UseML

Machine vision in industrial quality control, e.g. in surface inspection, generates an enormous amount of data. These data are input for machine learning structures that reproduce human decision making. The setup and optimization of such machine learning structures in industrial environments requires the solution of a few problems:
If new defect classes are taught to the system, it may be difficult to clearly define the boundaries to all other defect classes. This may negatively affect the performance of the whole system.

In cases where there is a mismatch between the decision of the human expert and the automatic system, it is necessary to understand why the system reached that decision. An explanation in terms of feature vectors and decision boundaries is not suitable for communication with the machine operator.
Finally, the training of such decision making systems often takes place in an online mode. Efficient methods are needed to reduce the training effort and also a certain “repair” functionality is required to correct mistakes during training.

The project addresses this topics by extracting information about the state of the machine learning system and by presenting this information in a form that is understandable for the machine operator. Only then there will be a useful interaction between the machine learning system and the operator.
The main result of the project will be software prototypes that are evaluated on a real-world inspection task.

Projcet Name:
Improving the Usability of Machine Learning in Industrial Inspection Systems

Funding:
FFG – IKT der Zukunft – 1. Ausschreibung (2012)

Duration:
01.06.2013 – 31.05.2015