基于语义分割的粮仓粮食数量变化动态监测方法  

Dynamic Monitoring Methodology of Grain Quantity Variation in Granaries Based on Semantic Segmentation

作　　者：李智 张艳飞[1,2] 杨卫东 但乃禹 张蕙[2] 陈卫东 荆世华 邵辉[6] 任飞燕 Li Zhi;Zhang Yanfei;Yang Weidong;Dan Naiyu;Zhang Hui;Chen Weidong;Jing Shihua;Shao Hui;Ren Feiyan(Henan Province Key Laboratory of Grain Photoelectric Detection and Control,Zhengzhou 450001;School of Information Science and Engineering,Henan University of Technology,Zhengzhou 450001;School of Artificial Intelligence and Big Data,Henan University of Technology,Zhengzhou 450001;School of Food and Materials Reserve,Henan University of Technology,Zhengzhou 450001;National Engineering Research Center for Grain Storage and Logistics,Zhengzhou 450001;Inspur Digital Grain Storage Technology Co.,Ltd.,Jinan 250098)

机构地区：[1]河南省粮食光电探测与控制重点实验室,郑州450001 [2]河南工业大学信息科学与工程学院,郑州450001 [3]河南工业大学人工智能与大数据学院,郑州450001 [4]河南工业大学粮食和物资储备学院,郑州450001 [5]粮食储运国家工程研究中心,郑州450001 [6]浪潮数字粮储科技有限公司,济南250098

出　　处：《中国粮油学报》2024年第4期131-139,共9页Journal of the Chinese Cereals and Oils Association

基　　金：国家重点研发计划项目(2017YFD0401001-02);河南省杰出青年基金项目(222300420004);河南省重大公益专项(201300210100)。

摘　　要：为了高效、准确且低成本监测粮仓粮食数量变化情况,研究提出一种基于语义分割的粮仓粮食数量变化动态监测方法,利用深度学习技术对粮仓仓内摄像机采集的图像进行分析,实现对粮仓仓内粮食数量变化情况的动态监测。通过将监测结果与仓内近期业务数据进行比对,可及时发现违法违规行为线索并预警,提高日常监管的针对性和效率。本研究选取粮仓仓内监控摄像机采集的图像作为数据集,构建了基于DeepLabV3+的粮仓粮食数量变化动态监测模型,通过提取判断粮面变化的参照边界,利用参照边界像素值的变化判断仓内粮食数量变化情况,并通过引入基于MobileNetV2的特征提取网络,提高了模型识别的准确性和计算效率。实验结果表明,该模型平均交并比和平均像素准确率分别达到89.57%和94.53%,参数量为5.818 M,MIoU分别比PSPNet模型和UNet模型高0.95%和0.88%。通过对50个粮仓的测试分析,模型识别得到的仓内粮食数量变化情况与实际情况的一致性为96%,验证了该方法的有效性,为粮仓粮食数量的动态监测提供了新的思路。In order to efficiently,accurately and cost-effectively monitor changes in the quantity of grain in grain warehouses.In this paper,a dynamic monitoring model was presented for grain storage based on semantic segmentation.In this approach,deep learning techniques were utilized to analyze and process images collected by cameras inside the granary,enabling the dynamic monitoring of the changes in grain quantities inside the granary.Finally,the monitoring results were compared with recent business data to identify any illegal behavior during routine supervision and provide timely feedback to the grain storage regulators.The proposed approach enhanced the targeting and efficiency of grain inventory inspections.In this paper,the images collected by the monitoring camera in the granary were selected as the data set,and a dynamic monitoring model of granary grain quantity change based on DeepPlabv 3+was constructed.By extracting the reference boundary to judge the grain surface change,the change of grain quantity in the granary was judged by the change of the pixel value of the reference boundary,and by introducing the feature extraction network based on MobileNetV2,the accuracy and calculation efficiency of model identification were improved.The experimental results indicated that this model had the mean intersection over union and mean pixel accuracy reaching 89.57%and 94.53%,respectively,with the number of parameters of 5.818 M.This model improved the Mean Intersection over Union by 0.95%and 0.88%compared to PSPNet and UNet models,respectively.Through the testing analysis of 50 grain silos,the consistency between the change of grain quantity in the silo obtained by the model identification and the actual situation was 96%,demonstrating the effectiveness of this method and providing a new idea for the dynamic monitoring of grain quantity in grain silos.

关 键 词：深度学习 DeepLabV3+ 粮面识别 语义分割 

分 类 号：S126[农业科学—农业基础科学] TP391.4[自动化与计算机技术—计算机应用技术]