| 摘 要: 在面部年龄估计任务中,传统方法常使用高斯分布模拟年龄的真实分布。然而,高斯分布不能充分模拟长尾特征,因此提出了一种分层拉普拉斯-指数标签编码(HLELE),并基于所提年龄编码构建了一个年龄估计方法。首先,所提方法通过分层拉普拉斯-指数标签编码将年龄标签转化为年龄分布,从而获取年龄标签之间的关系;其次,采用 VGG16(Visual Geometry Group16)卷积神经网络对面部图像进行特征提取,并对结果进行回归学习。实验结果显示,在 MORPHII(Multimedia Object Retrieval and Presentation in High Dimensional Spaces AlbumII)数据集上,该模型的平均绝对误差(MAE)达到2.55,相较于使用高斯分布的方法,该分层分布编码能够有效提升年龄估计模型的准确性。 |
| 关键词: 年龄估计;标签分布学习;年龄编码;损失函数 |
|
中图分类号: TP391
文献标识码: A
|
|
| Research on Age Estimation Based on Hierarchical Label Distribution Learning |
|
TANG Min, HU Chunlong
|
(School of Computer Science, Jiangsu University of Science and Technology, Zhenjiang 212114, China)
tangmin@stu.just.edu.cn; huchunlong@just.edu.cn
|
| Abstract: In facial age estimation tasks, traditional methods often employ Gaussian distributions to model the true distribution of age labels. However, Gaussian distributions fail to adequately capture long-tail characteristics. To address this problem, a Hierarchical Laplacian-Exponential Label Encoding (HLELE) is proposed, and an age estimation method is developed based on the proposed age encoding. Firstly, the proposed method converts age labels into age distributions through HLELE to capture relationships between age labels. Secondly, a VGG16 (Visual Geometry Group 16) convolutional neural network is utilized to extract facial image features, followed by regression learning on the results. Experimental results demonstrate that on the MORPH II (Multimedia Object Retrieval and Presentation in High-Dimensional Spaces Album II) dataset, the model achieves a Mean Absolute Error (MAE) of 2.55. Compared to methods using Gaussian distributions, the proposed hierarchical distribution encoding effectively improves the accuracy of age estimation models. |
| Keywords: age estimation; label distribution learning; age encoding; loss function |
