Deep transfer network for face recognition using 3D synthesized face

Face recognition has experienced a flurry of advances with deep learning. However, training a model requires a lot of data. In order to meet this condition, some researchers use the 3D rendering technique to synthesize fake face images to expand the training data. Experimental results have demonstrated that this method is an effective way. There exist, however, dataset bias between the real 2D real face images and 3D synthesized face images. In this paper, we use Deep Transfer Network(DTN) to reduce dataset bias. First, we utilize the 3DMM face model to synthesize face images with various poses and natural expression. We choose the Inception-Resnet-V1 as our benchmark model. Then, we optimize our DTN based on maximum mean discrepancy(MMD) of the shared feature extraction layers and the discrimination layers. Our experiments demonstrate that the model jointly trained using synthesized images and real images is more robust than using either dataset (2D real faces or 3D synthesized faces). Furthermore, the performance obtained by our approach is comparable to the-state-of-the-art results to the systems trained on millions of real images.

I. Introduction

Neural networks based deep learning approaches have achieved many inspiring results in machine learning and pattern recognition [1], especially in face recognition [2]–[6]. Google's Facenet [3] has achieved 99.63% on the Labeled Faces in the Wild(LFW) benchmark [7] with the novel Inception architecture and the intriguing Triplet loss, which is almost reaching near-perfect performances. Though impressive as CNN is, training a robust and reliable neural network needs a large amount of data. Therefore, harvesting and labeling large dataset has become an effective way to boost the performance of CNN. For instance, Deep-Face proposed by Facebook [2], VGG-Face [4] and Face++'s Megvii system [5] is trained by using 4.4 million faces, 2.6 million faces, and 5 million faces respectively. And we know that the more complicated the neural network is, the more data it needs in order to prevent overfitting. The state-of-the-art FaceNet [3] utilized 200 million faces with eight million unique identities.