基于深度学习的人脸美丽预测模型及其应用

摘要

为了进一步提高人脸美丽预测精度,本文构建了一个多尺度图像输入的人脸美丽预测深度卷积神经网络模型(Deep Convolution Neural Network,DCNN),以增强对人脸图像空间结构特征的提取能力.采用深度可分离卷积层代替普通卷积层、Max-Feature-Max(MFM)激活函数代替修正线性单元(Rectified Linear Unit,ReLU)激活函数,可减少网络训练参数并提取具有竞争性的网络特征.基于大规模亚洲女性人脸美丽数据库(Large Scale Asian Female Beauty Database,LSAFBD)的实验结果表明,本文所构建的人脸美丽预测模型取得了59.75％的正确分类率,优于现有DCNN模型的分类结果.%In order to improve the accuracy of facial beauty prediction, a DCNN model which can extract multi-ply spatial structure features by taking in multi-scale face images is constructed. The Max-Feature-Max (MFM) activation function is used instead of the modified linear unit (Rectified Linear Unit, ReLU) activation function to reduce the network training parameters and extract competitive network features. Experimental results based on the Large Scale Asian Female Beauty Database (LSAFBD) show that our DCNN model achieves the accuracy of 59.75％, which is superior to the existing DCNN model classification results.