摘要
为精确检测不同人体尺度的关键点,提出一种基于高分辨率表征的关键点尺度变换网络(high-resolution for scale transformation structure,HR-STS)。由高低分辨率并行子网络提取所有初步关键点特征,通过尺度变换结构把关键点特征标准化,经过逆空间变换得到关键点坐标。实验对比结果表明,改进后的算法在MPⅡ数据集和COCO数据集上的平均检测精度提升明显,网络参数量与浮点运算量也小于其它算法。
For accurate detection of different body scales and key points,a key point scale transformation network based on high-resolution representation(HR-STS)was proposed.All preliminary key point features were extracted through the high-low resolution parallel sub-network,and the key point features were standardized by the scale transformation structure.The predicted key point coordinates were obtained by inverse space transformation.Through experimental comparison,the results show that the precision of the improved algorithm on MPⅡ and COCO is significantly improved.Meanwhile,the quantity of parameters and the computational complexity is smaller than that of other algorithms as well.
作者
宋玉琴
曾贺东
高师杰
熊高强
SONG Yu-qin;ZENG He-dong;GAO Shi-jie;XIONG Gao-qiang(School of Electronics and Information,Xi’an Polytechnic University,Xi’an 710600,China)
出处
《计算机工程与设计》
北大核心
2022年第4期1045-1051,共7页
Computer Engineering and Design
基金
中国纺织工业联合会科技指导性基金项目(2019062)
西安市科技局科技计划基金项目(201805030YD8CG14(17))。
关键词
人体姿态估计
卷积神经网络
高分辨率表征
尺度变换结构
空间变换
human pose estimation
convolutional neural network
high-resolution representation
scaling transformation structure
space transformation
