Autonomous face recognition

Abstract

Most state-of-the-art machine learning systems are based on the supervised learning theory, which require separate pre-training procedure for enrolling every new face and updating existing faces. Therefore, an additional human supervisor is normally required. Users cooperation is expected as well during the training phase. However, a human vision system is far more intelligent. It has no difficulty to automatically memorize the faces they have interacted with for future recognition. All the enrollments, updates, and comparisons have been done completely in the brain without any outside assistance. Although the biological reasons behind it are not clear until now, it is not hard to imagine that the brains can combine all information that is useful for recognition, including image processing, video context, logic deduction, experiences etc. Inspired from the human vision system, we combined the conventional learning algorithms and image processing algorithms with predefined rules to increase the intelligence of machine recognition systems. As the first step, face detection is implemented by an industrial image-based face detector combined with novel temporal differencing algorithms. The face detection result, an industrial image-based classifier, temporal filtering and video context related rules are all combined for face recognition. The database can be constructed online and be adapted automatically according to the update rules. State machine is introduced to keep the system running automatically and stably. The major feature of the system is self-learning. No separate or pre-training is required. It can start with an empty database and get to know the faces of the people showing up in an unsupervised way. When known people enter again, the system can recognize them and adaptively update the corresponding databases to keep up with recent views. The proposed system can find promising applications in many fields especially for consumer electronics.