基于Faster R-CNN和数据增强的棉田苗期杂草识别方法

Cotton Field Seedling Weed Identification Method Based on Faster R-CNN and Data Enhancement

为解决棉花幼苗与多种类杂草交叉生长的识别率低、鲁棒性差等问题,以棉花幼苗和田间的七类常见杂草为研究对象,提出了一种基于Faster R-CNN和数据增强的棉田苗期杂草识别方法.采集不同背景下受光照影响的杂草图像4694张,包括晴天、阴天和雨天.通过对样本图像的数据增强和特征提取网络ResNet-101的参数优化,训练出了一种可识别棉花幼苗与多种类杂草交叉生长的Faster R-CNN网络模型.在相同样本和特征网络下,将该模型与YOLO模型进行对比.结果表明:Faster R-CNN模型在棉田苗期的多种不同杂草识别中具有明显的优势,可实现各种交叉生长的杂草目标识别,平均识别率为92.01%,平均识别时间为0.261 s.

In order to solve the problems of low recognition rate and poor robustness of cotton seedlings and various types of weeds,taking cotton seedlings and the seven common weeds in the field as the research object,a method based on Faster R-CNN and data enhancement is proposed.Method of identifying weeds in cotton field at seedling stage.4694 images of weeds affected by light under different backgrounds were collected,including sunny,cloudy and rainy days.Through the data enhancement of the sample image and the parameter optimization of the feature extraction network ResNet-101,a Faster R-CNN network model that can identify the cross-growth of cotton seedlings and various types of weeds is trained.Under the same sample and feature network,the model is compared with the YOLO model.The results show that the Faster R-CNN model has obvious advantages in the identification of many different weeds in the cotton field seedling stage,and can realize the identification of various cross-growing weeds,reaching an average recognition rate of 92.01%and an average recognition time of 0.261 s.