Research On Age Estimation Of Face Images Based On Convolution Neural Networks

Age estimation from face image is an important task in the field of human-computer interaction and computer vision,it has a wide range of practical application value.Concerning the problem that the accuracy of age estimation of face images is low,this thesis proposes a new method of face age estimation based on deep residual networks of residual networks(RoR)and large data sets pre-training to improve the accuracy of face image age estimation,which exhibits a robust and high performance.Firstly,this thesis introduces five convolution neural network models,and describes the principle and network structure.In order to further improve the learning ability of the model and suppress the gradient disappearance problem,this thesis constructs the RoR on the basis of the residual network,and the deep RoR is used as the basis convolutional neural network model to deal with the problem of face age classification.Secondly,this thesis presents the framework of age estimation.The RoR is pre-trained on the ImageNet dataset to learn the expression of basic image features.Then the IMDB-WIKI-101 dataset is established to fine-tune deep RoR,which achieves migration learning from general object images to face images,thus making the model adapt to the distribution of the age groups and improve the network learning capability.Thirdly,considering that face age estimation accuracy is affected by gender,the facial image features of the peer who have different gender are different.Compared with the general object images,there is a sequential relationship among the face images.Therefore,two special training methods: pre-training by gender and specific weight loss layer,are used in age group classification.In addition,stochastic depth algorithm is added during network training to reduce the over-fitting phenomenon.Finally,the evaluation methods and three datasets are introduced.Age estimation contains two categories: age classification and age regression.Adience is used for age group classification,and the new state-of-the-art accuracy and 1-off value are obtained,respectively 67.74% and 97.52%.Meanwhile,MORPH Album 2 and FG-NET are used for age regression,and the deep expectation is used to improve the accuracy,and the best mean absolute errors are obtained on the two datasets of 2.34 and 2.17 years respectively.The RoR-152 achieves new state-of-the-art results on all the three datasets.The experimental results show that the combination of deeper RoR and large data sets can effectively improve the accuracy of face age classification.