The Research Of Face Recognition System Based On HOG Feature

Face recognition has long been a concern in the field of computer vision andpattern recognition. In recent years, the face recognition technology has made rapiddevelopment, but Some shortcomings still need further development and improvement.Gradient orientation histogram has been proven very effective recognition featureextraction operator in the field of target recognition.This paper presents a kind of facedetection method and feature extraction method based HOG feature, we also built a facedetection and recognition system.The main research work in this article are as follows:1. We studied and implemented a new face detection algorithm based on HOGfeatures and support vector machine classifier, finished face classifier training. Thispaper proposes a classifier design method, this method firstly trained a large number ofSVM partial classifier which based on local HOG features, then preferably selectedsome classifier which had good classification results from these partial classifiers.Training sample’ clarified results through new trained SVM classifiers are combinedinto a new feature vector, finally the final SVM classifiers are trained again with thesevectors. During both the training and testing,the HOG integral image is used, therebythe detection and training speed will be greatly improved.2.We studied feature extraction method of fusion of the global HOG feature andlocal HOG feature. To remove redundant information and the singularity of scattermatrix between classes, firstly global HOG feature dimensionality is reduced withprincipal component analysis, and then we get feature vectors with a linear discriminantanalysis method. In addition, we divided face image into parts, and then every localHOG feature dimensionality reduction is finished with Fisher linear discriminantcriteria, and then we can get lacal feature vectors. Finally, the local features and globalfeatures are combined to obtain the final face classifier. The feature we get from thisalgorithm can maintain the uniqueness and stability of face，the feature vector can keeprobust to illumination, pose, expression and other factors.3.In order to overcome the problem of difficulty of human face detection under high light and the dark circumstance, this paper presents an image preprocessingmethod, this method use segmented Gamma transformation to realize gray correction.4. Face detection and recognition tests are carried out on multiple image librariesrespectively, the experimental results show that the performance of the face detectionand recognition algorithms work well.