Facial Attractiveness Prediction Using Deep Learning Method

Facial attractiveness has allured humans for centuries.There are many studies in the field of psychology,anthropology and biology,but relatively less in computational science.Traditional machine learning methods of facial attractiveness prediction usually extract geometric features or apparent features to represent facial beauty,which can easily lead to the loss of important information.Deep learning is a new area in the field of machine learning.As a kind of data-driven model,deep learning learns the data characteristics automatically.Deep learning uses a cascade of many layers of nonlinear processing units for feature extraction and transformation,which is believed to be more representable than traditional methods.Deep learning has made great progress in the study of image classification and speech recognition.However,in the subject of human facial attractiveness prediction,not much studies have been done.In this article,we take advantages of deep learning method to explore facial attractiveness prediction.The main work of this thesis are as follows:(1)Three kinds of deep learning methods are adopted to predict human facial beauty,which are MLP(Multi-Layer Perceptron),CNN(Convolutional Neural Network)and PCANet(Principle Component Analysis Network).Multiple structures are proposed to explore what kind of network is more representable.(2)A new kind of apparent feature is extracted based on cognitive psychology,which is called smoothness layer.It has been proved by a plenty of experiments that smoothness layer is more representable than traditional features such as eigenface,Gabor filters and LBP texture.(3)Two models for facial attractiveness prediction are proposed based on convolutional neural network and cognitive psychology knowledge,which are MC-CNN(Multi-column CNN)and CF-CNN(Cascaded Fine-tuning CNN).MC-CNN gets higher correlation result but CF-CNN is an end-to-end facial beauty prediction model.(4)A plenty of experiments are done on three different databases to test the performance of our proposed models and features.The analysis of the results reveals the effectiveness of our models.Furthermore,we visualized the features that the CNN has learned,and find that CNN focuses on facial skin,outline and features,which is consistent with human perception.