Facial Beauty Prediction Research Based On Multi-Feature Fusion

Facial beauty plays an important role in many social activities.It affects careers such as digital entertainment,models and performances.People's pursuit and yearning for facial beauty has attracted scholars from various fields to study facial beauty prediction.However,the traditional method of facial beauty prediction has spent a lot of time in features extraction,and the effect of prediction is not ideal.At present,researchers use the deep features extracted by convolution neural network instead of hand-ceafted features,which enables CNN to extract facial beauty features intelligently,thus making it possible for machine to automatically predict facial beauty.In this paper,a facial beauty prediction method based on multi-feature fusion is proposed,which use multi-feature fusion instead of single feature to enhance the ability of facial beauty prediction network.The main contributions of this work are as follows:(1)A facial beauty prediction model based on the fusion of geometric features and PCA network is used.The features extracted from PCANet are fused with geometric features to obtain more robust and discriminatory features.Finally,5-fold cross-validation experiments are carried out by using SVM regression and random forest regression.Experiments show that the prediction performance of the PCANet after fusing geometric features is better than that without features fusion,which shows that the fused geometric features can effectively improve the results of facial beauty prediction.(2)A new method of facial beauty prediction based on local binary pattern and convolution neural network is used.LBP operator extracts face images with local texture features and light invariance.By adding LBP texture images to CNN,learning the local structure features of LBP images,the network extracts more structural and hierarchical features.After channel fusion of input images,the channel feature maps are linearly combined with 1×1 convolution layer to realize cross-channel interaction and information fusion.The experimental results show that the combination of LBP texture images in CNN can effectively improve the results of facial beauty prediction.(3)A facial beauty prediction model based on multi-layer convolution features fusion is used.Unlike traditional convolutional neural network,convolutional neural network of multi-layer features fusion combines multiple shallow features and intermediate features with deep features to enhance the process of facial features extraction.Not only the semantic information of the deep convolutional layers,but also the detail information of the shallow convolutional layers are considered,which makes the facial beauty prediction more accurate.The experimental results show that the model fusing shallow and deep convolution features outperforms the single-layer model.(4)Sequeeze and Excitation(SE)blocks are embedded in the multi-layer convolution features fusion model,which can adaptively recalibrate channel-wise feature responses by modelling channel relationships,so that the model can selectively emphasize valuable features and suppress useless ones by global information.Our proposed method exceeds the best performance on the Large Scale Facial Beauty Database(LSFBD)in comparison with the other existing work.