Face attribute prediction with convolutional neural networks

In the background of big data and better computer hardware, convolutional neural network (CNN) has also made great progress. Since AlexNet model won the ILSVRC-2012 competition, CNN has been widely applied to computer version tasks including prediction task of face attributes. Different with single attribute prediction, multi-attribute task is subject to different attributes and is difficult to achieve particularly good performance. With the aim of building a CNN network with good performance and simple framework, ATNet is born as a based network with 4 convolution layers. Based on strong correlation, attributes have a positive effect on optimization of network weights, and conversely will be a negative effect. Therefore, 40 attributes of CelebA dataset are grouped to 3 groups. Because the correlation between groups is also different, groups are trained by branching from different depth. Meanwhile, 1×1 convolution layer is used to reduce dimension and limit the size of networks. Finally, we use a network named ATNet-GT with the above skills and get a good performance on CelebA dataset. The average accuracy of 40 attributes is 90.18%, and the standard deviation of accuracies is 7.25%.