Gender Recognition And Age Estimation Based On Facial Image

Gender recognition and age estimation based on facial image is an important branch of pattern recognition field,and it has great significance to achieve a friendly human-computer interaction.This thesis starts with two aspects of feature extraction and classification to research gender recognition and age estimation,and puts forward the improved method for these two aspects.The main work of the thesis is as follows:(1)Aim to the problem that the single feature can't describe the shape information and texture information of facial image fully,a fusion feature extraction method is proposed based on HOG(Histograms of Oriented Gradients)and ML-GMM(Multi-Level Gaussian Mixture Model).On the one hand,using HOG operator to extract shape features which can describe appearance profile information well.On the other hand,ML-GMM is proposed to extract texture features which can describe variety texture information in different regions of a whole image by constructing multi-level Gaussian mixture model.In the end,these two features are fused by cascade,and the fusion features can fully describe the shape information and texture information of the face image,which is helpful to the next step of classification.Then use SVM(Support Vector Machine)to classify the gender of the fusion features.The experimental results show that the proposed method of fusion HOG and ML-GMM feature achieves better gender recognition performance than the single feature or the common fusion features.(2)To reduce the limitations of the single estimation model in age estimation,two-layer age estimation model is proposed based on SVM-KNN(K-Nearest Neighbor)weighted.Firstly,SVM is used to predict the age range of facial images.Secondly,considering the contribution of the adjacent ages to estimate the real age,the K nearest neighbor theory is used to obtain the age class sequence which makes the minimum between sample and the continuous K age classes in the predicted age range.The distance between the sample and the age class is used as the weight.Finally calculating the final age of the sample by weight.The experimental results show that two-layer estimation model based on SVM-KNN weighted achieves better recognition performance than single estimation model of common used methods and double estimation model in age estimation,it reduces the mean absolute error greatly,and improves the accuracy of the age estimation.(3)In order to deal with the influence of gender recognition and age estimation on each other,a hierarchical model for automatic estimation of gender and age is proposed.In the first layer,SVM is used to predict the rough age range of sample.In the second layer,SVM is used to predict gender information of the sample with the known age range,which can reduce the interference of age information on gender recognition.In the third layer,using the known age range and gender class as prior knowledge to estimate age value of the facial image by weighted KNN algorithm,that can reduc the interference of gender information on age estimation.At the same time,the known age range can reduce the error range of age estimation.The experimental results show that the hierarchical model for automatic estimation of gender and age weaken es the influence of gender information and age information on each other,and improves the gender recognition and the age estimation accuracy at the same time.