Research On Key Techniques Of Faciallandmark Localization For Bayonet Surveillance

In recent years,face recognition technology has developed rapidly,and the face feature point localization algorithm has been widely used as the key technology of face recognition and verification,and one of the main areas of application is security monitoring.This paper focused on improving the accuracy of facial landmark localization for faces from bayonet surveillance.The main tasks include:(1)This paper established a bayonet surveillance face database with 68 facial landmarks.As the faces from bayonet surveillance such as train station and airport usually have the characteristics of low resolution,varied face posture,poor lighting,and other complex conditions,resulting in the existing facial landmark localization algorithms have difficulty in accurately locating the facial points for the bayonet surveillance system.The current public face database is mainly collected through the Internet crawler,and faces from bayonet surveillance system are very different from these public face datasets.In order to improve the accuracy of feature point localizations for bayonet surveillance system,this paper established a new face dataset.By dealing with a large number of train stations and bus station monitoring video,and then a semi-artificial method is used to estimate the location of facial landmarks of the monitoring bayonet data.This database includes 6647 face images of the train station bayonet surveillance and 1287 face images of the express bus station bayonet surveillance.All experiments in this paper depended on this bayonet surveillance dataset for training and testing.(2)Aiming at the overfitting problem of face feature points localization in the bayonet monitoring of local Binary Features(LBF),this paper proposes an algorithm for locating facial points based on incremental learning.The main idea of this method is to use the incremental learning method to introduce the face images from bayonet surveillance in the regression model,and to regress the new data to obtain a new shape increment in the final stage of the LBF algorithm cascade regression training,so that the model can be modified to achieve the desired effect.The number of newly added face images selected in this paper is one tenth of the number of training sets.The experimental results show that the proposed method is more effective than supervised Descent Method(SDM),Ensemble of Regression Trees(ERT)and the LBF in feature point localization for faces from bayonet surveillance.(3)Aiming at of the challenging characteristics of faces from bayonet surveillance such as overlooking,low resolution and the motion blur,this paper proposed a weight self-learning multi-task Cascaded Convolutional Networks method.Multi-task Cascaded Convolutional Networks(MTCNN)is a joint cascade convolution neural network framework for joint face detection and face alignment.It is trained in multitasking at the same time in multi-level network structure.Firstly,this paper manually adjust the weight of different tasks to increase the weight of face feature point localization.And then by adding weight self-learning module,making the network can automatically learn and modify the weight,so that it can obtain the optimal weight distribution of different tasks,so as to further improve the accuracy of MTCNN to locate the position of face features.The experimental results show that the accuracy of the method is higher than that of SDM,ERT,LBF,MTCNN and LLF based on LBF incremental learning in bayonet surveillance.Finally,the work of this paper is summarized,and the follow-up work of this paper is prospected.