr. Andreas Hutter (andreas.hutter@siemens.com)

One of the important tasks in industrial applications is visual inspection. Till date, a lot of the visual inspections are carried out manually. It is desired to automate these inspection tasks by setting up cameras. The goal is to leverage on latest breakthroughs in the area of AI and specifically deep learning to solve the problem of detecting anomalies. Classical techniques based on image analysis and deep learning based supervised techniques have been used to address the problem. The challenges of using these approaches are lack of generalization, lack of sufficient training data, lack of availability of sufficient annotations for the training data and class imbalance. One of the possible directions to address these challenges is to use weakly supervised or unsupervised approaches.

The focus of this Master Thesis is on the analysis and implementation of deep learning based weakly supervised or unsupervised algorithm(s) for anomaly detection for industrial applications. It is required to study and analyze state of the art deep learning techniques for anomaly detection followed by implementation and experiments using any of the popular deep learning frameworks.
Further details can be discussed once an interest is expressed.