Research On Face Identification Based On Multi-features Fusion

In recent years,face recognition has become a challenging and far-reaching research topic in the field of computer vision and pattern recognition.The research on the subject of face recognition can be divided into face identification and face verification.We can extract and analyze facial features through a computer,so that the computer can be used to recognize face images of different identities.Nowadays face recognition technology brings many conveniences to human life,but there are still many problems which require scientific researchers to continue to explore and re-search.This thesis expounds the background and significance of face recognition research,the current status of research at home and abroad,introduces face recognition related technologies,proposes a feature learning algorithm and fuses it with other features for face recognition research.The main research contents and works of this thesis are as follows:(1)First of all,this thesis conducts related research on the face image preprocessing operation.The Viola-Jones algorithm is used to perform face detection on the raw face image.Then we locate and correct the face through spatial normalization operations such as affine transformation.Finally,we normalize the processing operation by grayscale,face images with less interference infor-mation are obtained.(2)In order to reduce the dimension of face image and improve the accuracy of face recog-nition.In this thesis,we present a novel top-push constrained feature learning(TFL)method for face recognition.This algorithm learns low-rank approximations from face images after prepro-cessing in the first step so that the metric distance between face images of the same identity is smaller than the metric distance of different identities.The features learned by the algorithm have low dimensions,strong discriminative power and high recognition accuracy.We verify the effec-tiveness and robustness of the algorithm through experiments on four public face datasets.(3)We want to reduce the impact of the face image due to problems such as lighting and occlusion,we consider extracting the local features of the face image and integrating them with the global features learned by the TFL algorithm for face recognition research.Local feature ex-traction algorithms are widely used in the fields of face detection,facial expression recognition and other fields;in extracting local features of face images,rational division of regions and selec-tion of gradient directions are the keys to improving the generality of local features.Therefore,we use gradient of matrices(GOMs)as local features of face images for the first time.GOMs feature can not only describe facial contour information and gradient structure,but also provide reliable spatial information.Subsequently,we propose a generalized low-rank multi-view feature learning(GMVFL)algorithm based on the feature learned by the TFL algorithm and GOMs feature for the research of face recognition.The feature of images were trained by K-Nearest Neighbors(KNN)classifiers,and experiments were performed on four public face datasets.The accuracy of face recognition based on multi-feature fusion was proved better than the recognition accuracy of single feature through ten-fold cross-validation.