基于改进残差网络的人脸表观年龄估计

Estimation of face apparent age based on improved residual network

针对现有卷积神经网络(CNN)模型对人脸年龄特征表达能力不足、识别精度不高和模型参数量大的问题,提出一种融入注意力机制的残差网络表观年龄估计方法。首先,对卷积块注意力模块(CBAM)进行优化改进,在获取通道重要度权重的基础上,采用特征融合再训练的方法,混合计算两部分的池化压缩特征互信息,增强关键性通道特征的表示;然后,将改进的CBAM融入到ResNet模型的残差结构中,与残差结构的特征提取层以先后串行结构融合构建新的残差模块;最后,利用SoftMax进行分类获得最终估算年龄。在APPA-REAL和LAP2015两个数据集的实验表明:本文方法比原ResNet模型取得了更好的平均绝对误差指标,并且与其他相关方法的对比也证明了其有效性。

Aiming at the problems of insufficient expression ability of face age features,low recognition precision and large amount of model parameters in the existing convolutional neural network(CNN)models,an apparent age estimation method based on residual network integrating attention mechanism is proposes.Firstly,the convolutional block attention module(CBAM)is optimized and improved.Based on obtaining the channel importance weight,the feature fusion retraining method is used to calculate the pooled compressed feature mutual information of the two parts,in order to enhance the representation of key channel features.Then the improved CBAM is integrated into the residual structure of ResNet model,and fused with the feature extraction layer of residual structure to construct a new residual module in sequential serial structure.Finally,the final estimated age is obtained by classification using SoftMax.Experiments on APPA-REAL and LAP2015 datasets show that the proposed method achieves better average absolute error than the original ResNet model,and its effectiveness is also proved by comparison with other related methods.