Multi-task learning for visual quality control
Transfer learning from neural networks pre-trained on large-scale visual recognition datasets (e.g. ImageNet) has been the standard approach in computer vision for many years. However, this approach provides little benefit in situations where the image statistics deviate substantially from photographic images. One example of such a situation is visual quality control in industrial production lines. In this thesis, we will compile a large-scale dataset for visual quality control and evaluate to what extent pre-training on this dataset benefits image classification and segmentation problems in the context of visual quality control. This thesis will be carried out in collaboration with Layer7 AI GmbH in Tübingen. Possible work locations are Göttingen and Tübingen.
Requirements
- Good mathematical understanding (in particular statistics and linear algebra)
- Python programming
- Experience with deep learning (PyTorch or Tensorflow )
Contact
Alexander Ecker