In recent years, Deep Learning techniques have been used extremely successfully in the field of image processing. In addition to Deep Learning, there are many other Machine Learning methods such as the support vector machine. One challenge for the developers of all these methods is: how can these algorithms be used safely and stable for the optical quality assurance in production? This challenge arises from the special features of many learning methods:
Hybrid Methods and »classic« Solutions for Success
Methods such as »Deep Learning« require a large amount of annotated data, e.g. of defects found in a production facility. However, in a well-functioning production environment, many images of faultless products are available, but only a few of products with defects. One possibility is then to perform data augmentation, i.e. artificial defect databases are created on the basis of the real defect data. Another solution would be to describe the defects mathematically and then learn this model.
It is also difficult to change the inspection level of a machine-learning based inspection system during production, e.g. to set certain quality levels. We therefore frequently use hybrids based on »classical« parameterizable methods (filters, morphology, edge detectors) and machine learning.
In addition to solutions for production, we also offer »typical« Machine Learning solutions for image processing. These are often projects in which huge amounts of image data are currently processed manually and this process is now to be automated by software.