基于堆叠沙漏网络的量体特征点定位  

Anthropometric Feature Points Localization Based on Stacked Hourglass Network

作　　者：邹昆 王伟灿[2] 董帅 李文生 ZOU Kun;WANG Wei-can;DONG Shuai;LI Wen-sheng(School of Computer Engineering,Zhongshan Institute,University of Electronic Science and Technology of China,Zhongshan Guangdong,528402;School of Computer Science and Engineering,University of Electronic Science and Technology of China,Chengdu,611731)

机构地区：[1]电子科技大学中山学院计算机学院,广东中山528402 [2]电子科技大学计算机科学与工程学院,成都611731

出　　处：《电子科技大学学报》2020年第5期709-717,共9页Journal of University of Electronic Science and Technology of China

基　　金：国家自然科学基金(61502088);广东省自然科学基金(2016A030313018)。

摘　　要：为提高复杂背景和任意着装情况下的量体特征点定位精度,将堆叠沙漏网络(SHN)引入人体图像量体特征点定位中,并针对SHN模型输出特征图分辨率过低导致定位精度不足的问题,构建了一种Deconv-SHN模型。一方面用多个反卷积层代替初始模型的输出层以提高输出特征图的分辨率,另一方面基于Smooth L1和局部响应对目标函数进行了优化。在自建的6700幅正面人体图像数据集上对Deconv-SHN模型、SHN模型以及传统算法进行实验的结果表明,Deconv-SHN模型在复杂背景和任意着装情况下的特征点定位精度较传统算法有显著提升,也明显优于SHN模型,基本满足人体参数测量应用的要求。In order to improve the accuracy of anthropometric feature point localization in complex background and arbitrary dress cases,the stacked hourglass network(SHN)is introduced into the localization of anthropometric feature points in body images.However,the resolution of the SHN model’s output feature map is too low to obtain high accurate feature points.So,a Deconv-SHN model is proposed to address this problem.On the one hand,the output layer of the initial model is replaced by several deconvolution layers to improve the resolution of the output feature map.On the other hand,the objective function is optimized based on Smooth L1 and local response.According to the experimental results on the self-built dataset consisting of 6700 human body images,the localization precision of the Deconv-SHN model in complex background and arbitrary dress cases is significantly higher than that of the traditional algorithm,which is also obviously superior to the SHN model,and basically meets the requirements of anthropometric applications.

关 键 词：量体特征点定位 深度学习 非接触式人体参数测量 堆叠沙漏网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]