引用本文
  •    [点击复制]
  •    [点击复制]
【打印本页】 【下载PDF全文】 查看/发表评论下载PDF阅读器关闭

←前一篇|后一篇→

过刊浏览    高级检索

本文已被:浏览 236次   下载 248 本文二维码信息
码上扫一扫!
基于RGB-D图像的自动化膳食调查系统
李仕超1,高梓成1,郭浩2*,邓志扬3,马瑞1,雷杰1,张昊3
0
(1.中国农业大学 土地科学与技术学院, 北京 100083;2.中国农业大学 信息与电气工程学院, 北京 100083;3.中国农业大学 食品科学与营养工程学院, 北京 100083)
摘要:
针对传统膳食调查方法使用过程繁琐,耗时长的问题,对基于彩色深度图像的自动化膳食调查系统进行研究,采用DeepLabv3+语义分割模型对RealSense传感器获取的食物彩色图像进行分割,根据食物分割结果和对应的深度图像计算每一类别食物体积并结合食物配料表和营养成分估算营养素含量;同时采集人体深度图像序列结合蒙皮多人线性模型拟合三维人体模型,并利用模型点云和人体体重信息计算身体质量指数和腰臀比。结果表明:1)本系统在食物图像分割中的像素准确率为72.1%,像素准确率平均值为97.13%,平均交并比为82.03%;2)食物体积计算中的平均绝对误差均小于40 cm3;3)所有人体样本计算的腰臀比与真实值偏离程度均小于4%。本系统在实现自动监测食物各营养素摄入量的基础上,增加对人体体重状况以及中心性肥胖程度的初步判定,同时提高了膳食调查的便捷性和自动化程度,可以为营养干预提供依据和参考。
关键词:  膳食调查  RGB-D图像  深度学习  食物语义分割  食物体积估计
DOI:10.11841/j.issn.1007-4333.2023.02.17
投稿时间:2022-05-13
基金项目:国家自然科学基金项目(42071449)
A dietary survey system based on RGB-D images
LI Shichao1,GAO Zicheng1,GUO Hao2*,DENG Zhiyang3,MA Rui1,LEI Jie1,ZHANG Hao3
(1.College of Land Science and Technology, China Agricultural University, Beijing 100083, China;2.College of Information and Electrical Engineeing, China Agricultural University, Beijing 100083, China;3.College of Food Science and Nutrition Engineering, China Agricultural University, Beijing 100083, China)
Abstract:
To address the problems of tedious and time-consuming process of the traditional dietary survey methods, this study proposed to investigate an automated dietary survey system based on color depth images. DeepLabv3+ semantic segmentation model was adopted to segment the color images of food acquired by RealSense sensors by calculating the volume of each category of food based on the food segmentation results with corresponding depth images and combining the food ingredient list and nutrient composition to estimate nutrient content. Meanwhile, the human depth image sequences were collected to fit the three-dimensional human model with the skinned multi-person linear model, while the body mass index and waist-hip ratio were calculated by using the model point cloud and human weight information. The results show that: 1)The pixel accuracy of this system is 72. 1% in food image segmentation, the mean pixel accuracy is 97. 13%, and the mean intersection over union is 82. 03%; 2)The mean absolute errors are all less than 40 cm3 in food volume calculation; 3)The waist-hip ratio calculated for all human samples deviates from the true value by less than 4%. This system provides references for the nutritional interventions by increasing the preliminary determination of human body weight status and the degree of central obesity based on the realization of automatic monitoring of the intake of each nutrient in food, as well as improving the convenience and automation of dietary surveys.
Key words:  dietary survey  RGB-D images  deep learning  food segmentation  food volume estimation