Research Of Crop Disease Diagnosis And Leaf Morphological Measurement Parameters Based On Image Processing Technology

Pesticide is one of the largest quantities of the agricultural chemical in the world. Employing pesticide reasonably can prevent and cure plant disease effectively, increase crop yields, improve the income of farmers. However, employing pesticide too much will bring lots of disadvantages on environment and human health. It can pollute soil, water, produce and livestock, especially the man who directly or indirectly touch the pesticide. In order to make good use of the benefit of the pesticide and reduce its harms, the farmers must get the plant growing information exactly and rapidly. Curing the disease depends on the pathogen and the degree of plant disease.This paper adopted advanced image processing technique to diagnose the type of plant disease and measure morphological parameters of plant leaf, which was supported by the project of the national high-tech plan (863) "the research on nutrition information of the crop and shape parameter measuring technology" (2003AA209012).Firstly, the status and the progress about diagnosis on plant disease was presented based on imaging processing and then an image capture set was built including hand-carrier and digital camera. It is light and handy-taken which can reduce the interference between tresses and leaves, filtrate the soil background, make the post-image processing easily.Secondly, median filtering method was adopted to filter the components of RGB respectively, and then the filtered components was composed into RGB image .This method can filter noise well and can not make the image edge blur.Thirdly, Bp neural network and Fuzzy C means were adopted to segment the tomato leaf mold images and maize blotch disease images. In order to get rid of the effect of the light intensity, the RGB color space was transformed to HSI and YCbCr color spaces. H,Cb,Cr components which are not related with light intensity were extracted from HSI color space and YCbCr color space .Thus a kind of multi-color space was formed . H,Cb,Cr as features from every sample image were put into the Bp artificial neural network. The net error was satisfied depending on demand after 172 training times. As a result, segmentation effective percent was above 94.6% after validation. This precision of experiment met the demand and provided a good preparation for diagnosing the plant disease. This research adopted Fuzzy C means to segment maize blotch disease image in H color space successfully. As a result, segmentation effective percent was above 97.8% after validation. It was more exact than Otsu's method.Fourthly, 15 features were picked up including boundary, color and texture as inputs of the plant disease diagnostic model. Thus, eigenvector space of the plant disease which could represent all the information was formed.Fifthly, three methods were adopted including Bp neural network, PNN neural network and LSSVM to build the diagnostic model of maize blotch disease. At first, normalized eigenvector to [-1,1].3 layers united Bp neural network was built to diagnose the disease including 15 features as network inputs ,so the input cell number was 15 .The hidden cell number was 35 by Matlab and the output cell number was 1 depending on practice. The net error was satisfied depending on demand after 696 training times. The structure of PNN neural network was similar to Bp neural network, but its output cell number was 2 and radial distributing density was 0.01. LSSVM Toolbox compiled by Leuven University in 2002 was adopted to builddiagnostic model. Eventually, it got the punishment parameter C=106.7764 and radial radius sig2=l.091312 and made use of last two parameters to train and validate the SVM. The LSSVM was the best method to diagnose the maize blotch disease through error analysis. The diagnostic error was 93.75%~100%.Finally, the program about measuring leaf morphologic parameters was compiled using Matlab. The leaf actual area and perimeter was measured through consulting object .The relative error was 2.5%~9.8% for measuring area and 1.6%~9.2% for measuring perimeter. If the methods presented chapter 4 were combined, 19 parameters could be extract from leaf.This study investigated technologies of plant disease automatic diagnosis using digital processing, neural network, fuzzy theory and support vector machine based on arithmetic theory and Matlab language.