Face Attributes Extraction with Muti-Task Learning and Knowledge Distillation

The image of a human face contains a lot of key information, and extracting these information can effectively help us understand the facial feature information. In this paper, we propose a multi-task classification network for human faces, which is used to detect the key information of human faces. We use ResNet18 as the basic network and add an attention mechanism to each residual block. Our model includes seven binary classification branches and a regression prediction branch, these branch networks share the same parameters of the backbone network. Finally, we use the method of knowledge distillation to compress the network and reduce the amount of parameters. Since there is no suitable public dataset for our task, we built a dataset containing 76751 images ourselves. we have obtained excellent results in our face dataset. The accuracy of five binary classifications is above 90%, the accuracy of the remaining two binary classifications is above 85%, and the average error of continuous attribute prediction is 3.34.