Large Pose Face Alignment And Face Estimation Based On Hourglass Network Model

With the rapid development of computer hardware and software technology and computer vision technology,human-computer interaction and face recognition technology has attracted more and more attention of domestic AI scholars.Face pose estimation is an important part of human-computer interaction,and face alignment is a necessary preprocessing process of face recognition.In recent years,with the promotion of deep learning,there has been a breakthrough in the research of these two directions.However,in the unconstrained scene of natural light,face image is affected by various factors,there are still many problems and bottlenecks.Among them,face pose estimation,as an important factor,greatly interferes with the accuracy of face alignment.In order to effectively solve this problem,this paper proposes an hourglass network model based on the python deep learning framework based on the in-depth study and research of the popular large pose face alignment algorithm.The algorithm effectively reduces the influence of pose factors on face alignment,improves the accuracy of face alignment,and on this basis,explores the problem of large pose estimation based on face alignment.The main contributions of this paper are as follows:1.Summarize and explain the current challenges and research status of face alignment and pose estimation.The representative algorithm flows of these two types of problems are introduced in detail,and the performance,advantages and disadvantages of these algorithms are analyzed.2.Based on in-depth understanding and research on the popular face alignment methods,this paper proposes an hourglass network model algorithm based on the Inception-Resnet module.The algorithm uses an end-to-end convolutional neural network model,which avoids the bias effect caused by the previous method that required training targets to be processed in stages.The method in this paper combines the 1 * 1 convolution kernel,3 * 3 convolution kernel,and jump connection to form the Inception-Resnet module,and then the Inception-Resnet module forms afirst-order hourglass model.Four first-order hourglass models are connected in series to finally form this article.Fourth-order hourglass network model.The experimental results show that the normalized mean square error of the proposed algorithm on large pose face images is reduced to 5.92%,which is 12.7% lower than the existing method.The overall normalized mean square error of each pose is reduced to 4.41.%,Which is 8.9% lower than the existing method,and achieves better face alignment in various poses in the natural environment.3.Combining the facial feature points obtained by the face alignment algorithm in this paper,a new facial pose estimation algorithm based on feature points is proposed.The face pose estimation is based on the nonlinear least square method based on robust estimation.The experimental results show that when the face deflection angle is at(0 °,30 °),the average absolute error of the algorithm in this paper is 1.09 °.When the face angle is(30 °,90 °),the algorithm in this paper still maintains a high-precision estimation within 3 °,which improves the accuracy by about 50% compared with the traditional nonlinear least square method.On this basis,using the pose estimation algorithm of this paper combined with parallel depth estimation network,we get better results of face depth estimation.