A Study For Facial Beauty Prediction Based On Deep Learning

Facial beauty prediction is a prospective research topic in machine learning field. Currently, most modern research of facial beauty focuses on geometric features and apparent features by traditional machine learning methods. Geometric features rely heavily on accurate manual landmark localization of facial features and impose strict restrictions on training samples. Apparent features do not require manual intervention, and have a unique advantage compared with geometric features. In addition, geometric features and some apparent features focus on a particular aspect of face description, such as distance, proportion, and texture, causing a loss of face image information characterizing facial beauty. Facial beauty prediction is a spring-up research in the world, so there are few training sample libraries and more of them are not public. Therefore, the lack of training samples brings a great difficulty to the research. To solve these problems, we present a face beauty prediction model based on deep deconvolutional networks(ADN). The main work of this paper includes the following aspects:(1) This paper adopts Adaptive Deconvolutional Networks(ADN) with four layers for feature extraction and visualization analysis. ADN has the advantage of hierarchical feature extraction like convolutional neural network(CNN), which can extract multilayer apparent features from input images unsupervisedly. ADN’s feature extraction process is in accordance with the hierarchical visual perception mechanism of human brain. Meanwhile, ADN can observe the network parameters and features by visualization methods, which provide scientific and visually basis for facial beauty prediction. To ensure the integrity of facial information, ADN uses the error between the reverse reconstruction of high-level features and the original images via network training. Because the training of ADN is unsupervised completely, in the case of labeled samples difficult to obtain,a large number of unlabeled samples are added to improve the network’s understanding of data distribution.(2) Through visualization analysis, we found that ADN accorded with the hierarchical regional perception of the human cortex, and it was a process of continuous abstraction and extraction. Meanwhile, through the reverse reconstruction of single or multiple high-level features, we found that high-level features corresponded to the local of face image. Experimental results show that the facial beauty rank is the outcome of combined action of the local features, such as eyes, nose, mouth, eyebrows, face contour, etc. This paper breaks through traditional quantitative description for facial beauty, and we obtain more structured and hierarchical expression through multilayer network.(3) We get a lager number of images from social networking services and job websites through the web crawler technology, then filter these images and do geometric correction using face detection technology. This paper adopts Adaboost detection algorithm of OPENCV to filter images and do geometric correction based on the face region coordinates and the key characteristic point coordinates. Through filtering and correction, we get standard positive face images which will be used for the model of training and testing.(4) Three regression methods, which includes Logistic, SVM and KNN, are used to analyze the correlation between multi-layer apparent features extracted by ADN and the facial beauty degree. Through regression analysis, computer can simulate human brain and do statistical analysis on the facial features, so we can predict the beauty degree for an unknown face image. Experimental results show that SVM combined with ADN-3 works better and get a higher correlation, it also demonstrated that apparent features extracted by ADN are better to describe the experimental data.