A Continuous Learning System for Face Clustering and Recognition

This work is to design a high performance and light weight system that can be trained and work well on practical real- world face images. It can run on the mobile phones in the market, and it can automatically identify and group photos in a personal digital album. One target of the research is to produce a system that can be given in the hands of users for long-term. To achieve this goal, we employ the face recognition and object clustering techniques to build a system that can update the number of classes (faces) with a continuously growing number of input photos. This system can be used together with the personal storage (such as hard disk drives); therefore, it has the advantage of privacy. Hence, we propose a system pipeline to create a smart album that will automatically cluster photos and recognize previously seen faces when the user is adding pictures to it. We have done experiments on the public and self-collected datasets and the system is able to perform well even on the challenging images that are difficult to cluster by using the conventional techniques.