BMI estimation method fuses face image depth and appearance features
Body mass index(BMI)is an important indicator of human health.3D face information is estimated from 2D frontal images and a end-to-end BMI estimation framework is proposed to further improve the estimation performance of BMI.Firstly,468 3D key points of human face is calculated and facial depth image is drawn according to the depth of key points to head center of mass.Secondly,histogram of oriented gradient(HOG)of human face image is extracted and visualized to represent appearance features.Finally,convolutional neural network(CNN),VGGNet and ResNet are used for feature extraction on depth face image and HOG.Then,Hadamard product is used to fuse the features of the two backbone networks to estimate BMI.In the comparative experiments with existing methods on the two public datasets,the overall mean absolute error(MAE)of the proposed method is reduced by 0.38 and 1,respectively.The above experimental results show that the effectiveness of the proposed BMI estimation method which fuses the 3D face image depth and appearance characteristics.
body mass index(BMI)estimation3D facial landmarkface mesh modelhistogram of oriented gradient(HOG)deep convolutional neural network(DCNN)
NETL
NSTL
万方数据
维普
