Race Classification from Face using Deep Convolutional Neural Networks
Xulei Wu, Peijiang Yuan, Tianmiao Wang, Doudou Gao, Ying Cai · 2018
As a basic and key attribute of human beings, race plays an indispensable role in face analysis. Traditional machine learning methods all tackle the problem of race classification in combination with two separate steps: extracting artificially designed features and training a proper classifier with these features. Some convolutional neural networks have also been proposed to deal with this problem, but get unsatisfactory accuracies. In this paper, we propose an improved deep convolutional neural network based on an existing network. The network uses a branch structure to merge networks of different depths, such that it can see multi-scale features (features in the low layers are more global and general than those in the high layers). To train this network, we collect a private race database using the available search engines on the Internet, which is larger and more balanced than publicly available databases. Experimental results show that the proposed network can not only extract features and classify them simultaneously compared with traditional methods, but also to achieve state-of-the-art accuracy of almost 99% on both public and self-made databases. Finally, it is necessary to highlight the importance of the advanced face detection and face alignment for the final result.