面向人脸年龄估计的选择性多层融合神经网络

Selective Multilayer Fusion Convolutional Neural Network for Facial Age Estimation

针对使用传统神经网络算法进行人脸年龄估计可能会带来梯度弥散或梯度爆炸引起网络退化的风险,提出了选择性多层融合的卷积神经网络结构(SMLF-Net,Selective multilayer fusion convolutional neural network)。首先在SMLF-Net设计中融合了类似SE(Squeeze-and-Excitation Networks)结构增强了模型的非线性和特征重标定,其次为了提高深层网络模型的分类精度,引入了类似RE(ResNets)结构解决网络退化问题,最后根据卷积核移动步长动态选择网络结构避免网络参数过多。该方法通过构建RE与SE结构提取人脸的高维、中维、低维特征,并使用BN(Batch Normalization)操作降低网络训练难度。结果在扩充亚洲数据后的IMDB-WIKI数据集上对本文方法进行训练,并在树莓派上进行测试,结果显示SMLF-Net方法能够取得平均绝对误差(MAE)为3.09的估计精度,优于Google InceptionNet方法(MAE)为3.32和ShuffleNet方法(MAE)为3.54,充分证实了本文所提方法的先进性和有效性。

Facial age estimation using traditional neural network algorithm can bring gradient dispersion or gradient explosion,which leads to the risk of network degradation.Hence,a Selective multilayer fusion convolutional neural network structure is proposed.Firstly,the SMLF-Net with SE-like(Squeeze-and-Excitation Networks)structure is designed to enhance the model’s nonlinearity and feature recalibration.Secondly,in order to improve the classification accuracy of deep network model,a structure similar to RE(ResNets)is utilized to solve the network degradation problem.Finally,the network structure is selected dynamically according to the moved step of convolution kernel to avoid excessive network parameters.The proposed method builds RE and SE structures to extract high-dimensional,middle-dimensional and low-dimensional features of human faces.Meanwhile,it uses Batch Normalization to reduce network training difficulty.Imdb-wiki data set after Asian data expansion is used to train the proposed methed and tested on Raspberry Pi.The results show that the average absolute error(MAE)of SMLF-NET is 3.09.This result is better than Google InceptionNet with the average absolute error 3.32 and ShuffleNet with the average absolute error 3.54.The results demonstrate adequately the advance and effectiveness of the proposed method.