基于图像分割与卷积神经网络的水稻病害识别  

Recognition of Rice Disease Based on Image Segmentation and Convolution Neural Network

作　　者：万颖[1] 杨红云[2] 王映龙[1] 罗建军 梅梦 WAN Ying;YANG Hongyun;WANG Yinglong;LUO Jianjun;MEI Meng(School of Computer Information and Engineering, Jiangxi Agricultural University, Nanchang 330045,China;School of Software, Jiangxi Agricultural University, Nanchang 330045, China;Basic Teaching Department,Jiangxi Business School,Nanchang 330103,China)

机构地区：[1]江西农业大学计算机信息与工程学院,南昌330045 [2]江西农业大学软件学院,南昌330045 [3]江西省商务学校基础教学部,南昌330103

出　　处：《西北农业学报》2022年第2期246-256,共11页Acta Agriculturae Boreali-occidentalis Sinica

基　　金：国家自然科学基金(61562039);江西省教育厅科技项目(GJJ160374,GJJ170279)。

摘　　要：为了提高水稻病害计算机视觉识别的准确性,研究提出针对水稻白叶枯病、赤枯病、胡麻斑病和纹枯病4种病害进行分类识别的模型。利用计算机视觉和机器学习软件库opencv对病斑图像进行随机旋转、随机翻转、随机亮度变换及随机对比度等处理方式扩充样本,应用区域生长、基于水平集的CV模型、显著性检测3种算法对图像进行分割。通过Tensorflow深度学习平台,构建网络层分别为6层(输入层32×32×3,卷积核大小为5×5)和8层(输入层227×227×3,卷积核大小为11×11、5×5、3×3)的卷积神经网络,将图像分割后得到的3组数据,均以8∶2的比例分别作为卷积神经网络的训练数据和测试数据,训练后得到6个模型,并结合召回率、F1评价指标对模型进行评估。结果表明,6个模型中训练识别准确率最低为97.66%,测试识别准确率最低为95.31%,其中以显著性检测分割算法和8层网络层的卷积神经网络结合得到的模型效果最佳,其训练识别准确率为99.99%,测试识别准确率为99.88%,相较于端到端的卷积神经网络水稻病害识别结果也有所提升。In order to improve the recognition accuracy of rice diseases with computer vision,a model for classification and recognition of four kinds of rice diseases was established.Firstly,software library for opencv for computer vision and machine learning,was used to expand the samples of the images by random rotation,random flipping,random brightness transformation and random contrast.and then three algorithms were applied in image segment,which include regional growth algorithm,CV model based on level sets and saliency detection algorithm.Secondly,two convolutional neural networks with 6 layers(input layer 32×32×3,convolution kernel size of 5×5)and 8 layers(input layer 227×227×3,convolution kernel size of 11×11,5×5,3×3)were respectively constructed through a deep learning platform of Tensorflow,and then,three groups of data were obtained after image segmentation were used as the training data and test data of convolution neural network in the ratio of 8∶2,thereby six models were obtained,and the recall rate and F1 score were employed to evaluate the model,and the experimental results showed that the lowest training recognition accuracy was 97.66%and the lowest test recognition accuracy was 95.31%in the six models.Among them,the model obtained by combining the saliency detection segmentation algorithm and the convolutional neural network with 8-layers had the best performance,the training recognition accuracy was 99.99%and the test recognition accuracy was 99.88%,compared with the end-to-end convolutional neural network.

关 键 词：图像分割 卷积神经网络 水稻病害 识别 显著性检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]