Application Research Of Fusing Noise Label Learning And Multi-task Learning For Facial Beauty Prediction

Facial beauty prediction is frontier topics,which enables computers to have the ability to judge facial beauty and can be applied in many fields such as cosmetic surgery,social network recommendation,facial image beautification,and character prototyping.At present,facial beauty prediction still suffers from insufficiently supervised information and models are susceptible to noise labels.Noisy label learning can be used to reduce the impact of noisy labels in facial beauty prediction by building noisy models,designing loss functions or regularizers.Multi-task learning can improve the accuracy of facial beauty prediction by exploiting the effectively supervised information of multiple facial beauty prediction related tasks.By fusing noisy label learning with multi-task learning,not only can the problem of insufficiently supervisory information in facial beauty prediction be solved by using multi-task supervisory information,but also the impact of noisy labels in facial beauty database can be reduced.Therefore,the fusion of multi-task learning and noisy label learning is presented in this thesis and applied in facial beauty prediction.The research in this thesis consists of the following:(1)Deep convolutional neural networks applied in facial beauty prediction suffer from the problem of over-fitting noise labeled samples,which affects the generalization of deep convolutional neural networks.Thus,A self-correcting noisy labeling method applied in facial beauty prediction is presented in this thesis,which can decrease the negative effect of noisy labels.A mechanism of self-training teacher models and a mechanism of relabeling and retraining are used in this method.The teacher model is obtained by the self-training teacher model mechanism in a self-training manner,which helps the student model to select and train samples corresponding to clean labels until the student model outperforms the teacher model in generalization and becomes the new teacher model.The process is repeated consistently.With the mechanism of relabeling and retraining,the noisy labels are corrected by comparing the maximum prediction probability with the corresponding prediction probability of the labels.Thus,the label-corrected data is used to repeatedly adopt the mechanism of self-training teacher models.Experimental results show that the method can reduce the impact of noisy labels under the condition of synthetic noisy labels.Meanwhile,the prediction accuracy based on the Large Scale Asia Facial Beauty(LSAFB)database,SCUT-FBP5500 database is higher than that by conventional methods.The method also has the characteristics of selecting clean samples and making full use of all the data.(2)Multi-Task Attention Network(MTAN)is able to use multiple label types of data from a single database for supervised training,but ignores the problem of ineffective multi-task training which multiple databases having only one label type.Therefore,this thesis presents the Dual-Input Dual-Task Attention Network(DIDTAN)based on MTAN,which extends the Batch Normalization(BN)layer shared by tasks in MTAN to different task-specific BN layers.The Neural Discriminative Dimensionality Reduction(NDDR)module is introduced to constrain the expression of shallow layer features,while the Deep CORrelation Alignment loss function(Deep CORAL)is used to constrain the representation of fully connected layer.Experimental results based on LSAFB database,SCUT-FBP5500 database,and Celeb A database show that dual-input dual-task attention network can improve the accuracy of facial beauty prediction and is higher than that by the benchmark method.In addition,the network solves the problem of insufficient supervisory information by exploiting supervisory information from two singlelabel type facial beauty databases simultaneously.(3)The dual-input,dual-task attention network is able to use supervisory information from both input data to solve the problem of insufficiently supervisory information,but the model is still susceptible to noisy labels.With the mechanism of relabeling and retraining,the noisy labels are corrected by comparing the maximum prediction probability with the corresponding prediction probability of the labels.The label-corrected data is used to retrain the model,effectively reducing the impact of the noisy labels.Thus,by incorporating DIDTAN into the relabeling retraining mechanism,a fusion model with the relabeling retraining mechanism is constructed.The fusion model not only to utilize the supervised information of both input data,but also to reduce the impact of noisy labels.Experimental results show that the dual-input,dual-task facial beauty prediction based on LSAFB database,SCUT-FBP5500 database achieves the accuracy of 65.4%,which is higher than that by conventional method.