Face Recognition in Real-world Surveillance Videos with Deep Learning Method

Robust face recognition in real-world surveillance videos is a challenging but important issue due to the needs of practical applications such as security monitoring. While current face recognition systems perform well in relatively constrained scenes, they tend to suffer from variations in pose, illumination or facial expression in real-world surveillance videos. In this paper, we propose a method for face recognition in real-world surveillance videos by deep learning. First, a novel dataset from target real-world surveillance videos is constructed automatically and incrementally with the process of face detection, tracking, labeling and purifying. Then, a convolutional neural network with the labeled dataset is fine-tuned. On the testing dataset collected from the campus surveillance system, the network after fine-tuning achieves recognition accuracy of 92.1 %, which obviously outperforms the network without fine-tuning, which returns a recognition accuracy of 83.6%.