欢迎访问《中国粮油学报》编辑部网站！

李智,张艳飞,杨卫东,但乃禹,张蕙,陈卫东,荆世华,邵辉,任飞燕.基于语义分割的粮仓粮食数量变化动态监测方法[J].中国粮油学报,2024,39(4):131-139

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

投稿时间：2023-04-20 修订日期：2023-08-25

DOI：

英文关键词:deep learning DeepLabV3+ grain surface recognition semantic segmentation

基金项目: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)

作者	单位	邮编
李智^*	河南省粮食光电探测与控制重点实验室,河南工业大学信息科学与工程学院	450001
张艳飞	河南省粮食光电探测与控制重点实验室,河南工业大学信息科学与工程学院
杨卫东	河南省粮食光电探测与控制重点实验室,河南工业大学人工智能与大数据学院
但乃禹	河南省粮食光电探测与控制重点实验室,河南工业大学信息科学与工程学院
张蕙	河南工业大学信息科学与工程学院
陈卫东	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
荆世华	浪潮数字粮储科技有限公司
邵辉	浪潮数字粮储科技有限公司
任飞燕	河南省粮食光电探测与控制重点实验室,河南工业大学信息科学与工程学院

摘要点击次数: 56

全文下载次数: 0

中文摘要:

为了高效、准确且低成本的监测粮仓粮食数量变化情况,本文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,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的线索并预警,提高粮库清仓检查的针对性和效率。本文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5.818M,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%,验证了本方法的有效性,为粮仓粮食数量的动态监测提供了新的思路。

英文摘要:

In order to efficiently, accurately, and cost-effectively monitor changes in the quantity of grain in grain warehouses. This study presents a dynamic monitoring model for grain storage based on semantic segmentation. The approach utilizes deep learning techniques to analyze and process images collected by cameras inside the granary. This enables the dynamic monitoring of the changes in grain quantities inside the granary. Finally, the monitoring results are 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 enhances the targeting and efficiency of grain inventory inspections. In this study, granary images were obtained by utilizing surveillance cameras installed in the grain storage facility. The images were carefully screened to construct a high-quality granary image dataset. Subsequently, a dynamic monitoring model for grain storage based on DeepLabV3+ was developed, reference boundaries are extracted for determining grain surface changes, and the changes in pixel values of the reference boundaries are used to determine the changes in grain surface. The Xception backbone network in the original DeepLabV3+ model was replaced with the more computationally efficient MobileNetV2 to improved the model"s accuracy and computational efficiency. The experimental results show that this model has 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 agreement between the change of grain quantity in the silo obtained by the model identification and the actual situation is 96%, which demonstrates the effectiveness of this method and provides a new idea for the dynamic monitoring of grain quantity in grain silos.

关闭