结合特征融合和任务分组的人脸属性识别  

Face Attribute Recognition Combining Feature Fusion and Task Grouping

作　　者：刘英芳 王松[1,2] 马亚彤 LIU Yingfang;WANG Song;MA Yatong(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;Gansu Provincial Engineering Research Center for Artificial Intelligence and Graphic and Image Processing,Lanzhou 730070,China)

机构地区：[1]兰州交通大学电子与信息工程学院,兰州730070 [2]甘肃省人工智能与图形图像处理工程研究中心,兰州730070

出　　处：《计算机工程》2023年第11期211-219,共9页Computer Engineering

基　　金：国家自然科学基金(62067006);甘肃省自然科学基金(21JR7RA291);甘肃省教育科技创新项目(2021jyjbgs-05);甘肃省高校产业支撑计划项目(2020C-19)。

摘　　要：针对现有人脸属性识别模型存在的特征提取不足、划分属性组时未充分考虑属性相关性强弱等问题,为了实现对多个人脸属性的同时识别,建立一种结合特征融合和任务分组的人脸属性识别模型。在参数共享部分,通过多尺度特征融合模块将不同尺度的特征进行融合增强特征相关性,同时设计基于中心核对齐和谱聚类的属性分组策略进行属性识别,通过中心核对齐方法度量属性的相关程度,并以此为基础使用谱聚类算法得到属性的合理分组,使同一组内的属性相关性尽可能大,提高属性识别准确率。在分支部分,使用注意力机制加强对目标区域的关注,并通过不确定性加权方法表示任务间的相对难度,自动调整每组任务损失之间的相对权重,进一步优化模型性能。在CelebA公开数据集上的实验结果表明,所提模型的分类准确率相较于MOON、GNAS和DMM-CNN模型提升了0.78、0.09和0.02个百分点,参数量仅为上述对比模型的1.10%、17.08%和0.37%。To address the issues of insufficient feature extraction and inadequate consideration of attribute correlation in existing face attribute recognition models,a face attribute recognition model combining feature fusion and task grouping is proposed to achieve the simultaneous recognition of multiple face attributes.In the parameter sharing part,features of different scales are fused using multiscale feature fusion modules to enhance feature relevance.Simultaneously,an attribute grouping strategy based on Centered Kernel Alignment-Spectral Clustering(CKA-SC)is designed for attribute identification.The extent of correlation of the attributes is measured using CKA.Based on this,the SC algorithm is utilized to obtain reasonable attribute groups,maximizing the relevance of attributes within the same group to improve attribute recognition accuracy.In the branch part,attention mechanisms are employed to enhance the focus on target regions,and the Uncertainty Weighting(UW)method is used to represent the relative difficulty between tasks.This automatically adjusts the relative weights between the losses of each task group to further optimize the model performance.Experimental results on the publicly available CelebA dataset demonstrate that the proposed model achieves classification accuracy improvements of 0.78,0.09,and 0.02 percentage points compared with the MOON,GNAS,and DMM-CNN models,respectively,with parameter counts accounting for only 1.10%,17.08%,and 0.37%of the mentioned comparison models.

关 键 词：人脸属性识别 特征融合 中心核对齐 属性分组 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]