Face Age Image Recognition Based On Deep Learning

Face is the most common recognition feature,which reflects the perceptible information such as age,gender,identity,race and emotion.Face recognition has been widely used in information security,finance,intelligent security,public transportation,medical treatment and other fields due to its advantages of naturalness,concealment,non-contact,non mandatory and strong anti-counterfeiting ability,and has gradually become the dominant biometric technology.Although the rapid development of machine learning technology,the innovation of hardware and the increase of dataset scale have brought great impetus to the research,the aging of human face tissue,skin,fat,muscle and bone with the growth of age leads to the change of important features such as texture and shape of human face image,which makes it more difficult to recognize human face age image.In addition,the complexity of the process of age growth and the huge individual differences put forward a new topic for the research of face age image recognition.In recent years,as an important technical means in the field of computer vision,deep learning uses the hierarchical structure of neural network to automatically learn features from the database,so that the computer model can learn and represent data just like the brain’s perception and understanding of multimodal information,has made a breakthrough in image recognition.Therefore,the effective way to improve the level of face age image recognition based on deep learning technology is to select a robust deep model to represent the semantic information of the image,and use a high-precision classifier for feature classification.On the two frameworks of convolutional neural network(CNN)and generative adversarial network(GAN)respectively,this dissertation proposes two effective methods to improve the accuracy of age recognition and cross age face recognition.Model tuning,multi-level feature fusion,hybrid loss supervision and high-performance classifier are used to improve the age recognition accuracy of face image,and the disentangled representation learning and adversarial learning are introduced to improve the recognition accuracy of cross age face image.The following is the work of this dissertation:Firstly,in view of the high complexity of the face age recognition model and the low classification accuracy of the age feature classifier,the face age classification framework is improved in structure,and the FD-MJPC model composed of fine-tuning deep age feature extraction model(FDAFE)and maximum joint probability classifier(MJPC)is proposed.The FDAFE,which is fine-tuned by VGG-Face model,can extract age-related facial features to the maximum extent,and the MJPC can classify age features accurately.Experiments on four commonly used face image sets show that the output age features of different layers of deep model will affect the performance of age classification,and compared with other hybrid models,FD-MJPC can not only improve the accuracy of age classification,but also save about 10% of classification time.Secondly,to solve the problem of the incomplete representation of the features extracted from the model to the age information and the low accuracy of face age recognition,a multi-level fusion network with mixed loss supervision(ML-MLFN)is proposed.The network is composed of the multi-level convolution output of wrn28-10 and the sub network composed of global average pooling layer,full connection layer and mixed loss layer.The multi-level fusion network learns the features in a shared way,and maps the features extracted from the convolution layers of different depths through the sub network,and the training of the model is supervised by the mixed loss function,so that ML-MLFN can completely represent the age information of the image,and accurately estimate the age of the features.Experiments on three open face image sets show that the ML-MLFN proposed in this paper can improve the accuracy of age recognition,and even make the average absolute error of age estimation of the morph II set reach 2.51.Finally,in order to improve the accuracy of cross-age face recognition,an identity preserving disentanglement representation learning(IPDRL)network composed of encoder,generator and discriminator is proposed.The encoder disentangles any age change in the feature,and encodes the identity information of the image at the same time,and extracts features that are only conducive to identity identification,so that it realizes the disentanglement of identity feature and age feature.The generator generates the corresponding identity preserving age image according to the input age features,and the discriminator realizes the class distribution prediction of age and identity through adversarial learning and multi-task learning.The combination of disentanglement representation learning,adversary learning and multi-task learning makes the identity information of face image well preserved,which improves the accuracy of cross-age face image recognition,and achieve 99.53%accuracy in LFW.