首页|基于特征增强的高分辨率人体姿态估计网络

基于特征增强的高分辨率人体姿态估计网络

扫码查看
原文链接
  • 万方数据
  • 维普
在轻量级卷积神经网络进行高分辨率人体姿态估计时存在提取特征不充分,针对该问题,提出了一种基于特征增强的高分辨率人体姿态估计网络.首先利用空洞卷积补全操作提取图像特征,以避免特征信息丢失且保持模型参数基本不变;接着利用池化增强模块进行卷积提取特征的选择,以保留重要特征且减轻传统池化模块对提取特征造成的破坏;最后利用加强通道信息交互的深度可分离卷积模块进行特征提取,以保持该模块的参数量较少且能够提高其特征提取能力.在COCO2017数据集进行性能测试,本文算法和DiteHRNet30算法的AR值分别为77.9%和77.2%;在 MPII数据集进行性能测试,本文算法和DiteHRNet30算法的PCKh值分别为32.6%和31.7%.实验结果表明,本文算法在人体姿态估计精度和算法复杂度之间能够达到较好的平衡.
High-resolution human pose estimation network based on feature enhancement
In order to solve the problem of insufficient extracted features in high-resolution human pose estimation using lightweight convolutional neural network,a high-resolution human pose estimation network based on feature enhancement is proposed in this paper.Firstly,the dilated convolution completion operation was used to extract image features to avoid the loss of feature information and basically keep the model parameters unchanged.Then,the pooling enhancement module was used to select the features of convolution extraction,which retained important features and reduced the damage caused by traditional pooling module on extracted features.Finally,the depthwise separable convolution module that strengthens the channel information interaction was used for feature extraction,so as to keep the number of parameters of the module small and improve its feature extraction ability.The performance of the proposed algorithm and DiteHRNet-30 algorithm were tested on the COCO2017 dataset.The AR values of the proposed algorithm and DiteHRNet-30 algorithm are 77.9%and 77.2%,respectively.The performance of the proposed algorithm and DiteHRNet-30 algorithm are tested on the MPII dataset.The PCKh values of the proposed algorithm and DiteHRNet-30 algorithm are 32.6%and 31.7%,respectively.Experimental results show that the proposed algorithm can achieve a good balance between the accuracy of human pose estimation and the complexity of the algorithm.

human pose estimationlightweight networkhigh resolutiondilated convolutionpoolingdepth separable convolution

谢唯嘉、易见兵、曹锋、李俊

展开 >

江西理工大学信息工程学院 赣州 341000

人体姿态估计 轻量级网络 高分辨率 空洞卷积 池化 深度可分离卷积

国家自然科学基金江西省自然科学基金江西省教育厅科技项目江西省教育厅科技项目江西省教育厅科技项目江西省赣州市科技计划江西省研究生创新专项

6206601820181BAB202004GJJ210828GJJ200818GJJ180482YC2022-S640

2024

10.19651/j.cnki.emt.2314508

电子测量技术
北京无线电技术研究所

电子测量技术

CSTPCD北大核心
影响因子：1.166
ISSN：1002-7300
年,卷(期)：2024.47(2)
  • 36