基于多特征融合的人脸颜值预测

The Prediction of Facial Beauty Based On Multi- Feature Fusion

随着计算机技术的迅速发展以及人脸识别技术的成熟,人脸颜值受到越来越多的关注和研究。针对单一的人脸特征无法表征完整的人脸信息,提出一种几何特征、表观特征相融合的人脸特征,该特征综合考虑了人脸的关键特征点、几何距离及关键部位面积和全局特征对人脸颜值的影响。数据信息来源于LFW数据库和互联网的1500张非限制性东西方女性人脸图像数据进行建模,对源于社交网络的100张女性人脸图像在KNN和BP神经网络两种分类器下进行实验。实验结果表明融合特征在不同分类器模型下一致性更好,Pearson相关系数提升了2.3%,平均绝对误差下降了2.7%,均方误差下降了2.9%。因此本文提出的融合特征与单一特征相比,人脸颜值的预测性能有所提高。

With the development of computer science and face recognition technology in recent years,face beauty has been drawing more and more attention. Single facial feature unable to characterize the whole facial information,to explore face beauty from the perspective of face features,we propose a fused feature of geometry and apparent features. The geometric distance of face organ and global characteristics are considered to determine the face beauty. LFW database and 1500 unrestrained women face images from the internet are used to build prediction model in this paper. Face beauty of 100 women face images from social network are predicted under the model with KNN and BP neural network classifiers. Experimental results show that better consistency was found with fused features in different classification model,and the Pearson correlation coefficient improved by 2. 3%,the average absolute error decreased by2. 7%,the mean squared error decreased 2. 9%. The experiment results show the performance of the face beauty prediction with our proposed fused feature is better compared with the results with independent features.