Image-based Facial Age Estimation

Age estimation based on facial images has a vast market space in security surveillance systems,customer management systems,facial recognition technologies,image and video retrieval systems,criminal case investigations and other fields,so the study of facial age is of great significance.The age estimation of facial images aims to dig out the relationship between facial images and age by means of computer vision,machine learning,and deep learning,which is used to calculate the specific age or age range of the facial images.Based on the existing age estimation technology,in this paper,the age estimation research of facial images was carried out,and two estimation methods were proposed.The main research work is as follows:(1)The fusion model of ResNet50?SVR was constructed and applied to facial age estimation.In the aspect of facial age feature extraction,the ResNet50 network was used as a method to extract facial age features in this paper.The ResNet50 network has a strong feature learning ability,but its learning feature dimension is relatively higher,which makes the linear model is difficult to perform the corresponding fitting task.A new SVR(Support Vector Regression)model was proposed to fit the age estimation task in this paper.Firstly,the ResNet50 network was pre-trained on the Image Net dataset.Secondly,the pre-trained ResNet50 network was fine-tuned on the MORPH 2 dataset.Finally,the output result of the Pooling layers in the ResNet50 network as an age feature was sent to the SVR model to estimate the age of the facial image.The ResNet50?SVR model constructed in this paper has a MAE(Mean Absolute Error)of 3.5 on the FG-NET dataset.Compared with the standard CNN model,the ResNet50?SVR model is improved the accuracy by 33.96%,which shows that the model has strong generalization ability for face age estimation.(2)The fusion model of ResNet50?TPR?C2AE was constructed and applied to facial age estimation.In order to comprehensively consider the similarities and correlations between age features,ResNet50 network and TPR(Two-Points Representation)method were used to extract age features.Based on the C3AE(Compact Cascade Context-based Age Estimation)model,the C2AE(Compact Cascade Age Estimation)model was proposed and used for age estimation,aiming to solve the difficulty in estimation of face images between adjacent ages.The C2 AE model adopted the compact structure and cascaded training strategy of the C3 AE model.In the ResNet50?TPR?C2AE model,a compact structure of three cascaded tasks was used to obtain the age estimation model of the face image,where three loss functions were defined for the three cascaded tasks.The first cascade task,Categorical Crossentropy was used to measure the error between the initial age features.KL(Kullback-Leibler)divergence was replaced for Focal loss to solve the difficulty in estimation facial images between adjacent ages in the second cascaded task,so as to obtain age distribution characteristics closer to the age labels.In the third cascade task,MAE was used to measure the prediction error of the C2 AE model for face age estimation,in order to improve the performance of the age estimation model.First of all,the ResNet50 network was pre-trained on the Image Net dataset,and was fine-tuned the pre-trained ResNet50 network on the MORPH 2 dataset.Then,the output of the Pooling layers in the ResNet50 network was used as the initial age features,and the TPR method was used to further extract age distribution features from the initial age features.Finally,the age distribution feature was sent to the C2 AE model to estimate the age of the facial image.The ResNet50?TPR?C2AE model constructs in this paper has a MAE of 1.9 on the MORPH 1 dataset.Compared with the standard CNN model,the ResNet50?TPR?C2AE model is improved the accuracy by 64.15%,indicating that this model has high estimation accuracy for face age estimation.The age estimation methods in this paper were verified on international public age databases,and the results show that the methods proposed in this paper have a high estimation accuracy for facial age.