Research On Multi-pose Face Recognition Algorithm

With the development of science and technology in recent years,face recognition technology has entered people's daily life.It has been widely used in many aspects such as face payment,identity verification,smart attendance,etc.And with the advancement of software algorithms and hardware capabilities,face recognition technology has become increasingly mature.However,face recognition technology is often disturbed by factors such as lighting,facial expressions,clothing occlusion,and posture changes in real life.Among them,the multi-pose of the human face is one of the important points and difficulties on face recognition technology.(1)Starting from the method based on human eye positioning,the commonly used ASEF(Average of Synthetic Exact Filters)algorithm is researched,analyzed and improved,and the corresponding human eye calibration data is obtained by positioning the human eye position through the weighted ASEF algorithm.According to the human eye calibration data,the clustering algorithm is used to classify and estimate the different poses of the human face,and then the PCA + Adaboost and ICA + Adaboost algorithms are used for identity recognition.(2)In view of the lack of samples in the current multi-pose face database,the generated adversarial network is used to expand the experimental samples.This paper improves the algorithm based on the DCGAN(Deep Convolution Generative Adversarial Networks)network and conducts experiments to generate virtual samples.On the premise of not affecting the experimental effect,the number of convolution kernels in each layer of the network in this paper is halved compared with the traditional DCGAN network,and its purpose is to speed up its calculation speed.(3)Based on the real-virtual face database obtained by the above generated confrontation network,the pose of the face is detected and judged,and then the identity of the face is recognized.This paper uses the improved VGG-16 convolutional neural network model for recognition experiments,then analyze the practicality of the algorithm.Combining the proportion of the generated virtual images in the training database,this paper has conducted many experiments and comparisons,and gave suggestions on the proportion of virtual images in the real-virtual face database.The above multi-pose face recognition algorithms are all tested and verified on the MultiPIE database.Among them,the multi-pose face recognition algorithm based on clustering can achieve a recognition rate of 95.1%,and the multi-pose face recognition algorithm based on deep learning can achieve 98.5% accuracy.In addition,based on the experimental analysis of the expanded real-virtual face database,it can be concluded that the experiment can achieve better results when the proportion of virtual images does not exceed 30%.