基于最优二叉树支持向量机的蜜柚叶部病害识别

摘要

Honey pomelo, one of the most important fruits in China, always suffers a variety of diseases during the whole process of planting, such as maculopathy, anthracnose, scab and dark mildew, which seriously affects the fruit quality and yield. The accurate recognition of honey pomelo leaf diseases is the premise of the treatment of honey pomelo diseases, and the precision directly affects the efficiency in controlling diseases. However, most of the current researches on disease recognition aimed at the global information of the study objects, but ignored the disease’s local feature extraction in multi-scale and multi-direction; in addition, the present researches generally used the method of“one to one”or“one to many”when building many types of support vector machine (SVM) in the disease classification model, few researches used the method about SVM based on directed acyclic decision tree. So, leaf diseases recognition of honey pomelo based on SVM of directed acyclic decision tree was put forward in this paper. At first, statistical analysis on components of color characteristics of collected honey pomelo leaf diseases was carried on, and the conclusion was drawn according to the statistics of component B, component 2G-R-B, component (G+R+B)/3 and component Q in YIQ color model, which were easily distinguished among the 4 diseases, and so the 4 color components were used as disease color features. Secondly, honey pomelo leaf disease images were converted into 4 grayscale images of component B, component 2G-R-B, component (G+R+B)/3 and component Q in YIQ color model. Gabor wavelet with 5 dimensions and 8 directions was used for convolution calculation with 4 grayscale component images, and 16-dimension energy sub-band was got, the mean value of which was used as eigenvector. Disease recognition model of three-level directed acyclic decision tree SVM was constructed by 6 SVM classifiers, in order to recognize 4 honey pomelo diseases, i.e. maculopathy, anthracnose, scab and dark mildew. According to the test results of cross validation method, the recognition accuracies of maculopathy, anthracnose, scab and dark mildew respectively reached 90%, 96.66%, 93.33% and 96.66%, and the average recognition rate of the 4 diseases was 94.16%, showing that the method could effectively recognize the 4 honey pomelo leaf diseases. Optimal binary tree SVM proposed in this paper was compared with BP neural network, one-to-one SVM and one-to-many SVM in different characteristic dimensions, and the results showed that the training time of the proposed method in this paper and other 3 methods was respectively 740 ms, 420 ms, 450 ms and 370 ms, and the disease recognition accuracy of the 4 methods was respectively 86%, 91.5%, 90% and 94.16%. The method proposed in this paper is superior to the other 3 algorithms in training time and recognition precision. So the proposed method can provide technical support for the accurate recognition of honey pomelo leaf diseases, in favor of the prevention and treatment of pomelo diseases, and also provide references for the prevention and cure of other plant’s leaf diseases.%为了提高蜜柚叶部中晚期病害的识别准确率，确保蜜柚叶部病害对症施药与病害防治的效果，该文提出了一种基于最优二叉树支持向量机（support vector machine，SVM）的蜜柚叶部病害识别方法，该方法首先将蜜柚叶部病害图像转换为B分量、2G-R-B分量、(G+R+B)/3分量以及YIQ颜色模型中的Q分量的4个灰度图像，再利用5尺度8方向的Gabor小波分别与4个分量灰度图像进行卷积运算，获得5个尺度下不同方向的幅值均值作为病害的特征向量，并结合提出的最优二叉树支持向量机病害识别模型，对黄斑病、炭疽病、疮痂病、煤烟病等4种蜜柚叶部病害进行分类识别。通过交叉验证的方法进行分类识别测试，结果表明：黄斑病、炭疽病、疮痂病、煤烟病识别准确率分别为90%、96.66%、93.33%、96.66%，平均识别率达到94.16%，并将该方法与BP神经网络、一对一SVM与一对多SVM进行比较，试验结果表明该方法可有效识别4种蜜柚叶部病害，在训练时间和识别精度上都优于其他3种方法。该方法可为蜜柚病害准确识别与防治提供有效的技术支持。