基于低空无人机影像和改进Faster R-CNN的棉田杂草识别方法

Weed detection method in cotton field based on low altitude UAV image and improved Faster R-CNN

为实现杂草精准防控、快速识别棉田杂草,本文基于低空无人机技术拍摄棉田苗期影像,以幼苗和杂草为研究对象,提出了一种改进Faster R-CNN的棉田杂草识别方法。改进1:特征提取网络采用融合FPN的ResNet50代替VGG16来训练模型,并对比VGG16、ResNet50和MobileNetv2的分类效果;改进2:针对Faster R-CNN模型设计适合小目标的锚尺度,其中对应的anchor尺度为{8×8,16×16,32×32},比例为{1∶2,1∶1,2∶1};改进3:训练过程在通过特征提取阶段后,采用双线性插值操作,避免两次量化对物体识别产生的影响;改进4:添加2个Dropout层,分别在特征提取网络之后的2个全连接层上,避免模型产生过拟合现象,增加了网络的鲁棒性。研究表明:融合FPN的ResNet50训练的的平均精确率比VGG16提高了3.82%,与ResNet50和MobileNetv2相比分别高出5.05%和18.38%,说明Faster R-CNN使用融合FPN的ResNet50具有最佳的性能,改进的Faster R-CNN模型对单张图像平均识别时间为0.289 s,平均精确率为89.19%;与YOLOv5s检测算法相比,改进的方法平均精确率提高了4.93%,表明本文方法在平均检测精度和平均识别时间占据一定的优势,具有更强的拟合性能。本文方法为精准除草提供一定的可行性和推广性,为实现田间杂草防控奠定理论基础。

In order to achieve precise weed control and fast weed identification in cotton fields,this paper proposed an improved Faster R-CNN weed identification method based on low-altitude UAV technology to take seedlings and weeds as research objects.Improvement1:The VGG16 network was replaced by ResNet50 with FPN as the feature extraction network,and the classification effects of VGG16,ResNet50 and MobileNetv2 were compared.Improvement2:The anchor scale suitable for small targets was designed for the Faster R-CNN model,where the corresponding anchor scale were{8×8,16×16,32×32}and the ratio were{1∶2,1∶1,2∶1}.Improvement3:In the training process,bilinear interpolation operation was adopted after feature extraction stage to avoid the influence of double quantization on object recognition.Improvement 4:Two Dropout layers were added to the two fully connected layers behind the feature extraction network to avoid over-fitting of the model and increase the robustness of the network.Research shows that:The average accuracy of ResNet50 trained by fusing FPN was 3.82%higher than VGG16,5.05%higher than ResNet50 and 18.38%higher than MobileNetv2,respectively,indicating that Faster R-CNN used ResNet50 of fusing FPN had the best performance.The improved Faster R-CNN model had an average recognition time of 0.289 s and an average accuracy of 89.19%for a single image.Compared with the YOLOv5 s model,the average accuracy of the improved method was improved by 4.93%,indicating that the proposed method had certain advantages in the average detection accuracy and average recognition time,and had stronger fitting performance.The method in this paper provided certain feasibility and popularization for accurate weed control,and laid a theoretical foundation for realizing weed control in field.