Facial Beauty Prediction Based On Multi-scale Image Deep Learning

Facial beauty prediction is a research project of using intelligent information processing technology to realize intelligent prediction of facial beauty degree by extracting the aesthetic features of face.Because of the subjectivity of human face beauty,people lack objective understanding of face beauty and unified evaluation standard,which leads to a big gap comparing with both face recognition and facial expression recognition in the research progress.There are two methods for facial beauty feature extraction,one is manual feature design based on the theory of human face aesthetics,the other is network feature extraction by constructing a deep learning network.The facial beauty prediction based on the theory of human face aesthetics,such as geometry and texture etc,has poor prediction accuracy and generalization ability due to lack of a complete description of face beauty.However,facial beautiful prediction method based on deep learning through supervised training with a large amount of tagged data,can extract potential structural and abstract aesthetic characteristics from face images.Therefore,the full automation and intelligence of facial beauty prediction can be realized,which opens up a new direction for the research of facial beauty prediction.Based on the above analysis,this paper uses the deep learning method to construct the network to automatically extract the aesthetic features and classification of the face,and introduces the multi-scale image technology to enhance the feature extraction ability of the deep learning network.The main contents of this study are as follows:(1)In view of the difficulty of network training and poor fitting effect of small scale facial beauty database,a multi-scale Principal Component Analysis Network(PCANet)facial beauty prediction model is constructed in this paper.Three different scales of facial beauty images are generated by using the multi-scale technique of face beauty images.The aesthetic features of the three sets of faces are extracted by using PCANet as feature extractor.New features with global structure can be obtained by feature fusion.Finally,linear-svm and stochastic forest regression are trained with new features,and the highest correlation coefficient of 0.8627 regression prediction was obtained.(2)With the method of depth learning,a neural network model of deep Convolution Neural Network(DCNN)is constructed,which is suitable for the training of large scale face beauty data sets.The large-scale facial beauty database not only contains more images,but also has a more reasonable distribution of facial beauty.It is helpful to reduce the difficulty of network training.Secondly,by studying the Inception Model of GoogleNet and utilizing the multi-scale feature extraction ability of the unit,the network can enhance the ability of extracting facial image feature details.Then through deep convolution and pool processing,we can extract the feature expression of face image,which is more hierarchical and abstract.Experiments on Large-scale Database of Asian Women's Face Database(LSAFBD)show that the deep convolution neural network in this paper obtains the correct classification rate of 63.5%.Compared with the existing DCNN models,the model of this paper is more suitable for facial beauty prediction.(3)By studying the Mobile-Net model and replacing the traditional convolutional layer with the depth separable convolution layer of Mobile-Net,a facial beautiful prediction model of DCNN is constructed through further optimizing the network structure and simplifying the network parameters,which can run quickly on the embedded device.With the popularization and application of embedded electronic devices,such as mobile phone and tablet,mobile applications are paid more and more attention.By constructing a small scale DCNN prediction model and transplanting the model to embedded devices to run,it has great potential to promote the transformation and marketing of the research on facial beauty prediction.In this paper,the parameter scale of the constructed DCNN model is only 683.7K,the experimental results on Large-scale Database of Asian Women's Face Database(LSAFBD)show that the model can quickly run after transplanting into embedded devices by losing certain accuracy and obtaining better discriminating ability of facial beauty classification,which can improves the application field of face beauty prediction.