Abstract: As affected by the complex background of images from the complex environment, it is difficult to recognize disease from images of tomato leaves. This research intended to classify the most common tomato diseases of powdery mildew, early blight, and Liriomyza sativae using leaf images captured by mobile phonesina greenhouse environment. A new method of CCLSVM was designed by fusion color and texture feature and SVM method. A sliding window was applied to cut an original image to small region images to support small region image data set and high recognize performance under complex environments. Small region images of tomato leaves were selected as training and validation samples. After extract color and texture features from subimages, the SVM method was used to classify early blight, powdery mildew, Liriomyza sativae, and healthy leaf four classes. Experiment results showed that recognition accuracy was significantly improved from 60.63% to 97.5% after introducingcolor and texture features. Compared with deep learning methods, the recognition accuracy of CCLSVM wasalso higher than VGG16 and Alexnet, and the test time of a single frame was much lower than that of a deep learning network. It was considered that the designed CCLSVM method could greatly reduce the effect from the background environment, and also with low computation cost and system requirement.

Key words: tomato, complex environment, sliding window, disease recognition, CCLSVM

摘要： 针对复杂环境下番茄叶部图像因其背景复杂导致病害识别较为困难,以温室大棚内采集的番茄叶部图像作为研究对象,对番茄白粉病、早疫病和斑潜蝇三种常见病虫害,提出一种结合颜色纹理特征的基于支持向量机(SVM)的CCL-SVM的复杂环境番茄叶部图像病害识别方法。CCL-SVM方法为实现小样本及复杂背景环境下的快速识别,首先采用滑动窗口将原始番茄叶部病害图像切割成小区域图像,选取不包含背景的小区域图像样本作为试验样本,从而实现样本数量和样本多样性的增加,并降低样本复杂背景的影响。通过对样本数据抽取颜色纹理特征(CCL),采用SVM模型对番茄早疫病、白粉病、斑潜蝇和健康叶片分类识别。试验结果表明,提出的CCL-SVM方法比Gray-SVM对番茄叶片病害种类的识别性能得到大幅提升,识别率从60.63%提升到97.5%;CCL-SVM方法识别精度高于对比的深度学习网络VGG16和Alexnet方法,且每个小区域图像的平均测试时间远低于深度学习网络。本文设计的CCL-SVM方法具有减小复杂背景影响,计算量小及系统要求低的优点,为复杂环境下番茄病害快速识别提供一种新的思路。

关键词: 番茄, 复杂环境, 滑动窗口, 病害识别, CCL-SVM