基于双有序性约束的人脸年龄估计研究

目的 人类年龄是人类识别和搜索任务中的重要特征,现有研究一般将人脸年龄估计视为传统的分类任务,忽略了年龄之间的有序特征,导致估计年龄与真实年龄之间的差距较大,因此,有必要寻找一种方法以缩小估计年龄与实际年龄的差距。方法 提出一种基于双有序性约束卷积神经网络模型(DO-CNN)的人脸图像年龄估计方法。首先,DO-CNN使用基于广义Logistic分布的有序回归模型作为卷积神经网络的分类器,并验证比其他有序分类器在人脸估计任务上的优越性;接着,进一步提出有序竞争比损失函数,在传统竞争比损失函数上,通过引入风险项使损失函数注意到预测年龄与真实年龄的误差,进而指导模型缩小估计年龄与真实年龄的差距。结果 在开源人脸图像年龄数据集FGNET和AgeDB上的对比实验显示：相比现有研究方法,DO-CNN分别提升约12%和3%的准确率,当允许的误差范围扩大后,该优势依然保持。此外,基于广义Logistic分布的有序回归分类器相比基于其他分布的有序回归分类器具有明显提升。结论 实验结果表明：基于双有序性约束的卷积神经网络模型可以明显提升人脸年龄估计的准确率,并减少年龄估计的实际误差。

Study on Face Age Estimation Based on Double-ordinality Constraints

Objective Human age is an important feature in human recognition and search tasks. Existing researchgenerally treats age estimation in facial images as a traditional classification task ignoring the ordered characteristics ofage and resulting in a large gap between the estimated age and the actual age. Therefore it is necessary to find a methodto reduce the gap between the estimated age and the actual age. Methods This paper proposed a method for ageestimation of face images based on a double-ordinality constrained convolutional neural network DO-CNN model.Firstly DO-CNN used an ordered regression model based on the generalized Logistic distribution as a classifier forconvolutional neural networks and verified its superiority over other ordered classifiers for face estimation tasks. Then anordered competitive ratio loss function was further proposed. By introducing a risk term into the traditional competitiveratio loss function the loss function considered the error between the predicted age and the actual age thus guiding themodel to reduce the gap between the estimated age and the actual age. Results Comparative experiments on the opensource facial image age datasets FGNET and AgeDB showed that compared with existing research methods DO-CNNimproved the accuracy by about 12% and 3% respectively and this advantage remains even when allowing for a largererror range. In addition the ordered regression classifier based on the generalized Logistic distribution exhibited significant improvements compared with ordered regression classifiers based on other distributions. Conclusion Theexperimental results show that the convolutional neural network model based on double-ordinality constraints cansignificantly improve the accuracy of age estimation in facial images and reduce the actual errors in age estimation.