基于卷积神经网络与迁移学习的稻田苗期杂草识别

Recognition of Weeds at Seedling Stage in Paddy Fields Using Convolutional Neural Network and Transfer Learning

杂草类别信息获取是实现杂草智能化田间管理的基础,为实现自然光照和大田复杂背景下的稻田苗期杂草自动识别,提出了一种基于卷积神经网络(Convolutional Neural Network,CNN)与迁移学习相结合的稻田苗期杂草识别方法,可将预训练CNN图像分类模型的参数迁移到稻田苗期杂草识别任务。工作时,采集6种稻田苗期杂草图像共928幅,包括鳢肠、丁香蓼、空心莲子草、千金子、野慈姑和稗草,随机选择70%的图像样本作为训练集,剩余30%的图像样本作为测试集。基于3种经典CNN图像分类模型AlexNet、VGG16和GoogLeNet进行参数迁移学习,这3种参数迁移模型对6种稻田苗期杂草测试样本的正确识别率分别为96.40%、97.48%和91.01%。试验结果表明:基于预训练CNN图像分类模型与迁移学习相结合的稻田苗期杂草识别方法切实可行,可为其他农业应用中小样本图像的识别提供参考。

Weed category information is the basis of the field intelligence management.In order to realize the automatic recognition of weed in the paddy field,this paper presents a new recognition method for weeds at seedling stage in paddy fields using convolutional neural network(CNN)and transfer learning.In this proposed method,parameter-transfer with the well trained CNN model in image classification can move to identify model with weeds.The paper selected 6 weed species in paddy fields,including Eclipta prostrata,Ludwigia adscendens,Alternanthera philoxeroides,Leptochloa chinensis,Sagittaria trifolia,and Echinochloa crus-galli,which were captured in early growth stages with natural background and variable illumination.A total of 928 images were taken.Randomly selected sample of 70%were used as training data,and the samples of rest were used as testing data.On the one hand,three types of weed classification models were established based on pre-trained CNN model in image classification and transfer learning of parameter,which were AlexNet,VGG16 and GoogLeNet.The result indicates that the recogniton accuracy of three CNN model respectively reached 96.40%,97.48%and 91.01%with testing data of weeds images.The results demonstrate that the method of weed recognition using pre-trained CNN and transfer learning can improve the classification accuracy of weeds with the complex background and variable illumination in paddy fields.It shows that the method is practicable which can provide reference to other image recognition with the small sample data in agriculture.