Research On Greenhouse Cucumber Disease Identification System Based On Image Processing Technology

With the development of computer technology and image processing technology,automatic identification of disease has become a research focus. Science and technologyapplied to agricultural production has not only freed the labor force, to improve crop yieldsand the quality of agricultural products, but also protect the environment, so it is of greatsignificance to the development of modern agriculture.In this thesis, the downy mildew, powdery mildew and anthracnose, the commondiseases of greenhouse cucumber, are taken as the main research object. A recognitionsystem of greenhouse cucumber disease is designed, and realized in the Visual C++environment. The system mainly includes pretreatment, disease spot segmentation,(color,shape, texture) feature extraction, disease recognition four modules, and the algorithm ineach module can be changed or revised.The gray level transformation, smooth and sharpening processing are used in theimage pre-processing part. The weighted average method is used in the gray leveltransformation. After the mean filtering method and the median filtering method have beencompared during the image smooth processing, we concluded that the median filter ismuch more suitable for the texture feature extraction while the mean filter is suitable forthe color feature extraction. The gradient sharpening and Laplace’s sharpening areanalyzed, and we found that the effect of the gradient is better and more suitable to be usedin this thesis.In disease spot segmentation module, the iteration threshold segmentation method andthe OTSU method were compared. The effects of disease spot extraction are unsatisfactory.Based on the iteration threshold segmentation method, an improved segmentationalgorithm is presented. The advantage of the improved algorithm is that it can give a moreaccurate positioning of the lesion.In feature extraction module, nine color features, four shape features, eight texture features are extracted, and the data of all features are normalized.In disease recognition module,21characteristics are optimized and screened, the trialand error method is used for assigning the weights. According to the importance of thecharacteristics above, the characteristics were assigned to different weights. Experimentalresults show that identify diseases accuracy can reach more than88%when we only usethe mean of blue pixel values, and when it combined with color, texture, shapecharacteristics, the identification diseases accuracy can reach more than96%. So these twomethods are effective ways.