Face Detection And Recognition Based On MTCNN And SVM Algorithm

In order to further studying the problem of face recognition,this paper researches the problem from two aspects:face detection and face recognition.The experiment is mainly divided into four stages:image preprocessing,face detection,face image feature extraction and classifier training.In the pre-processing stage,the image is filtered and contrast enhanced to improve the recognition accuracy.Using multi-task convolutional neural network(MTCNN)for face detection,widening the network structure and parameter optimization properly,the model performance is further improved.The classification of face and non-face and the regression of face boundary frame in three cascaded networks are analyzed detection rate and recall rate in three tasks.The results show that the accuracy of training classification is 95.57%,the accuracy of bounding box regression is 96.60%,and the accuracy of face key point location is 95.97%.Three algorithms,including principal component analysis(PCA),directional gradient histogram(HOG),linear discrimination analysis(LDA),are used to extract face features in face recognition.Support vector machines(SVM)is used in the classification phase,where 75%of the samples in the database are using for training and 25%for testing.Two kinds of schemes are designed as follows.Scheme one:first,the directional gradient histogram(HOG)of the face image is calculated to extract the most obvious contour information of the face image.Each feature vector of the output is stacked vertically into a two-dimensional matrix.Then principal component analysis(PCA)is used to reduce the correlation and noise between features.Support vector machine(SVM)is used to identify the classifier,and the recognition accuracy of the classifier is calculated as the variance threshold of the PCA algorithm increases.The final results showed that the highest recognition accuracy was 96.0%at 0.8.A ROC curve is used to evaluate the merits and demerits of the method.The area under the curve is 0.998.Scheme two:first,principal component analysis(PCA)is used to transform the face image into a new feature space,eliminating the correlation between image features and noise to extract the global features of the face.Then the linear discriminant analysis(LDA)algorithm is used to further project the transformation to reduce the data dimension in order to taking more projection directions in the experimental stage to keep the original information as much as possible.Finally,support vector machine(SVM)is used for classification recognition.Combining the advantages of PCA LDA and SVM algorithms,the experiment carries out simulation experiments on the ORL Yale AR database and Data detected.The results show that the recognition rate of this method is the highest when the variance threshold is 0.8 Over 99.0 percent.The area below its ROC curve is 0.99224.