基于混合域注意力机制的服装关键点定位及属性预测算法  

Clothing key points location and attribute prediction algorithm based on mixed domain attention mechanism

作　　者：雷冬冬 王俊英[1,2] 董方敏[1] 臧兆祥[1] 聂雄锋 LEI Dongdong;WANG Junying;DONG Fangmin;ZANG Zhaoxiang;NIE Xiongfeng(Hubei Key Laboratory ofIntelligent Vision Based Monitoringfor Hydroelectric Engineering,Three Gorges University,Yichang443002,China;Hubei Province Engineering Technology Research Centerfor Construction Quality Testing Equipments,Three Gorges University,Yichang443002,China)

机构地区：[1]三峡大学水电工程智能视觉监测湖北省重点实验室,湖北宜昌443002 [2]三峡大学湖北省建筑质量检测装备工程技术研究中心,湖北宜昌443002

出　　处：《东华大学学报（自然科学版）》2022年第4期28-35,共8页Journal of Donghua University(Natural Science)

基　　金：国家自然科学基金新疆联合基金重点项目(U1703261);湖北省水电工程智能视觉监测开放基金项目(2017SDSJ04)。

摘　　要：针对服装形变和模特复杂姿态影响服装视觉分析准确率的问题,提出一个基于混合域注意力机制的服装关键点定位与属性预测算法,该算法利用循环十字交叉注意力(recurrent criss-cross attention,RCCA)模块得到服装图像的每个像素的上下文信息,从而捕获服装关键点之间潜在的空间几何关系,再融合服装图像的空间联系和通道交互信息来获得更好的服装关键点定位和属性预测效果。服装的空间特征由空间注意力分支网络在关键点热图的基础上学习得到,而通道交互信息通过局部跨通道交互策略生成通道注意力来捕获。试验结果表明,所提算法降低了服装关键点定位的归一化误差,并在一定程度上提高了服装的分类与属性预测效果。Aiming at the problem that clothing deformation and complex posture of models affect the accuracy of clothing visual analysis,an algorithm of clothing key points location and attribute prediction based on mixed domain attention mechanism is proposed.The algorithm uses the RCCA(recurrent criss-cross attention)module to obtain the context information of each pixel of clothing to capture the potential spatial relationship among the clothing key points,and fuses the spatial connection and channel interaction information of clothing image to obtain better clothing key point positioning and properties prediction effect.The spatial features of clothing are learned by the spatial attention branching network based on the heat maps of key points,while the channel interaction information is captured by the local cross-channel interaction strategy to generate channel attention.Experimental results show that the proposed algorithm reduces the normalization error of the positioning of clothing key points,and improves the effect of clothing classification and attribute prediction to a certain extent.

关 键 词：服装关键点定位 属性预测 混合域注意力机制 非局部空间连接 局部跨通道交互 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]