基于改进YOLOv4的田间密集小目标检测方法  

Method for detection of farmland dense small target based on improved YOLOv4

作　　者：杨军奇 冯全 王书志[2] 张建华[3] 杨森 YANG Junqi;FENG Quan;WANG Shuzhi;ZHANG Jianhua;YANG Sen(School of Mechanical and Electrical Engineering,Gansu Agriculture University,Lanzhou 730070,China;School of Electrical Engineering,Northwest University for Nationalities,Lanzhou 730030,China;Agricultural Information Institute,Chinese Academy of Agricultural Sciences,Beijing 100081,China)

机构地区：[1]甘肃农业大学机电工程学院,兰州730070 [2]西北民族大学电气工程学院,兰州730030 [3]中国农业科学院农业信息研究所,北京100081

出　　处：《东北农业大学学报》2022年第5期69-79,共11页Journal of Northeast Agricultural University

基　　金：国家自然科学基金项目(32160421,31971792);甘肃省教育厅产业支撑项目(2021CYZC-57);甘肃省青年科技基金项目(20JR10RA544)

摘　　要：为研究农业场景下密集小目标难识别问题,提出一种基于卷积注意力的CBAM-YOLOv4密集小目标检测方法。该方法在YOLOv4模型骨干网络Add层和Concat层后嵌入卷积注意力模块(CBAM),在保证模型检测效率基础上提高模型检测精度。通过设置多种试验条件,使用不同密集程度、光照条件及天气状况下采集的密集葡萄叶片数据测试模型检测效果,试验对比EfficientDet、YOLOv3、YOLOv4及CBAM-YOLOv44种网络,采用统计AP值的评价方法评估各模型差异。结果表明,在密集葡萄叶片数据集中,CBAM-YOLOv4模型识别效果提升明显,对于高度密集叶片数据集,该模型AP值为82.04%,相比YOLOv4提高1.94%,高于其他检测模型3%~4%。将该模型应用于无人机果树资源调查中,对果树的计数精度为90.43%。以上试验结果表明,该方法对农业场景下密集小目标检测具有较高精度。In order to study the problem of dense small target recognition in agricultural scene,this paper proposes a CBAM-YOLOv4 dense small target detection method based on convolution attention.In this method,Convolutional block attention module(CBAM)is embedded behind the add layer and concat layer of the original model backbone network to improve the detection accuracy of the model while ensuring the efficiency of model detection.By setting a variety of experimental conditions,the detection effect of the model was tested by collecting the data of dense grape leaves under different densities,different light and different weather.The four algorithms of EfficientDet,YOLOv3,YOLOv4 and CBAM-YOLOv4 were compared,and the statistical AP evaluation method was used to evaluate the differences of each model.For the highly dense leaf datasets,the AP of the dense small target detection model proposed in this paper was 82.04%,which was 1.94%higher than YOLOv4 and 3%-4%higher than other detection models.The model was also applied to the UAV fruit tree resource survey,and the accuracy of tree counting was 90.43%.The above experimental results showed that this method had high accuracy for dense small target detection in agricultural scenes.

关 键 词：农业场景 密集小目标 YOLOv4 卷积注意力 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]