Identification Of Grape Varieties Using An Optimal Feature Subset

Classification and identification of grape varieties is of great significance in promotionof improved varieties of grapes, international contract protection of the quality mark of origin,protection of grape genetic resources and the identification of new hybrids. Compared withmanual work, computer-aided classification of grape varieties can deal with the classificationtasks with high efficiency, thus bringing direct and potential economic value to the society.With superior and inferior varieties planted together, the quality of grape wine areseriously constrained, inferior varieties thus need to be removed, superior ones to bepromoted. However, such demands as recognizing the grape varieties are difficult to besatisfied because of the high requirements on the grapes professional knowledge. Theultimate purpose of this research is to establish a special platform to identity them quicklyand accurately for grape growers and wine producers.The objective of the study is for single and batch identification of grape varieties usingtheir leaves with high accuracy. The main contents and conclusions of this target are asfollows:(1) Robustness of two shape features (relative geometrical features and modified Fourierdescriptors) is researched. Under transformation of rotation, translation and scale, the valuesof relative geometrical features varies slightly and exhibits robustness, while modifies Fourierdescriptors are sensitive to those transformations and thus not suitable for grape varietyclassification.(2) For the first time the wavelet GLCM features which fuses the wavelet decompositionand GLCM statistical features are used in the grape variety classification. In the meantime, inorder to compare, the Hu invariant moments, Gabor wavelet feature and local binary pattern(the Local Binary Patterns, LBP) feature are also calculated. Research shows that the waveletGLCM features can best represent grape leaf textures. By using this texture, a much higheraccuracy of79.90%is achieved than other textures.(3) The principal component analysis (Principal Component Analysis, PCA) andmulti-class linear discriminant analysis (Linear Discriminant Analysis, LDA) are adopted toreduce the dimension of the feature set. The recognition rate of the PCA method is merely 56.99%, while multi-class LDA reaches79.90%, indicating that the multi-class LDA methodcan generate an optimal feature subset with stronger classification ability.(4) SVM (Support Vector Machine, SVM) classifier is adopted to classify the grapevarieties, and K-nearest neighbor classifier (K-the Nearest Neighbors, the K-NN),probabilistic neural network (Probabilistic Neural Networks PNN) classifier are for validationof the optimal subset. Experimental results show that it can be used to characterize the featureof grape leaf with a good classification rate.(5) Under Matlab2009a, a system for classification of grape varieties using their leavesis designed. Functions like appropriate pretreatment as well as several shape and textureextractions are realized, after which individual identification and batch classification can thenbe achieved, the average recognition rate is79.90%.