基于部件检测与检索的行人精细化分割  

Fine Pedestrian Segmentation with Parts Detection and Retrieval

作　　者：王枫[1] 厉智[1] 刘青山[1] 孙玉宝[1] WANG Feng;LI Zhi;LIU Qing-shan;SUN Yu-bao(B-DAT Lab,Collaborative Innovation Center,School of Information & Control,Nanjing University of Information Science and Technology,Nanjing,Jiangsu 210044,China)

机构地区：[1]南京信息工程大学信息与控制学院江苏省大数据分析技术实验室,江苏南京210044

出　　处：《电子学报》2019年第2期502-508,共7页Acta Electronica Sinica

基　　金：国家自然科学基金(No.61532009)

摘　　要：针对行人图像外观的多样性以及结构、姿态、场景的复杂性,提出一种有效的精细化行人部件分割方法.该方法实现把一幅行人图像分割成不同的语义区域,主要包含三个阶段,前两个阶段单独训练两个Fast R-CNN(Fast Region-based Convolutional Neural Network,快速区域卷积神经网络)模型,分别用来检测整个人体以及各个部件以获得各类别部件的大体位置;第三个阶段使用基于检索过分割图像的方法来对检测到的各个部件进行分割,最后把各部件分割结果还原到原图坐标上以得到最终的分割结果.实验表明所提方法在三个公开的数据库上,与其他算法相比,分割准确率更高,边缘效果更好.Focused on the diversity of appearance and the complexity of configuration,laying,and occasion in human images,a coarse-to-fine method was proposed for effective human parsing.It can decompose a human image into semantic regions which consists of three phases.In the first two phases,two effective models were trained with Fast Region-based Convolutional Network(Fast R-CNN)to respectively detect human body and clothing items.In the third phase,parsing clothing items based on retrieving similar over-segmented images and morphing them into absolute image coordinates.Experiments are conducted on three public databases ,and the experimental results show that proposed method has higher accuracy and promising performance.

关 键 词：行人分割 快速区域卷积神经网络 过分割 部件检索 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]