Research On Facial Landmark Detection Based On Deep Convolutional Neural Networks

In recent years,the understanding of the importance of facial landmark detection,how to optimize facial landmark detection technology has become a research hotspot in face recognition.A lot of research scholars have proposed some face facial landmark detection algorithm used to solve various types of face recognition problems,some of them have been successfully applied in the actual project.However,in the complex environment,subject to a variety of external factors,such as posture,lighting conditions,and occlusion factors,the traditional facial landmark detection algorithm will appear a significant decline in accuracy.Based on the comprehensive introduction of the theoretical knowledge and research status of facial landmark detection,this dissertation mainly studies how to use the convolution neural network to design a significant facial landmark detection algorithm.The major researching work of this dissertation can be summarized as follows:(1)Aiming at the shortcomings of the traditional convolution neural network model in dealing with the problem of facial landmark detection and redesigning a convolution neural network model suitable for facial landmark detection.For example,the traditional convolution neural network model is used to deal with classification problems and easy to over-fit,we proposed Deep Convolution Neural Network with Small Filter(DCNNSF).This algorithm introduces the idea of small filter and the deep convolution neural network model with the depth of "network depth",and redesigns the training for facial landmark detection to improve the effectiveness and applicability of the algorithm.In this dissertation,the proposed algorithm DCNNSF is used to predict the 5-point face feature points on the ALFW and AFW face data sets and to compare them with other classical algorithms.The experimental results show that the deep convolution neural network DCNNSF based on small filter has good accuracy and robustness in predicting face facial landmark(5 points)detection.(2)Aiming at the problem that the single convolution neural network model is not suitable for predicting multiple feature landmarks,we propose DCNNSF with Coarse-to-fine Cascade on the basis of the original DCNNSF(DCNNF-CFC).The DCNNSF-CFC algorithm divides the 68 feature points detection problems into the detection problem of the external contour points and the internal feature points of the face.The facial landmarks are further classified into various facial components(left and right eyebrows,left and right eyes,Nose,mouth),this from the whole to the local prediction method can effectively predict the 68 points of facial landmark detection problems.In this dissertation,the algorithm DCNNSF-CFC is experimentally compared with the other eight representative classical algorithms in the 300-W unconstrained face database containing LFPW,Helen,AFW and IBUG.The results show that:our proposed algorithm,DCNNSF-CFC,has higher accuracy and better robustness than most classical algorithms.