河南农业科学 ›› 2024, Vol. 53 ›› Issue (11): 156-163.DOI: 10.15933/j.cnki.1004-3268.2024.11.017

• 农业信息与工程·农产品加工 • 上一篇    下一篇

基于U-Net 的粮仓储粮高度定量检测方法

任飞燕1,2,张蕙1,2,李智1,2,杨卫东1,张艳飞1,2,陈卫东3,谈云建4,柳瑞芸4   

  1. (1.河南省粮食光电探测与控制重点实验室,河南 郑州 450001;2.河南工业大学信息科学与工程学院,河南 郑州 450001;3.河南工业大学粮食和物资储备学院/粮食储运国家工程研究中心,河南 郑州 450001;4.华信咨询设计研究院有限公司,浙江 杭州 310051)
  • 收稿日期:2024-06-20 出版日期:2024-11-15 发布日期:2024-12-19
  • 通讯作者: 李智(1985-),男,河南郑州人,副教授,博士,主要从事计算机视觉研究。E-mail:lizhi@haut.edu.cn
  • 作者简介:任飞燕(2000-),女,河南洛阳人,在读硕士研究生,研究方向:计算机视觉。E-mail:renfeiyan2023@163.com
  • 基金资助:
    河南省杰出青年基金项目(222300420004)

Quantitative Detection Method of Grain Storage Height in Grain Silos Based on U‑Net

REN Feiyan1,2,ZHANG Hui1,2,LI Zhi1,2,YANG Weidong1,ZHANG Yanfei1,2,CHEN Weidong3,TAN Yunjian4,LIU Ruiyun4   

  1. (1.Henan Province Key Laboratory of Grain Photoelectric Detection and Control,Zhengzhou 450001,China;2.College of Information Science and Engineering,Henan University of Technology,Zhengzhou 450001,China;3.School of Food and Strategic Reserves,Henan University of Technology,National Engineering Research Center of Grain Storage and Logistics,Zhengzhou 450001,China;4.Huaxin Consulting Design and Research Institute Co.,Ltd.,Hangzhou 310051,China)
  • Received:2024-06-20 Published:2024-11-15 Online:2024-12-19

摘要: 我国储粮分布具有点多、线长、面广的特点,传统的粮食数量监管面临着效率低、成本高、严重滞后等问题,亟需研发可对粮仓储粮数量实时快速检测的新技术。针对该问题,提出一种基于U-Net的粮仓储粮高度定量检测方法,通过对仓内监控相机拍摄的图片中粮面和通风窗进行分割处理,得到粮面边缘和通风窗上下边缘像素值,并以通风窗上下边缘的离地高度为基础,计算得到粮仓储粮的实际高度,再结合粮仓长度、宽度和粮食密度等基础数据,即可得到粮仓储粮实际数量。分别使用U-Net、DeepLabV3+和PSPNet 3种算法对分割后的粮仓图片进行分析处理并计算储粮高度。结果表明,U-Net的平均交并比(MIoU)和平均像素准确度(MPA)分别为93.25%和95.88%,MIoU相较于DeepLabV3+和PSPNet 分别提高1.82、2.69 百分点,MPA 相较于PSPNet 提高2.42 百分点;U-Net 的定量分析误差为3.51%,比DeepLabV3+和PSPNet分别低1.34、0.43百分点,较为适合作为粮仓储粮高度定量计算分割算法。该方法无需提前设置测量标尺,仅依靠仓内监控相机即可实现粮堆高度测量,可基本满足粮仓储粮数量的快速检测需求。

关键词: 粮仓储粮数量, 储粮高度, 语义分割, U-Net, 定量分析, 计算机视觉

Abstract: Due to the distribution of grain storage in China is characterized by many points,long lines and wide areas,the traditional method of monitoring grain quantity is struggling with low efficiency,high costs,and significant delays.There is an urgent need to research and develop a new technology that can perform real‑time and rapid detection of the quantity of grain stored in grain silos.To address the problem,we proposed a quantitative detection method of grain storage height in grain silos based on U‑Net.By segmenting the grain surface and ventilation windows in pictures captured by monitoring cameras in the grain warehouse,we obtained pixel values of the edge of the grain surface and the upper and lower edges of the ventilation window.Taking into account the height of the ventilation window’s upper and lower edges above and below the ground,we could determine the actual height of grain storage in the grain silo. Subsequently,merged with the length,width,grain density,and other fundamental data,we obtained the actual quantity of grain stored in the grain silo.In this study,the height of grain storage was calculated by analyzing and processing the segmented soil images using U‑Net,DeepLabV3+,and PSPNet algorithms respectively.The experimental results showed that the mean intersection over union(MIoU)reached 93.25%,which was 1.82 and 2.69 percentage points higher than that of DeepLabV3+and PSPNet,respectively.The mean pixel accuracy(MPA)reached 95.88%,which was 2.42 percentage points higher compared with PSPNet.The quantitative analysis error of U‑Net was 3.51%,which was 1.34 and 0.43 percentage points lower compared with DeepLabV3+ and PSPNet,respectively.U‑Net was more suitable as a segmentation algorithm for quantitative calculation of grain storage height in granaries.It is not necessary to set up the measuring scale beforehand and with this method.By relying solely on the monitoring camera inside the warehouse,the height measurement of grain pile can be achieved,which can effectively meet the fast detection requirements for the quantity of grain stored in the warehouse.

Key words: Quantity of grain stored in granaries, Grain storage height, Semantic segmentation, U?Net, Quantitative analysis, Computer vision

中图分类号: