基于双分支残差网络的人脸年龄估计

Face Age Estimation Based on Two-branch Residual Network

摘要:

针对现有卷积神经网络难以捕捉人脸长距离依赖关系的问题,提出一种基于大卷积核的双分支残差网络,并用于人脸年龄估计。为突破传统小卷积核感受野较小的局限,首先,采用大卷积核对深度学习模型(ResNet)的残差模块进行改进,以此扩大网络的有效感受野,从而更高效地捕捉人脸图像中的全局信息与长距离依赖关系。其次,考虑到人脸细微特征对年龄估计的关键作用,引入细节下采样模块,该模块能在网络初始阶段最大程度减少细节信息的丢失。在网络结构设计上,创新性地将ResNet的原始残差模块与改进后的大卷积残差模块进行并联,构建形成双分支残差网络,并利用注意力模块实现两个分支之间的特征融合。为深度挖掘人脸年龄特征,在双分支残差网络之后串联两个大卷积残差模块,通过递进式特征抽象强化模型对复杂年龄模式的建模能力。最后,针对年龄估计中标签序数特性带来的挑战,将所构建的双分支残差网络与序数回归方法结合,通过将年龄值转化为有序标签序列进行建模,有效提升了模型对年龄变化的分辨能力。实验结果表明,所提出的方法在UTK-FACE数据集上的平均绝对误差最多降低了0.46,在FG-NET数据集上的平均绝对误差最多降低了0.09。

Abstract:

To address the problem that existing convolutional neural networks have difficulty in capturing long-range dependencies in human faces, a dual-branch residual network based on large convolution kernels is proposed and applied to face age estimation. Firstly, to overcome the limitation of small receptive fields of traditional small convolution kernels, large convolution kernels are adopted to improve the residual module of the deep learning model (ResNet), thereby expanding the effective receptive field of the network and more efficiently capturing the global information and long-range dependencies in face images. Secondly, considering the crucial role of facial fine features in age estimation, a detail downsampling module is introduced, which can minimize the loss of detailed information at the initial stage of the network. In the design of the network structure, the original residual module of ResNet and the improved large convolution residual module are innovatively connected in parallel to form a dual-branch residual network, and an attention module is utilized to achieve feature fusion between the two branches. To further enhance the deep mining of face age features, two large convolution residual mo-dules are concatenated after the dual-branch residual network, and through progressive feature abstraction, the model's ability to model complex age patterns is strengthened. Finally, in response to the challenges brought by the label ordinality characteristics in age estimation, the constructed dual-branch residual network is combined with ordinal regression methods. By converting age values into ordered label sequences for modeling, the model's ability to distinguish age changes is effectively improved. Experimental results show that the proposed method reduces the mean absolute error (MAE) by up to 0.46 on the UTK-FACE dataset and by up to 0.09 on the FG-NET dataset.