A Large-scale Facial Beauty Prediction Research Based On Deep Feature

Facial beauty prediction(FBP)as an important task of artificial intelligence has potential market value,but also exists great challenges.There are considerable demand of entertainment,beauty,virtual media and other business areas for FBP,but it also faces with many problems,in which data size is insufficient,the face image is hard to classify,and the deep feature lacks research.So,a database of LSAFBD(Large Scale Asian Female Beauty Database)will be constructed in this paper,which contains total of 20,000 5-classes images that provide a better verification platform in evaluation and improvement of FBP algorithms.Taking LSAFBD as the benchmark test database,the key points of study in this paper mainly includes the following contents.(1)A discussion of traditional FBP.The traditional method of FBP is researched in early phase.Currently,many achievements in scientific research has been obtained.In this paper,LBP(Local Binary Pattern)and LPQ(Local Phase Quantization)using for feature extraction and SVM(Support Vector Machines)using for classification are considered as breakthrough of research.Then,the Raw Pixel feature and the shallow feature of K-means are introduced for comparison.The experimental results show that the shallow feature achieved better prediction results.Besides,we get a conclusion that the image details play an important role on FBP.Therefore,the review of traditional FBP method and reproducing its experimental result could direct the following algorithm improvement.(2)FBP based on CNN(Convolutional Neural Network).CNN has shown a great advantage in the field of image recognition in which largely due to its end-to-end learning ability in image feature extraction.Therefore,CNN methods are widely studied by scholars.Based on the CNN principle,a preliminary deep network Proposed Net was designed by this paper which is suitable for FBP.The experimental results show that,compared with other main stream CNN models,Proposed Net took a good prediction result,and its Pearson correlation was only slightly bellowed to Net B,and its classification rate achieved the best results.By comparison of the results,it is concluded that the CNN method is fully superior to the traditional method both in various data sets and prediction method of regression or classification.(3)Improvement of CNN model.First,a double activation layer is introduced to optimize the structure of Proposed Net.Secondly,a cost-sensitive loss function is proposed to suppress the overfitting problem when the CNN model is trained by small database.Finally,a deep feature fusion training method is study in this paper,the problem of insufficient training dataset will be solved by transform learning from a large-scale face database,and the prediction accuracy is further improved by deep fusions' feature.Experimental results show that the proposed methods could greatly improve the prediction performance of Proposed Net.Its classification accuracy was 64.8%,which was higher than Net B of 61.7%,and its regression correlation was 0.882,which was higher than Net B of 0.858.The experimental results also indicate that the color characteristics of images obtained the better performance in FBP.(4)Design of real-time FBP system.Based on the open source library of Caffe,Opencv and Dlib,a Multifunctional real-time FBP system integrated the proposed algorithm was designed to classify the gender or predict the value of facial beauty.The system uses multiple preprocessing methods to filter unqualified faces,introduces a gender classifier and single-gender FBP models to increase system prediction accuracy rate.The experimental results show that the system has a higher real-time running performance and strong robustness in gender or facial beauty prediction.Through the design of real-time system and the analysis of experimental results,it provides a perfected theoretical support for the advanced FBP system promotion and application.