Robust Face Recognition via Transfer Learning for Robot Partner

Face recognition is crucial for human-robot interaction. Robot partners are required to work in real-time under unconstrained condition, yet, do not restrict the personal freedom of human occupants. On the other hand, due to its limited computational capability, a tradeoff between accuracy and computational load needs to be made. This tradeoff can be alleviated via the introduction of informationally structured space. For this paper, transfer learning is employed to perform unconstrained face recognition, where templates are constructed from domains acquired from various image-capturing devices, which is a subset of sensors from the informationally structured space. Given the environmental conditions, appropriate templates are used for recognition. Currently, different database images are used to simulate different environmental conditions. The templates can be easily learned and merged via a reformulated joint probabilistic face verification method, which reduces significantly the processing load. Tested on standard databases, experimental studies show that specific and small target domain samples can boost the recognition performance without imposing strain on computation.