Project description

Computer vision is an attractive technology for entering the scalable digitalisation of industry. The technology is already being used intensively in the end customer sector, but a comparable success in production processes of the so-called "Industrial Computer Vision" (ICV) has so far failed to materialise despite high potentials. Within the scope of the project, the testing of the diverse visual inspections in production with the help of computer vision is to be evaluated. To reduce the effort for generating training data, synthetic training data will be generated from digital twins. For the detection of errors, production backlogs and unknown process deviations, unsupervised learning approaches are to be used, which are trained to the normal state in order to detect deviations. Furthermore, a concept for the use of automated AI-based inspections as official inspection tools is to be developed by researching the explainability of AI models. Furthermore, suitable learning systems for the use of computer vision solutions are to be designed and tested for the further training of employees. When using AI modules for computer vision applications, methods for evaluating and optimising resource consumption should be researched. In this way, operating costs for the use of cloud-based systems can be determined in advance in order to minimise energy requirements. Explainability approaches for the user are to be tested for the comprehensibility of the results and the operation of computer vision applications. This should enable non-specialists to understand their applications and to be able to adapt them in a targeted manner. Within the framework of long-term tests, data drift and stability as well as reliability will be evaluated and examined. In addition, backbone networks for computer vision applications for production environments will be pre-trained exploratively with the collected data sets.