多尺度YOLO人脸年龄估计方法研究

通过观察人脸估计年龄较为常见,但如何准确预测年龄则是一个难题。为提高人脸图像年龄估计的准确率,提出一种基于YOLO(You Only Look Once)模型的目标检测方法。将多尺度回归思想应用于卷积神经网络(Convolutional Neural Network,CNN),通过多尺度卷积改善模型对小尺寸目标的提取能力,结合特征通道分权重思想,改善特征提取操作中特征信息丢失的问题,构造决策树回归得到年龄估计。这种方法在人脸年龄图像库FG-NET上获得平均绝对误差(MAE)3.43,在GROUP数据集获得区间匹配度(AEM)62.4%。实验结果表明,通过多尺度特征回归以及通道权重分配,可以较为准确地进行人脸信息检测,并由此建立鲁棒性更强的人脸年龄估计模型。

Research on Face Age Estimation Based on Multi-Scale YOLO Model

To improve the accuracy of age estimation in face images, a feature extraction method based on YOLO model is proposed. The idea of multi-scale regression is applied to Convolutional Neural Network（CNN）, which improves the ability of extracting small-scale targets by multi-scale convolution. Combining the idea of feature channel weighting, the problem of feature information loss in feature extraction operation is improved. A decision tree regression is constructed to obtain age estimation. The Mean Absolute Error（MAE） is 3.43 on FG-NET, and the interval matching（AEM） is 62.4% on GROUP dataset. The experimental results show that the face information can be detected more accurately by multi-scale feature regression and channel weight allocation, and a more robust face age estimation model can be established.