Face Attributes Prediction Based on Deep Learning
Ying Fan, Wang Xianliang, Yannan Wu, Zhaoxing Zhang, Shi Yilin · 2018
Predicting face attributes is of great value in business user management and video surveillance as well as other fields. The face attributes mentioned in the topic include the factors such as gender, age, glasses, ethnic and expression. This paper proposes a kind of `end-to-end' machine learning method to predict these five attributes. The convolution neural network MobileNet is adopted, and different loss functions are designed according to the characteristics of each attribute as well. At the same time, during the process of training, the five attributes are trained by sharing parameters. At last, in terms of the tests of 10,000 samples, the attribute prediction has achieved high performance which means that the accuracy rate of gender reached 97.8 % the average age error was 3.2, the accuracy rate of glasses was 99.3 % and the accuracy rate of nationality was 96.3 % as well as the accuracy rate of expression was 68.9 %.