基于改进U-Net的料塔料位检测方法研究  

Research on material level detecting method of material silos based on improved U-Net

作　　者：解浩亮 刘仁鑫[1,2] 周波 容能威 XIE Haoliang;LIU Renxin;ZHOU Bo;RONG Nengwei(School of Engineering,Jiangxi Agricultural University,Nanchang 330045,China;Research Center of Animal Husbandry Facility Technology Exploitation,Nanchang 330045,China)

机构地区：[1]江西农业大学工学院,南昌330045 [2]畜牧设施技术开发研究中心,南昌330045

出　　处：《黑龙江畜牧兽医》2022年第11期54-59,137,138,共8页Heilongjiang Animal Science And veterinary Medicine

基　　金：江西省现代生猪产业技术体系专项(赣财文指[2019]7号)。

摘　　要：在规模化养殖中,传统料塔的料位检测设备较少,检测方法主要是人工检测和经验判断,难以精确、实时检测料塔余料,无法满足现代化农业发展的需求。为了实现精细化管理,摒弃传统的检测方法,采用基于U-Net的语义分割模型对料塔内饲料进行分割,并通过处理图像输出结果,检测料塔内余料容量,同时引入残差结构和空间注意力机制来提升分割效果,优化U-Net语义分割模型。经过对比试验结果显示,F1分数由0.942提升到了0.951,料位检测结果正确率由0.920提升到了0.960。说明基于改进U-Net的料位检测方法具有一定的可行性,能满足检测的基本要求。In large-scale breeding, the traditional silos have fewer material level detection equipment, and the detection methods are mainly manual inspection and empirical judgment. It is difficult to accurately and real-timely detect the remaining feed of the silo and cannot meet the needs of modern agricultural development. In order to achieve refined management and abandon the traditional detection methods, the network model based on U-Net was used to segment the feed in the silo, and the result were output as image processing to detect the remaining feed capacity in the silo. In order to improve the segmentation effect, the residual structure and spatial attention mechanism were introduced, and the U-Net semantic segmentation model was optimized. After comparison experiment, the result showed the F1 score was improved from 0.942 to 0.951, and the accuracy of feed level detection was improved from 0.920 to 0.960.The results indicated that the material level detection method based on improved U-Net had a certain feasibility and could meet the basic requirements of detection.

关 键 词：料塔 料位检测 现代农业 U-Net网络模型 图像处理 图像输出 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]