基于稀疏表示的自动年龄估计

来源期刊:中南大学学报(自然科学版)2015年第3期

论文作者:梁毅雄 李玲芝 艾玮 刘凌波

文章页码:878 - 886

关键词:模式识别;年龄估计;稀疏表示;空间金字塔匹配;弹性网

Key words:pattern recognition; age estimation; sparse representation; spatial pyramid matching; elastic net

摘    要:将稀疏表示同时应用于人脸图像年龄特征提取和年龄自动估计2个关键环节,提出一种基于稀疏表示的年龄估计新方法。该方法首先对人脸图像进行稠密采样,提取底层的SIFT描述子,训练字典对其进行稀疏编码,再进行空间金字塔表示,并将其作为刻画年龄属性的图像特征,然后采用线性稀疏回归模型同时选择特征进行年龄估计。针对人脸老化过程具有非线性特点,提出一种基于分段线性策略的层次模型,即首先训练若干个分类器将人脸粗分类到不同的年龄段,然后在该年龄段中训练对应的线性模型对年龄进行精确估计。在权威的FG-NET和MORPH人脸库上对该方法的有效性进行实验验证。研究结果表明:所提出的方法在FG-NET人脸库上年龄估计偏差的平均绝对误差为3.79,远比当前最好方法的平均绝对误差低,而在MORPH人脸库上的平均绝对误差为6.46,与当前最好方法的平均绝对误差相当。

Abstract: A novel sparse representation-based age estimation method was proposed. The proposed method integrated sparse representation into two key components of a successful age estimation system, i.e. age-related feature extraction and age estimation by regression. Specifically, during the feature exaction stage, the mid-level spatial-pyramid face representation based on sparse codes of dense SIFT features (ScSPM) was used to characterize the age-related variance, while during the age estimation step, the linear sparse regression models were learned which can not only select the most discriminative features but also perform the age estimation. The hierarchical strategy, which first coarsely classified the faces into age groups and then finely estimated the detailed age by the linear regression model of this group, was adopted to deal with the non-linearity attribution of aging to improve the performance of the age regression model. The effectiveness of the proposed method was validated on two benchmarks, i.e. the FG-NET database and the MORPH database. The proposed approach achieves the MAE of 3.79 in the FG-NET database and the MAE of 6.46 in the MORPH database, which shows that the proposed approach outperforms the state-of-the-arts on the FG-NET database and achieves competitive performance on the MORPH database.

有色金属在线官网  |   会议  |   在线投稿  |   购买纸书  |   科技图书馆

中南大学出版社 技术支持 版权声明   电话:0731-88830515 88830516   传真:0731-88710482   Email:administrator@cnnmol.com

互联网出版许可证:(署)网出证(京)字第342号   京ICP备17050991号-6      京公网安备11010802042557号