Research On Algorithm Of Face Gender Recognition And Age Estimation

With the development of research on Face Recognition Technology,face gender recognition and face age estimation based on visual information has gradually become one of the hot topic in the field of computer vision.Compared with other biometrics,the richness of face feature is high,and the acquisition cost and difficulty are low.However,its diversity and complexity also impose more requirements on the accuracy and robustness of recognition algorithm.Based on the analysis of the current situation of face recognition and age estimation at home and abroad,this thesis studies the methods of face detection,face feature extraction and age feature extraction,gender estimation and age classification.First,the thesis introduces the principle of face detection based on Adaboost algorithm and its training algorithm,and Adaboost algorithm has a good effect on face detection.In order to normalize the face,this thesis proposes a binocular localization method based on a two-level localization framework.The eye region is segmented by using the Adaboost algorithm at first,and then the eye's rough location is achieved.After that,by finding the integral projection curve,Trough point to achieve accurate positioning of the human eye,experiments show that the method can quickly and accurately locate the eye center point.Secondly,the LBP feature which can reflect the texture information of facial image is extracted by different block methods.The HOG feature which can reflect the shape information of the face image is extracted by using different size pixel block retrieval image,and both can be used as the face gender information In this thesis,the concept of depth learning is used to extract the gender features of face images by convolutional neural network,and the above results are used to classify human face.The results show that the HOG feature can obtain the best result.At last,the author constructs the convolutional neural network to extract the age feature of human face.When the number of training set samples is not large,that the training set directly to train CNN does not optimize the network performance.The solution is to initialize the network weights by using a database with sufficient number of samples to train the CNN,and then with this training the network will be re-training and learning,through the network to fine-tune the CNN to meet the needs of this thesis.The thesis use CNN to extract the age feature of human face and test its classification accuracy The LBP and HOG features of human face were extracted by the same face feature extraction method,which was used as the feature of human face age.Face age estimation experiment can test its performance.The experimental results show that the accuracy of CNN extraction is the highest,which shows that the depth learning method can extract better feature than the artificial feature when the sample is sufficient and the classification problem is complex.In conclusion,the author applies the Adaboost algorithm to detect and locate human faces;and the author propose a binocular localization method based on two-level localization framework in order to determine the position of the center of the binocular eye and realize the geometrical normalization of the human face.Using artificial features and depth-And then the SVM classifiers were used to perform the gender recognition and age estimation experiments respectively.The experimental results show that the artificial feature is more suitable for the feature extraction in the thesis.The feature extraction based on convolutional neural network is more prominent in the problem of age estimation with more sample samples.