Research On Face Visual Detection And Gender Recognition

With the deepening research of the face recognition technology,face gender recognition based on visual information gradually become a hot topic in the field of computer vision.Compared with other biological features,facial features contain more information,require lower acquisition costs and difficulty.But,because of the diversity and complexity of facial feature,the recognition algorithm with higher accuracy and robustness is required to achieve gender recognition.Aiming at the above problems,in this thesis,face detection,feature extraction and gender classification algorithm were studied on the basis of the analysis of the face gender recognition research status at home and abroad.For achieving face detection,Haar feature and Adaboost algorithm were applied and generate a cascade classifier,Experiments show that the methods have good performance for frontal face detection.In order to achieving geometry normalization of face images,a eye location method based on the two levels of location framework was proposed in the thesis.This method firstly segment the rectangular region of eyes using Adaboost learning algorithm for the rough location of eyes,and then draw the integral projection curve of the eye region,achieve precise eye location by looking for the coordinates of the curve valley point.Experimental results showed that the new method can quickly and accurately locate the center of eyes.A feature extraction method fusing LBP feature and HOG feature was proposed in this thesis.LBP algorithm was applied to characterize the difference between man and woman in terms of facial smoothness and beard.HOG algorithm was applied to characterize the difference between man and woman in terms of eyebrow width and lip thickness.For traditional LBP feature extracted in uniform face block,a method of making blocks partial overlapping was proposed to improve the traditional method,which can avoid dividing one organ into different blocks.Experimental results showed that the proposed fusion method and the improved partitioning method can better improve gender recognition accuracy.Using the SVM classifier for training and learning feature to achieve gender classification.In views of feature dimension and redundancy problem,a improved mRMR feature selection algorithm was proposed in this thesis.The traditional mRMR algorithm make the results tend to one side due to the imbalance of the value range on both sides of the formula,aiming at this problem,the algorithm was improved combining symmetrical uncertainty model of mutual information,which better balance relevance and redundancy and have better feature selection results.Experimental results showed that the improved mRMR algorithm better select max-relevance and min-redundancy features,not only reduce feature dimension and improve the accuracy of gender recognition.Add it all up,in this thesis,face detection and location was achieved by using Adaboost algorithm,and a eye location method based on the two levels of location framework was proposed to locate the center coordinates of eyes and achieved geometry normalization of face region.A method fusing LBP feature and HOG feature was proposed to achieve feature extraction of face region,and improved mRMR algorithm to select feature.At last,Experimental results showed the effectiveness and superiority in respect of improving the accuracy of gender recognition.