Research On Detection Algorithm Of Facial Landmarks

In recent years,with the continuous development of image processing technology,techniques such as facial recognition unlocking,face scan payment,and facial beautification have been widely applied in our life.Among them,facial landmark detection is crucial.The detection of facial landmarks provides important information for facial recognition and facial analysis by detecting the position of facial parts or facial contours.The efficiency and accuracy of facial landmark detection influence the time cost and accuracy of the whole system.Therefore,it is extremely important to obtain a fast and accurate facial landmark detection algorithm.The main work and innovations are summarized as follows:(1)A variety of LBP features have been implemented for Deformable Part Model(DPM),which is a facial landmark detection algorithm.In this paper,feature extraction models such as circular LBP feature,uniform parttern LBP feature,MB-LBP feature,mean LBP feature,K-LBP feature and the adaptive threshold gradient LBP which is proposed in this paper are chosen to conduct experiments,the results show that mean LBP outperforms other LBP features.(2)A posteriori algorithm of neural network is proposed to solve the problem of DPM algorithm's false detection of landmarks on human faces.Aiming at the false detection situation when DPM algorithm detects the landmarks of non faces,posterior algorithm,which eliminates false detection points through the neural network,is added to the DPM algorithm.The results on the LFW experimental dataset show that the false detection probability is reduced by about 3%.(3)A landmark detection based on the cascade of DPM and Convolutional Neural Networks(CNN)is proposed.In order to improve the precision of the DPM algorithm,a coarse-to-fine idea is adopted.Firstly,the DPM algorithm is used to detect the landmarks,then the CNN is used to modify each landmark.Comparing with the CNN algorithm,the experiment results on the LFW and LFPW datasets show that the cascade algorithm not only improves the detection accuracy,but also reduces the time by half.(4)An application for synthesizing baby pictures using parental pictures is implemented.In order to accurately transfer the features of the parents in pictures to the baby's pictures,the landmarks should be located,and then the baby's pictures are synthesized by geometric transformation and skin color fusion.To sum up,the facial landmark detection accuracy based on the DPM algorithm is improved by selecting features,adding posterior algorithm and cascaded CNN algorithm.In addition,the application of landmark detection in facial fusion is implemented.