基于朴素贝叶斯分类器的棉花盲椿象危害等级识别

针对自然条件下棉花盲椿象危害区域提取和危害等级识别难的问题,提出了棉花盲椿象危害等级自动识别方法。该方法以自然条件下采集的不同盲椿象危害等级棉叶图像为对象,利用最大类间方差阈值分割和多颜色分量组合方法进行作物与土壤分离和病斑分割,并利用分水岭分割方法对粘连棉叶进行分离并提取盲椿象危害棉叶区域,提取图像的颜色、纹理和形状特征,结合朴素贝叶斯分类器,依据划分的棉花盲椿象危害等级标准,对盲椿象危害等级进行识别。不同盲椿象危害等级识别试验结果表明:该模型平均识别正确率达90.0%,通过比较试验表明,该模型在识别精度比BP神经网络高2.5%,运行速度比支持向量机高11.7%,可较好的对棉花盲椿象危害等级进行识别,可为棉花盲椿象的防治提供技术支持。

Recognition of hazard grade for cotton blind stinkbug based on Naive Bayesian classifier

Abstract: Cotton, one of the most important economic crops in our country, always suffers a variety of pest during the whole process of planting. Blind stinkbug, which seriously affected the cotton quality and yield during BT cotton, is planted in large areas of the Yellow River and Xinjiang province in China. Traditional cotton blind stinkbug hazard ration identification method relies too much on experience, but recognition accuracy and recognition speed are low. In view of complex background of cotton blind stinkbug hazard region and the difficulty in segmentation and classification under natural conditions, an automatic classification method of Cotton blind stinkbug hazard level was proposed. On the basis of the classification standard of plant diseases and insect pests and hazard characteristics of cotton blind stinkbug, as well as the harm degree distribution of bug to cotton by artificial statistics, the cotton blind stinkbug damage grade was divided and the damage grade standard of bug to cotton was put forward. The processing steps of the cotton leaf image in different bug damage grade acquainted in natural conditions were as follows. Firstly, by using Q color component and Otsu segmentation method, the image background was divided. Secondly, in order to remove burrs after splitting, morphological opening operation and internal filling algorithm were used to deal with the segmented image, and the largest connected component was extracted, which can eliminate the influence of weeds. Thirdly, the disease regions of cotton were extract by H+a*+b* component and Otsu segmentation method based on blind stinkbug hazard cotton leaves. The adhesion cotton leaves were separated by Watershed segmentation method Forth, and extracted and selected contents including the color, texture and shape features of and cotton leaf hazard by blind stinkbug. In accordance with the principle of distinction and difference, color feature, texture feature and shape feature was the input indicators classifier. Based on the statistical results of color, texture and shape feature of bug damage image to cotton, R component, G component, B component, I1 component, S component and V component were selected as the color feature, Contrast and Correlation were selected as the texture feature, Pa value were selected as the shape feature. Finally, based on Matlab R2008 platform, combined with the bug feature variables and naive Bias classifier extraction, this method had the aim to distinguish the cotton blind stinkbug damage grade based on the cotton bug division of the harm grade standard. In this experiment, 120 cotton blind stinkbug damage leaves image with 6 levels were used for simulation, in which 60 images were the training set and the others were the validation set. Different bug harm level recognition experiment results showed that, the model has advantages in accuracy and speed with average rate of correct recognition as 90% and average operation time as 0.278 seconds, which was better than Support vector machine and BP neural network model. The proposed cotton blind stink bug hazard grade standard can provide a theoretical basis for the study of harmful cotton blind stinkbug. The proposed classification method of cotton blind stinkbug hazard rating will not only promote technical level for the prevention and treatment of the cotton blind stinkbug, but also it provides a reference for the identification and control of other pests.