Extracting Face Feature Based On Deep Learning

Today, face recognition, whether in security, monitoring, banking and other fields or in people's daily lives, are playing an increasingly important role. Face recognition is mainly composed of face detection, facial landmarks detection and face features extraction. Currently, both face detection and facial land-marks detection have made good progress. How to obtain more expressive face features becomes one of the most important. Face features extractor based on the low-level image features of manual design is gradually being replaced by the deep learning method. Mainly three reasons can explain. First of all, it is easy to get massive amount of face images on the Internet. Secondly, machines can compute much faster. Last but not least, the continuous research of the deep learning algorithm leads to the excellent performance of the deep learning algorithm in face recognition.In this paper, we deeply analyze the application of deep learning in face features extraction. We study face features extraction mainly in two aspects.One aspect is the design of the convolution neural network architecture, and the other one is selecting the best loss function of distance metric learning.Firstly, we introduce the commonly used convolution neural network, and re-design two convolution neural network architectures, which are more suitable for face features extraction task. Then, we introduce two commonly used loss functions of distance metric learning in face features extraction, and introduce the lifted structured loss function from domain of image search to obtain better results. Finally, the proposed face features extractor achieves an equal error rate of 99.55% in the LFW dataset, and achieves a recognition rate of 70.23% on the MegaFace dataset, both of which are comparable to the state-of-art works.In this paper, we also propose a method which combines distance metric learning and auto-encoder to obtain more expressive features of faces in both the ID card photos and daily photos, aiming at the problem that the face features extractor trained by Internet image is not effective enough in the comparison of ID card photos and daily photos. The face verification task achieves an accu-racy of 90.02% at a false acceptance rate of 10-3 in the ID card photos and daily photos dataset built by our laboratory members, and the face search task achieves an accuracy of 70.10% among a database of about 190 thousands ID card photos.Finally, based on the above method, we construct two types of system which make comparison between ID card photos and daily photos. One of them is mainly focused on face verification tasks, such as passing the customs, open-ing bank accounts and other applications. The other is mainly focused on face search tasks, such as searching for criminals. In an ordinary single-core CPU,the first system can finish a comparison within about 340 milliseconds, and the second system can search in a database of about 190 thousands ID card photos within about 206 milliseconds.