Facial Beauty Assessment Methods Based On Deep Learning

The golden ratio hypothesis,the “three courts five eyes” hypothesis and so on are all human's explorations on the evaluation standard of facial beauty.With the development of the society,more and more people concern facial beauty and yearn for having a more attractive face.In recent years,the deep network of facial beauty have been trained because of the development of deep learning,espically in the field of image processing,and the construction of face beauty data set.This thesis studied the facial beauty assessment algorithm based on deep learning and the SCUT-FBP5500 data set,which mainly includes the following three aspects:1.Facial beauty assessment network based on long short-term memory unit and key landmarks.The features of different positions in the face values different for the problem of facial beauty assessment.Besides,the average pooling reduces the difference of the features.In this thesis,the long short-term memory unit was used to replace the average pooling layer.The features corresponding to coordinate of the key landmarks were extracted,which were used as the input of the long short-term memory unit.This thesis obtained the better experimental results by comparing the longitudinal experimental results with different number of key landmarks,and also proved the effectiveness of the structure by the transverse experimental results based on ResNeXt-50 and ResNet-50.2.Facial beauty assessment network based on attention mechanism.Attention module can make the network find the area with higher relevance to the target task,increase its weight,and make the network pay more attention to the area.Because the dataset is relatively small,it is easy to produce overfitting phenomenon,this thesis compared the experimental results of embedding spatial attention module and channel domain module in different convolution blocks.The experimental results showed that embedding spatial attention module in shallow layer and channel attention module in deep layer are the best for facial beauty assessment.3.Facial beauty assessment network based the attention mechanism and faceial characteristic image.Facial characteristic image(facial texture image or facial key landmarks image)could make the network focus on the areas with higher relevance to facial beauty.Therefore,this thesis constructed a branch network with facial feature image as input and learned the space weight added in the shallower backbone network,which can further reduce the error and improve the relevance between the predictions and the groundtruths.