Research On Face Alignment Method Based On Convolutional Neural Network

Face alignment plays a pivotal role in the field of computer vision.It is one of the key steps in face recognition.It is promoted in many fields such as security,finance,and entertainment.Therefore,fast and accurate face alignment has a huge effect on face recognition technology.At present,the use of complex face transformation,image background and acquisition equipment and other factors still face major challenges in the application of face alignment.In this paper,the hourglass model is used for real-time research,but the feature extraction is still limited.Therefore,the improved hourglass model — —high-resolution model is further used to conduct research and discussion on accuracy.Firstly,given the complex structure and time problems in the application of the hourglass structure to face alignment,a lightweight hourglass network with depth and separability is proposed.The network first constructs a lightweight hourglass network through knowledge distillation to solve the complex network problem.Then,deep separable convolution is used in the hourglass model to simplify the complex network and solve the problem of high time overhead.Experimental results show that compared with recent face alignment models,the alignment accuracy of this model is unaffected basically and the alignment speed is improved significantly on the classic 300 W and WFLW data sets.Secondly,aiming at the improved hourglass model——high-resolution network used for face alignment and lack of correlation between local and global key points when extracting features,a high-resolution model with global information retention is proposed.This model adjusts the network dimensions,and uses new residuals to extract feature information of different scales to balance the loss of information.In addition,it combines feature fusion to increase detailed description and retain more effective information such as spatial background.Finally,the loss of each layer of the resolution network is output through the intermediate supervision mechanism to prevent the gradient from disappearing and reduce the noise.Based on the application of the hourglass model to the data set,the COFW data set with more changes is added.Compared with the face alignment model in recent years,based on maintaining real-time,combined with global information to highlight feature information,this model has a certain alignment accuracy.