深度PCANet模型的人脸美丽预测

深度卷积神经网络（Deep Convolution Neural Network, DCNN）在人脸识别、图像分类和目标检测领域已取得较好效果，并得到广泛应用；但是，在人脸美丽预测中却存在拟合效果欠佳、网络训练难度大等问题。深度PCANet模型，将深度主元分析网络（Principal Component Analysis Network，PCANet）作为特征提取器；采用无监督预训练提取网络参数，具有网络训练时间短、图像特征提取快等特点，能有效避免DCNN存在的问题。为此，本文将深度PCANet引入人脸美丽预测，对训练集图像采用多尺度预处理，训练深度PCANet。该模型可提取人脸图像的结构性全局特征，采用特征增强方法可生成更具表征能力的特征；运用线性支持向量机（Support Vector Machine, SVM）和随机森林（Random Forest, RF）回归器进行训练和预测。基于SCUT-FBP人脸美丽数据库的实验结果表明，深度PCANet模型具有结构简单、特征提取快和无需网络调参优化等特点；选择合适的图像尺度与采用特征增强方法可提高人脸美丽评价结果，证明了所提方法的有效性和可行性。 

The Facial Beauty Prediction of the Model of Deep PCANet

Deep convolution neural network(DCNN) has achieved good results in face recognition, image classification and object detection, and has been widely used. However, DCNN has some problems in the facial beauty prediction, such as bad fitting effect, hard training etc. Deep PCANet model, using Principal Component Analysis Network (PCANet) as feature extractor. The model can get parameters through unsupervised pre-training with the advantages of less time in training and faster speed in feature extraction, which can effectively avoid the problems of DCNN. For this reason, the deep PCANet is introduced into the facial beauty prediction, and the training set image adopts multi-scale preprocessing to train the deep PCANet. The model can extract the structural global features of the face image, and the feature enhancement method can generate features with more representative capabilities. Finally, linear Support Vector Machine (SVM) and Random Forest (RF) regressors were used for training and prediction. Experimental results for facial beauty prediction on SCUT-FBP database show that deep PCANet has advantages of simpler structure, faster feature extraction, and few parameter adjustment; choosing the right image scale and using feature enhancement methods can improve the face aesthetic evaluation results, which proves that the network is valid and feasible.