Study On Digital Image Recognition Of Prunus Armeniaca Leaf Diseases And Insect Pests

Current recognition methods for seedling diseases and insect pests employ subjective diagnoses according to characteristic symptom of diseases and insect pests or pathology assay. In this thesis, a digital image processing and pattern recognition method was proposed to diagnosing carmine spider mite (CSM, Tetranychus telarius Linnaeus), apricot bacterial shot hole (ABSH) and OK (health leaf) for apricot (Prunus armeniaca) leaves. The main contains and acquired advances are as follows.(1) Images of diseases & insect pests were acquired by using CCD camera in sunlight. This thesis employed color features, color texture features and gray texture features to evaluate different diseases and insect pests. Firstly, a group of statistical features, based on hue and saturation histogram, were used to describe color feature in HSI color space, including mean, variance, skewness, kurtosis, range, median, mode, quartiles(first quartile, second quartile, third quartile and interquartile rang ). Secondly, a small number of chromaticity moments based on the concept of chromaticity as defined within the CIE XYZ color space were extracted. In addition, a much similar concept, color moments, was suggested within the HSI color space according to the chromaticity moment. Finally, in order to reflect the relationship among pixels, the gray level concurrence matrix (GLCM) was further used to evaluate the gray texture features in the three situations. 5 typical features of GLCM were calculated, including local calm, correlation, energy, entropy and inertia quadrature.(2) Inter class distance method, correlation analysis and Mann-Whitney U were employed to choose and to extract features. Using the three methods and their different groups, many feature parameter groups were constructed. The comparison analysis shows 1) Statistical features of hue histogram perform well for classifying seedling diseases and insect pests as color features, but some features appear highly linear correlation. 2) The distribution of color moments (HSI) and the trace of chromaticity moments (XYZ) in color texture features are good for classifying in this field. However, different moments manifested very high correlation too. 3) GLCM can hardly recognize health leaves and diseases & insect pests'leaves, not only using inter classes distance method but also using Mann-Whitney U. But the three features (local calm, correlation and inertia quadrature) present good quality for classifying CSM and ABSH.(3) Support vector machine (SVM) and K nearest neighbor (K-NN) were employed to classify CSM, ABSH and OK. Two classification applications were performed: classifier A that was used to classify a diseases group and a health group, while classifier B that was used to classify CSM and ABSH. Using all features (unselective), the classification rate of different kernel functions results were compared. The rate of radial basis kernel function was highest to 88.5% and 84.0% for classifier A and B. The results show that radial basis kernel function is better than other kernel functions in this study. For classifier A, classification accuracy ranged between 88.5% and 91.0% using SVM, while it ranged between 80.0% and 85.0% using K-NN. And for classifier B, classification accuracy ranged between 84.0% and 91.0% using SVM, while it ranged between 80.0% and 85.0% using K-NN. The classification accuracy of SVM significantly exceeds K-NN's. In particularly, using the method based on the combination between Mann-Whitney U and correlation analysis, classification accuracy was 89.0% and 91.0% for classifier A and classifier B, and only need 13 and 11 features respectively in this study. Consequently, this new method does reach a higher accuracy. What's more, it can effectively decrease the number of features, keep the separability features and assure nonlinear correlation among features.(4) Based on Visual C++ development platform, some functions were programmed to recognize diseases and insect pests of apricots.The result of this paper attaches great importance to improve the application of information techniques in forestry. It accelerates the application of digital image processing technology in forestry field.