摘 要:为了解决人体轮廓识别精度不高,鲁棒性不好的问题,在保证实时性的要求下,提出了一种基于深度学习的人体轮廓识别方法,该方法采用U-Net 神经网络框架,建立特定视角的数据集,利用高斯滤波进行图像预处理操作,设计使用Dice和交叉熵函数相结合的损失函数进行训练。实验表明,该方法的重合度为91.85%,单次识别耗时为50.56 ms,在保证精度和实时性的前提下,也保证了对不同环境的适应性,在实际应用中有良好的价值。
关键词:人体轮廓;U-Net 神经网络;Dice 损失函数;高斯滤波
中图分类号:TP391.41;TP181 文献标识码:A 文章编号:2096-4706(2020)23-0090-04
Human Contour Recognition Based on Deep Learning
JIA Qunxi,ZHANG Weimin,SUN Zhanpeng,HU Xiaojian
(Luoyang Institute of Science and Technology,Luoyang 471023,China)
Abstract:In order to solve the problem of low accuracy and poor robustness of human contour recognition,a human contour recognition method based on deep learning is proposed under ensuring the real-time requirement. This method uses the U-Net neural network framework to establish a data set of specific perspective,uses Gaussian filtering to execute image preprocessing,and designs a loss function combined with Dice and cross entropy function for training. The experiment shows that the coincidence degree of the method is 91.85%,and the single recognition time is 50.56 ms. Under the premise of ensuring the accuracy and real-time,it also ensures the adaptability to different environments,which has good value in practical application.
Keywords:human contour;U-Net neural network;Dice loss function;Gaussian filtering
基金项目:2020 国家级大学生创新训练计划项目:基于人工智能的智能制衣系统(202011070006);2020 河南省高等学校重点科研项目计划(21B520012)
参考文献:
[1] 吴泽斌,张东亮,李基拓,等. 复杂场景下的人体轮廓提取及尺寸测量 [J]. 图学学报,2020,41(5):740-749.
[2] 路煜. 基于全卷积的编解码网络的人体轮廓提取方法研究 [J]. 信息与电脑(理论版),2020,32(3):47-48+52.
[3] 王鹏,方志军,赵晓丽,等. 基于深度学习的人体图像分割算法 [J]. 武汉大学学报(理学版),2017,63(5):466-470.
[4] 杨丹,刘国如,任梦成,等. 多尺度卷积核U-Net 模型的视网膜血管分割方法 [J]. 东北大学学报(自然科学版),2021,42(1):7-14.
[5] HE K M,GKIOXARI G,DOLLÁR P,et al. Mask R-CNN [C]//2017 IEEE International Conference on Computer Vision (ICCV).New York:IEEE Press,2017:2980-2988.
[6] SHEN X Y,TAO X,GAO H Y,et al. Deep automaticportraitmatting [C]// Computer Vision-ECCV 2016,2016:92-107.
[7] SHEN X Y,HERTZMANN A,JIA J Y,et al. AutomaticPortrait Segmentation for Image Stylization [C]//37th Annual Conferenceof the European Association for Computer Graphics,Goslar:Eurographics Association,2016:93-102.
作者简介:贾群喜(2000—),男,汉族,河南驻马店人,本科在读,研究方向:自动化。