Adaptive incremental learning of image semantics with application to social robot

Intelligent human–robot interaction is an interesting and challenging research topic both in areas of computational intelligence and robot control. In this paper we propose a novel adaptive and incremental image semantics learning framework based on the specific application platform of social robot. This endows the robot with the ability to learn to recognize new images based on previous human–robot interactions. In contrast with most of the intelligent image semantics learning works, which typically focus on how to recognize large scale of training data, this paper deals with how to learn image semantics from zero beginning and enrich the knowledge incrementally with human–robot interactions. In our framework, the user first presents unlabeled images to a humanoid robot for recognition; then the robot answers the user what׳s the image based on the semi-supervised incremental learning framework; thirdly user “teaches” the robot the right label of the image if the robot gives a wrong answer. Users can present more unlabeled images to the robot in the framework for teaching and learning. Extensive experiments and comparisons have validated the proposed methods with encouraging results. Our framework has a broad range of applications including education and rehabilitation.