Application Research On Facial Beauty Prediction Based On Semi-supervised Learning

Facial beauty prediction is a forefront subject in the field of artificial intelligence,which aims to investigate how to enable computers to judge or predict the beauty of a face like humans.Facial beauty prediction is widely used in various industries,such as animation industry and plastic and beauty industry,etc.And it has a vast potential for progress and expansion.At present,the field of facial beauty prediction still faces some difficulties,such as poor generalization ability,insufficient amount of training data and easy overfitting.To address the issues with conventional deep networks,research is being conducted in relevant areas.Facial beauty prediction based on deep learning method also has insufficient label data samples and lack of the ability to capture global information in feature extraction.Therefore,three important problems in the process of facial beauty prediction are studied in this thesis.Firstly,facial beauty prediction label data samples are insufficient.Secondly,the characteristic information of different scales is not well fused.Thirdly,facial beauty prediction lacks the ability to capture global information in feature extraction.This thesis focuses on the following key research areas:(1)Semi-supervised learning can utilize labeled data and unlabeled data to improve the accuracy of the model.The purpose of semi-supervised learning is to continuously improve the study capabilities of the model through newly annotated data to maximize it.And semi-supervised learning minimizes the cost of annotated data as much as possible.In view of the problem that unlabeled data can’t be effectively utilized by supervised learning,semi-supervised learning is applied in facial beauty prediction in this thesis,and a large number of relatively cheap unlabeled data are used to improve the performance of facial beauty prediction.The method applies two similar network models,namely Convolutional Neural Network(CNN)enhanced by Dropout data and a CNN network enhanced by the other data.The tagged and untagged facial beauty images were fed into the two networks at the same time,producing different predictions.At the same time,the square variance of the two prediction results is taken as the consistency loss function,and the cross entropy loss function is calculated by the labeled face image.The total loss function is the weighted sum of the two loss terms,and the final facial beauty prediction result is output.Experimental results based on Large Scale Asian Facial Beauty Database(LSAFBD)and SCUT-FBP5500 database show that the semi-supervised learning method can effectively improve the accuracy of facial beauty prediction.It can be widely applied in image classification and image recognition,etc.(2)Semi-supervised learning can address the issue of insufficient amount of labeled data,but it can’t learn the features of different scales well when extracting facial beauty image features.Aiming at the problem that semi-supervised learning can’t learn features of different scales well,this thesis introduces attention feature fusion,which can fuse multi-scale features to enrich the information of feature graph and give corresponding weights to the feature maps of different levels to balance the feature information with different scales.In order to better integrate features of different scales and extract context information from different scales,the method of semi-supervised learning and attention feature fusion is applied in facial beauty prediction.Firstly,data enhancement is carried out for labeled sample images and unlabeled sample images respectively.Secondly,the features of facial beauty images are extracted by way of attention feature fusion through convolutional neural network.Finally,the conformance loss function is introduced to adjust the model parameters of the CNN and the attention feature fusion network.Experimental results based on LSAFBD database and SCUT-FBP5500 database show that the image features extracted from the network model are more conducive to face beauty prediction,and the maximum accuracy of the presented method is better than that by conventional method.It can be widely applied in the fields of image classification and image recognition,etc.(3)Semi-supervised learning and attention feature fusion can utilize local information to extract facial beauty features,but lack the ability to extract global information.Vision Transformer can leverage the self-attention mechanism to capture global context information and build global dependencies on facial beauty features to extract stronger features.In view of the lack of global information extraction ability of semi-supervised learning and attention feature fusion,this thesis combines semi-supervised learning with Vision Transformer and applies it in facial beauty prediction.Semi-supervised learning can solve the problem of insufficient tag data in facial beauty database.Vision Transformer can solve the problem of insufficient global dependence on facial beauty features,so as to better extract features from facial images.Firstly,data enhancement is carried out for labeled sample images and unlabeled sample images respectively.Secondly,Vision Transformer is used to extract features of facial beauty images.Finally,the consistency loss function is applied to adjust the network model,and the facial beauty images are forecast.Experimental results based on LSAFBD database and SCUT-FBP5500 database show that the presented method is better than conventional method,and the highest accuracy is better than that by conventional method.It can be widely applied in the fields of image classification and image recognition,etc.