Gender Classification Based On Face Images

Face image analysis based on machine learning is one of the active research topics in the field of computer vision.Many related applications about human-computer interaction have seen facial attribute classifications based on face image analysis,such as face recognition,facial expression recognition,and gender classification and so on.This thesis summarizes and analyses the state-of-the-art about facial gender classification,explores and proposes new algorithms based on dictionary learning and deep learning.1.Gender classification based on shape-texture super-complete joint dictionary and ensemble representation.For a given face image,the original shape and texture features are extracted based on Fourier descriptor and spatial block MQ-LBP coding.The dimensionality of texture features is reduced based on feature selection and feature extraction.Then the super-complete hybrid dictionary learning based on shape-texture is realized by combining shape features with dimension-reduced texture features.Based on the estimated ensemble representation and the principle of minimum reconstruction error,the gender class of given face images can be predicted.Gender classification experiments based on CelebA face database demonstrate the effectiveness of the algorithm.2.Gender classification based on graph-constrained low-rank shared dictionary learning.With the structural consistency of face images,graph constraints is introduced and combined with the framework of low-rank shared dictionary learning,and a gender classification algorithm based on graph constrained low-rank shared dictionary learning is proposed.In this algorithm,class-specific information can be modeled based on class discriminant dictionary learning,while the inter-class common components can be modeled based on low-rank shared dictionary learning.Finally,the gender class of a given face image can be realized based on the principle of minimum error reconstruction and the nearest-distance method.Experiments based on CelebA face dataset from unrestricted scenes and AR face dataset in restricted scenes demonstrate the effectiveness of the proposed algorithm.3.Gender classification based on Deep Convolutional Neural Network.First,three typical deep convolutional neural network models,namely VGG16,ResNet50,and Inception-ResNet-v2,are realized and trained for gender classification based on face images.On the basis of three pretrained convolutional neural network models,two different modes of gender classification models via transfer learning and ensemble learning are proposed and realized.Mode 1 borrows the idea of deep transfer learning and realizes multi-level feature extraction based on feature extraction modules in pre-trained deep convolution network model.By combining multi-level features with linear SVM classifier,face gender can be predicted.In mode 2,the pre-trained deep convolutional neural networks such as VGG16,ResNet50,and Inception-ResNet-v2 are regarded as level-0 models,while a binomial logistic regression model is trained as level-1 learner.Finally an ensemble model based on stacked generalization strategy is developed for gender classification.The effectiveness of the proposed method is verified by the CelebA face database on unrestricted scenarios.