The Research Of Face Attributes Recognition On Convolutional Neural Network

The face attribute recognition,which uses computers to analyze the face of facial features,such as gender,age,ethnicity and other general biological information,and can also distinguish specific categories such as hair color,face type.It can also effectively check whether the object to be identified is wearing a hat,glasses,earrings and other jewelry,and whether the face makeup,expressions and so on.In addition,face attribute recognition has a wide range of application prospects in the fields of human-computer interaction,face retrieval and security systems.Because face images are often affected by uncontrollable factors such as face posture and lighting,so face attribute recognition still faces many challenges.On the other hand,deep learning is a very popular research topic in recent years.Among them,convolutional neural networks have been widely used in many fields and have achieved better performance.Therefore,the research on face attribute recognition using convolution neural network is a challenging and meaningful work.The specific work of this paper is as follows:Firstly,this paper analyzes the research status of face attribute recognition from two different aspects based on traditional learning methods and deep learning methods,and then analyzes the face recognition algorithm of single tasks and the face recognition algorithm of multitasks according to different perspectives of attribute recognition classification tasks.Finally,the traditional feature recognition methods of face and the feature extraction process and attribute classification process of face recognition method based on deep learning are introduced.Second,the traditional face attribute recognition algorithm can only extract low-level face features by hand,and can not cope with the challenge brought by complex face changes caused by posture,lighting and other conditions to face attributes recognition.To solve this problem,this paper introduces a multi-level subnetwork and proposes a single-task face recognition algorithm based on multi-level subnetworks.This method uses "Network in Network" in the design of the network structure.In each sub-network,there are three convolutional layers.After multiple convolutions,the size of the feature map in the sub-network remains unchanged.At the same time,in order to reduce the loading time of the model,the Slice layer and the Eltwise layer are introduced,which greatly reduces the volume of the network model.The algorithm adopts an end-to-end network structure to further improve the training capabilities of the network.The accuracy has been improved 3.1% and4.3% than current state of the art algorithm on the CelebA and LFWA datasets,separately.It is found from the existing research results that the accuracy of partial attribute recognition are low.To solve this problem,this paper proposes an algorithm that combines face parts and multi-label networks.Firstly,the face is divided into six parts,eye,nose,mouth,chin,face and full picture according to the 68 landmarks.Forty attributes and theircorresponding components are combined fully.Secondly,the relationship between features,which are global or local,and attributes should be combined with the corresponding parts.Then a multi-label network is designed based on six face parts to predict the accuracy of each attribute of the face.Finally,the algorithm is verified on the CelebA and LFWA datasets,and the recognition accuracy is significantly improved.