复杂背景下基于改进的Faster R-CNN杂草识别研究
Research on Weed Identification Based on Improved Faster R-CNN in Complex Backgrounds
宋国翠 1晏华成 1龙涛元 1吴立鸿1
作者信息
- 1. 中山火炬职业技术学院,广东中山 528436
- 折叠
摘要
杂草是阻碍农业丰产的重要因素,如何快速准确识别杂草是智能除草的关键.为了提高杂草识别的准确性与实时性,将Faster R-CNN经典的目标检测模型用于杂草识别,选取不同土壤环境、不同杂草密度下的花生幼苗-杂草为试验对象,制作了 PASCAL VOC格式数据集,为了消除复杂背景对植物目标识别的影响,文章提出了采用分割背景用于改进Faster R-CNN中RPN网络的分类层.为了验证该方法的有效性,该研究采用VGG-16为特征提取网络,对比原模型,引入分割背景,提高前景注意力机制后的花生苗-杂草识别模型的mAP提高了0.203.试验结果表明:所提方法对复杂背景下花生苗-杂草识别有较好的检测效果,可为实时精准除草提供参考.
Abstract
Weeds are an important factor that hinders agricultural productivity.How to quickly and accurately identify weeds is the key to weed control.In order to improve the accuracy and real-time performance of weed identi fi cation,the intelligent object detection model of Faster R-CNN is applied to weed identification.We select peanut seedlings and weeds under different soil environments and different weed densities as the test objects,and create a PASCAL VOC format dataset.In order to eliminate the influence of complex backgrounds on target recognition,this paper proposes to use segmentation backgrounds to improve the classification layer of the RPN network in Faster R-CNN.In order to verify the effectiveness of this method,this study uses VGG-16 as the feature extraction network,compared with the original model,introduces seg-mentation backgrounds,and improves the mAP of the peanut seedling-weed recognition model after adopting the foreground attention mechanism.The study shows that the method has good detection performance for peanut seedling-weed recognition in complex backgrounds,and can provide a way for real-time and accurate weed control.
关键词
杂草识别/Faster/R-CNN/RPN网络/背景分割Key words
Weed identification/Faster R-CNN/RPN network/background segmentation引用本文复制引用
出版年
2024
NETL
NSTL
万方数据