Research On Facial Pose Estimation And Landmarks Localization Based On Deep Learning

With the development of artificial intelligence and computer vision,intelligent human computer interaction and accurate face recognition have more and more important academic value and broad market prospects.Head pose estimation is an important component of human-computer interaction,and facial facial landmark detection is the core of face recognition.Facial pose estimation,also known as head pose estimation,refers to the process of locating the three angle parameters of the head in space according to the facial image,while facial Landmark detection is the process of locating a number of individual defined landmarks according to the facial image.Driven by the wave of deep learning,these two computer vision tasks have made breakthroughs.However,in practical applications,there are many disturbance factors,which lead to the decline of the prediction effect of the model.Therefore,it is still of great significance to design an effective algorithm for facial pose estimation and facial landmark detection.In this thesis,we will focus on the two problems of facial pose estimation and facial landmark detection.By studying the shortcomings of the existing methods,we propose an improved scheme for the existing algorithms.The main contributions of this thesis are as follows:(1)This thesis proposes a multi-modal facial pose estimation algorithm.Firstly,from the perspective of model construction,the objective function of mixture model is deduced in the framework of probability,and the training method of iteration alternation is introduced.On open datasets,a number of comparative experiments are designed.Compared with the current classical algorithms,the mixture regression algorithm proposed in this thesis achieves lower mean absolute errors.Finally,a new noise dataset is generated by adding different intensity noise or occlusion to image samples.The experimental results show that the mixture regression model is robust to both noise and occlusion.(2)An improved deep alignment network is proposed in the thesis.On the basis of the original network,the original convolution neural network is improved by using the well-designed residual blocks of various sizes,which makes the network more capable of feature extraction.The improved model achieves 9.1% improvement on the open data set300 W.The storage overhead and operation bottleneck of the networks are analyzed.Two different lightweight network modules are used to compress the model at the algorithm level.The validity of the model compression is verified by experiments.(3)A facial landmark detection algorithm based on semi-supervised learning is proposed in the thesis.Supervised training usually requires a large number of labeled samples,which leads to the overhead of data acquisition.To solve this problem,this thesis introduces a semi-supervised learning mechanism based on deep alignment network,which trains the model with a large number of unlabeled data and limited label samples.This method unifies supervised learning and unsupervised learning,and improves the effect of supervised learning through unsupervised learning.Experiments show that this semi-supervised learning mechanism significantly improves the accuracy of the original model.