基于分类与回归混合模型的人脸年龄估计方法

Facial age estimation method based on hybrid model of classification and regression

针对现有人脸年龄数据库样本数量少、各年龄段分布不均匀的问题,提出了一种基于分类与回归混合模型的人脸年龄估计方法。该方法主要包含两个方面:特征学习和估计模式。在特征学习方面,利用已有的深度卷积神经网络(CNN),先在粗糙年龄标注数据集上预训练,再在现有的精确年龄标注数据库上微调,分别得到一个年龄段判别模型和两个年龄估计模型;在估计模式方面,该方法采用由粗到细的策略:首先,将人脸分入青少年、中年、老年和两个重叠区域这五个年龄段;然后,对于青少年和老年采用分类模型估计,对于中年采用回归模型估计,对于重叠区域采用两个模型估计的均值。所提方法在测试集上的平均绝对误差(MAE)为2.56。实验结果表明该方法受不同肤色和性别的影响较小,有较低的误差。

Focusing on small size and uneven distribution of current facial age database, an approach based on a hybrid model combined with classifier and regressor was proposed for facial age estimation. This approach mainly consisted of two aspects： feature learning and estimation method. In the aspect of feature learning, based on an existing Convolutional Neural Network （CNN）, an age group classifier and two age estimators were pretrained on the coarse dataset and then fine tuned on the accurate database. In the aspect of estimation method, a coarse-to-fine strategy was adopted. First, a facial images were classified into teenaged, middled-aged, elderly and two overlap groups. Next, the teenaged and elderly groups were estimated by the classifier model, the middled-aged group was estimated by the regressor model, and the two overlap groups were estimated by both models. The proposed approach can achieve a Mean Absolute Error （MAE） of 2.56 on the test set. The experimental results show that the proposed approach can reach a low error under different races and genders.