Fast and Accurate Face Alignment Algorithm Based on Deep Knowledge Distillation

Most existing face key point detection algorithms only consider improving the generalization performance of the model, while ignoring the model efficiency. However, models with good generalization performance usually contain a large number of parameters, which hinders the deployment of face key point detection algorithms on devices with limited resources. In order to overcome this problem, a Fast and Accurate Face Alignment Algorithm is proposed in this paper. First, a teacher network that stacks multiple hourglass modules is designed. Its model structure is robust to meet the positioning accuracy requirements in actual use. Then, a lightweight student network is designed, which has fewer model parameters and faster feature extraction speed, which meets the speed requirements in practical use. Finally, through the knowledge distillation strategy, the knowledge of the teacher network is transferred to the student network, which makes the student network take into account both the model inference speed and the key point positioning accuracy.