Gender Classification Based On Similarity Measuring

Gender classification is an import part of the task of face recognition.Re-cently,face recognition has become a research hotspot.It has been widely ap-plied in the field of verification for payment,human-machine interaction,visual surveillance,and so on,which greatly promotes the development of technology and economy.The goal of face recognition is determing face attributes,and gender is an import attribute.Images shot in limited conditions are chosen as object in the past researches which get relatively ideal results.However,they are not effective enough for real-world images.Thus,this paper mainly study real-world images,which is meaningful.In order to solve the problem,this paper chooses deep learning model to extract features from images and classifies them.By tuning the structure of deep network,features can be more expressive.Deep Residual Network is used for feature extraction in this paper,since it will not degrade when the network is deep.Moreover,this paper uses Large-Margin Softmax Loss for classification task,which improves the accuracy.Based on this,this paper adds similarity measuring method to deep network.The purpose of this method is to map features to a new space,where samples are more discriminative with longer intra-class distance and shorter inter-class distance.Further improvement of accuracy is obtained by this addition.The method proposed by this paper obtains an average accuracy of 91.6%on cross-validation of Adience databases.