基于图卷积神经网络的人脸属性识别

Face Attribute Recognition Based on Graph Convolutional Neural Networks

人脸图像的多属性识别和多标签之间的依赖性建模研究,是计算机视觉和机器学习领域备受关注的研究课题。为借助多标签间的依赖关系提升识别效率,提出了一种基于图卷积神经网络的多标签人脸属性识别模型。该模型通过数据驱动的方式构建人脸属性间的有向图,并由图卷积神经网络将每个属性映射到对应属性分类器,以此对类别间的依赖关系进行建模。模型对图卷积神经网络中的相关矩阵和特征矩阵等关键元素进行了深入分析,使其能够胜任多标签人脸属性识别问题。实验结果表明,该模型在多标签人脸属性识别权威数据集CelebA上表现良好并能保持有意义的语义结构。

The research on multi-attribute recognition of facial images and the modeling of dependencies between multiple labels is a highly concerned research topic in the fields of Computer Vision and Machine Learning.A multi-label facial attribute recognition model based on Graph Convolutional Neural Networks is proposed to improve recognition efficiency by leveraging the dependency relationships between multiple labels.This model constructs a directed graph between facial attributes in a datadriven manner,and maps each attribute to the corresponding attribute classifier using a Graph Convolutional Neural Networks to model the dependency relationships between categories.The model has conducted in-depth analysis on key elements such as correlation matrix and feature matrix in Graph Convolutional Neural Networks,enabling it to handle multi-label facial attribute recognition problems.The experimental results show that the model performs well on the authoritative dataset CelebA for multilabel facial attribute recognition and can maintain a meaningful semantic structure.