A Research Of Facial Keypoint Localization Under Occlusion Based On Deep Learning

Because of environmental complexity,localizing the facial keypoints under occlusion is one of the most difficult problems in the field of face alignment.Many researchers proposed some strategies to deal with the above issue,but most of these methods cannot handle occlusion explicitly,which leads to poor performances under occlusion.In this thesis,it is believed that the bad performances result from the limited fitting ability of face alignment models.Once occlusion states of the facial keypoints are detected,the model accuracy will be improved by adjusting the fitting ability.To tackle the above problem,a novel face alignment method is proposed in this thesis.Constructing the facial keypoint occlusion detection model,the first module of this method is to obtain the occlusion states of facial keypoints.With the help of the occlusion states,the next module completes the task of adopting the target offset scaling strategy(TOSS)in the cascaded regression model of face alignment.Finally,the inverse offset method is aplied to obtain the final results.The strategy reduces the difficulty of localizing those keypoints under occlusion to reduce the predicting errors of the method.In order to improve the accuracy of the facial keypoint occlusion detection model,this thesis introduces the heatmap mehthod and the deep regression model.To some extent,the heatmap can assist the model to reduce the influence of background noise,improving the recall rate of the facial keypoint occlusion detection model.To extract the more representative features,the classification model is replaced by the regression model to accomplish the task of occlusion detection in human face,which can further improve the performance of the model.Introducing these two mechanisms leads to better performances of face alignment.To show the effectiveness of TOSS in the field of face alignment,the strategy is adopted in two other traditional cascaded regression models to prove its effectiveness.These two models are based on Random Forests and Support Vector Regression(SVR)respectively.It can be seen that TOSS can improve the performance of the face alignment model,showing that the strategy is effective on these models.