Explainable AI for computer vision applications in embedded systems

Artificial intelligence (AI) enables us to mimic the processes behind human intelligence and exploit the potential of intelligent behaviour in man-made systems. Computer vision (CV) is one central application of such AI technology, for which we focus on the development of intelligent models that simulate a human’s processes of perceiving and understanding its visual environment, e.g., in the context of an embedded system. Automated or autonomous embedded systems as deployed in many application domains, including retail, healthcare and manufacturing, can hugely profit from corresponding solutions. Deep neural networks (DNNs) are state-of-the-art models used for various CV problems like image segmentation, object detection, object classification etc. However, although they show incredible performance, one feature where they are currently severely lacking is that of explainability, or in other words the property of being able to offer a transparent explanation for a model’s behaviour and achievable results. Current research on improving explainability thus has to provide end users with more information about the CV models resulting in an increase in the end users’ trust in its performance and acceptance of its deployment. In this cooperation, we will investigate existing concepts for tackling this problem, and we will develop novel approaches for explainable CV applications. To this end, we will unite expertise from the partners in terms of computer vision, machine learning as well as model-based formal methods and reasoning. For evaluating our findings, we will consider smart embedded systems like a robot facing the task of having to assess and understand its visual environment.