基于动态集成的黄瓜叶部病害识别方法

Recognition Method of Cucumber Leaf Diseases with Dynamic Ensemble Learning

对作物病害类型的准确识别是病害防治的前提。为提高病害识别的准确度,以黄瓜叶部病害识别为例,提出一种基于动态集成的作物叶部病害种类的识别方法。首先利用图像分块策略提取病害图像的75维颜色统计特征,然后采用不一致度量方法对构建的10个BP神经网络单分类器进行差异性度量,并按照差异性大小进行排序,最后根据分类器的可信度,动态选择差异性大的分类器子集对病害图像进行集成识别。在由512幅白粉病、霜霉病、灰霉病和正常叶片4类黄瓜叶片组织图像构成的测试集上,所提方法的识别错误率为3.32%,分别比BP神经网络、SVM、Bagging、Ada Boost算法降低了1.37个百分点、1.56个百分点、1.76个百分点、0.78个百分点。试验结果表明:所提方法能够实现黄瓜叶部病害种类的准确识别,可为其它作物病害的识别提供借鉴。

Crop disease is one of the most important influencing factors for agricultural high yield and high quality. Accurate classification of diseases is a key and basic step for early disease monitoring, diagnostics and prevention. The optimal individual classifier design is currently the common limitation in most crop disease recognition methods based images. To improve the accuracy and stability of disease identification, a disease recognition method of cucumber leaf images via dynamic ensemble learning was proposed. The approach consisted of three major stages. Firstly, totally 75-dimension color features of leaf image were extracted with image block processing. Secondly, a disagreement approach was used to measure the diversity among 10 classifiers of neural networks with an ensemble technique, where the classifiers were ordered according to the diversity. Finally, with the confidence of classifiers, a classifier subset was dynamically selected and integrated to identify the images of crop leaf diseases. To verify the effectiveness of the proposed method, classification experiments were performed on images of four kinds of cucumber leaf tissues, including 512 samples composed of powdery milder, downy mildew, gray mold and normal leaf. The experimental results showed that the recognition error rate of the proposed method was 3.32% , compared with those of BP neural network, SVM, Bagging and AdaBoost methods, it was reduced by 1.37 percentage point, 1.56 percentage point, 1.76 percentage point and 0.78 percentage point, respectively. The proposed method identified the diseases accurately from cucumber leaf images. Moreover, the method was feasible and effective, and it can also be utilized and modified for the classification of other crop diseases.