基于深度学习的抗年龄干扰人脸识别EI北大核心CSCD

随着年龄的增长,人脸的形状、纹理等特征会随之发生较明显的改变从而造成显著的类内干扰,这使得人脸识别的性能大大降低.为了解决上述问题,本文基于深度卷积神经网络将年龄估计任务和人脸识别任务相结合,提出了一种抗年龄干扰的人脸识别新方法AD-CNN(Age decomposition convolution neural network),首先将卷积块注意力模型(Convolutional block attention module,CBAM)嵌入到残差网络中以学习更具有代表性的面部特征,随后利用线性回归指导年龄估计任务,提取出年龄干扰因子,通过多层感知机将整个面部特征与年龄干扰特征投影到同一线性可分空间,最后从面部稳定的特征中将年龄干扰分离,得到与年龄无关的面部特征,并采用改进后的角度损失函数基于年龄无关的身份特征进行人脸识别任务,从而达到抑制年龄干扰的目的.本文在MORPH和FGNET数据集上的识别正确率分别达到了98.93%,和90.0%,充分证实了本文所提方法的先进性和有效性.

Age Invariant Face Recognition Based on Deep Learning

Facial appearances such as shape and texture are subject to significant intra-class variations caused by the aging process over time, resulting in the performance reduction of face recognition. To overcome this problem, this paper proposes a novel method (age decomposition convolution neural network, AD-CNN) based on deep convolution neural network to learn age-invariant face features. Firstly, the AD-CNN utilizes convolutional block attention module (CBAM) to extract facial features and estimates age factors by linear regression. Then, the facial features and age factors are projected into the same linear separable space by multi-layer perceptron. Finally, the age-invariant face features can be obtained by separating age factors from the whole facial features. Here, the improved angle loss function is considered to guide the training process. The proposed AD-CNN achieves 98.93%, and 90.0% recognition accuracy on MORPH and FGNET datasets, respectively, which demonstrates the AD-CNN with a great potential for age-invariant face recognition.