基于CPM及亲和度向量的服装关键点检测方法  

A Two-stage Clothing Key Points Detection Method Based on Affinity Vector and CPM

作　　者：汤启凡 黄晋 何儒汉 彭涛[2] 陈佳[2] TANG Qifan;HUANG Jin;HE Ruhan;PENG Tao;CHEN Jia(Hubei Provincial Engineering Research Center for Intelligent Textile and Fashion, Wuhan 430200, China;School of Computer and Artificial Intelligence, Wuhan Textile University, Wuhan 430200, China;Engineering Research Center of Hubei Province for Clothing Information, Wuhan 430200, China)

机构地区：[1]纺织服装智能化湖北省工程研究中心,湖北武汉430200 [2]武汉纺织大学计算机与人工智能学院,湖北武汉430200 [3]湖北省服装信息化工程技术研究中心,湖北武汉430200

出　　处：《郑州大学学报（理学版）》2022年第4期78-85,共8页Journal of Zhengzhou University:Natural Science Edition

基　　金：湖北省教育厅科学技术研究计划重点项目(D20141603);国家自然科学基金面上项目(61170093)。

摘　　要：服装关键点检测是时尚大数据分析和应用的关键技术之一,受到工业界和学术界的共同关注。对多类别、姿态复杂、遮挡等服装关键点检测难点进行了研究,使用目标检测方法对服装分类并消除背景干扰,提出基于亲和度向量的卷积姿态机实现服装关键点检测,利用关键点间空间约束以提高检测的准确度,实现了一个两阶段的服装关键点检测框架。实验表明,该框架能够更准确地检测服装关键点,并对服装关键点分散性、遮挡和重叠都具有较强鲁棒性。Clothing key point detection is one of the key technologies of fashion big data analysis and application,which has attracted the common attention of industry and academia.The difficulties of clothing key point detection such as multiple category,complex posture and occlusion were studied.A target detection method was proposed to classify clothing and eliminate background interference.A convolution posture machine based on affinity vector was proposed to detect clothing key points,and the space constraints between key points were used to improve the accuracy of detection.A two-stage clothing key point detection framework was implemented.Experiments showed that the framework could detect clothing key points more accurately,and had strong robustness to dispersion,occlusion and overlap of clothing key points.

关 键 词：服装关键点检测 卷积姿态机 亲和度向量 目标检测 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]