Research On Face Feature Extraction And Location Based On Transfer Learning

With the rapid development of machine vision technology,facial image characteristics extraction and positioning recognition have become one of the most popular research areas.However,factors such as lighting,facial skin color,facial posture,and other accessories(such as glasses)will reduce the accuracy of face recognition in complex natural scenes.In this paper,the method of facial image characteristics extraction and positioning recognition is studied by learning theory and method,in order to improve the accuracy of facial image characteristics extraction in complex scenes.(1)On the basis of expatiating the research of face feature extraction,the development process of face recognition and its research status are analyzed.At the same time,the main content of this design is given.This paper outlines the basic theories and technologies of face feature extraction and localization based on deep learning.(2)Based on the commonly used face datasets such as BIOID,LFW and LFPW,the enhancement method of training sample data is studied.The sample data are enhanced by means of left-right rotation,random clipping,color perturbation,scaling,Gaussian noise,distortion correction,mix-up data enhancement,etc.to solve the problems of limited and unbalanced training sample data in the process of deep learning.(3)According to the relevant algorithm based on Adaboost CNN(Convolutional Neural Network),the flow chart and algorithm steps of the improved algorithm based on CNN+Adaboost face recognition are proposed.Face images are randomly extracted from the face data set.Through comparative experiments,it is found that the recognition rate of the improved algorithm based on CNN+ Adaboost can reach 96.45% for the face images with large light changes,which is about 20 percentage points higher than the recognition rate of the algorithm based on CNN.(4)In this paper,we propose a multi-face localization and mask recognition algorithm based on Resnet50 residual network and transfer learning.The algorithm uses Adaboost iterative formula to classify the HARR features of the input image,searches and locates multiple faces in the target image,and uses Resnet50 residual network model to transfer and learn on the large Imag Net data set,so as to speed up the training of network model and prevent the over-fitting problem caused by too few samples.The model is used to train the face data sets with masks and those without masks,so as to realize the function of multi-face mask recognition and whether the mask is worn properly.In the simulation experiment,3840 annotated images are used for model training,and 580 images are used as the model verification set.After testing,the recognition accuracy of the model on the training set could reach 98.5%,and the recognition accuracy on the verification set was 95.4%.In addition,the migration model and decision tree classification model are proposed on the original model.Through comparative experiments,it is concluded that the accuracy of the Res Net model for feature extraction and random forest for sample classification is 1.23% higher than that of the Res Net model.The method of face feature extraction and positioning based on deep learning not only effectively improves the accuracy of face recognition,but also has portability,which is suitable for intelligent detection related to the industrial field.