Research And Application On Attention-based Facial Landmark Detection Technology

Facial landmark detection is an important part of face recognition task,which is widely applied in face verification,expression recognition and other scenes.Considering the wide application of face recognition task,facial landmark detection has been a hot issue in the field of computer vision.In face recognition task,face images are often collected under unconstrained scenarios.The collection process is often affected by uncontrollable factors such as pose,expression,illumination and so on,which bring challenges to high-accuracy facial landmark detection.Besides,the real-time performance of face recognition system requires facial landmark detection with high efficiency.How to reconcile the detection accuracy with speed,thus achieving a good accuracy/speed trade-off,is one of the challenges in facial landmark detection currently.Focusing on the challenges in the field of facial landmark detection,the researches have been carried out in three aspects in this paper: Facial landmark detection model designing,facial landmark detection model training and facial landmark detection model applications.The main contributions are summarized below:1.For model designing,a facial landmark detection model based on attention mechanism is proposed on basis of the convolution neural network.The model has a two-branch feature extraction structure,which obtains rich feature information from face images.In addition,an attention-based feature selector is designed,which enhances the attention of the insensitive key areas,and guides feature learning in the feedforward process.The experimental results indicate that the model achieves a good accuracy/speed trade-off.2.For model training,two novel schemes are proposed: 1)A pose-based data augmentation is proposed to optimize the proportion of samples with different degrees of difficulty in trainset,so as to improve the detection accuracy on the challenging faces;2)A loss function combining heatmap regression and numerical coordinate regression is designed to train the model through combined regression,so as to improve the detection accuracy/speed trade-off.3.For model application,two face recognition systems based on facial landmark detection are carried out: 1)Fatigue monitoring: According to the working process of the driver fatigue monitor system,the facial landmarks are detected in the video stream.Fatigue features are extracted based on the landmark locations,and the fatigue degree judgment result is finally given;2)Face beauty: Make use of the landmark location information to beautify the face organs and shape.A face beauty tool is built by GUI interface.