Study On The Algorithms Of Effective Spectral Feature Selection And Classification In Flue-cured Tobacco Leaf Grading

At present, the flue-cured tobacco leaves are graded by artificial classification inour country. It is not only time-consuming, laborious, but also subjective.Classification accuracy depends on the personnel experience and the environment, sothe flue-cured tobacco auto-grading is imperative.The research on tobacco auto-grading mainly focuses on the classifying methodsbased on image features. The colors, texture, geometric features which are related tofactors of the artificial classification are extracted from the image of tobacco leavesand some pattern recognition methods are used to grade them. However, imagefeatures are difficult to reflect the factors of thickness, leaf structure, oil content andothers that are closely associated with tobacco classification. Spectrum can reflect thestructure and composition of materials, which is widely used in many fields. Thenear-infrared spectrum of flue-cured tobacco can reflect much information, but thehigh dimension data results in a little long grading time. This dissertation focuses onthe follows:1. Research on the spectral type, scope, interval and preprocessing methods.The reflectance spectrum between1500nm and2400nm with10nm (maximumsampling interval) is selected to grade the tobacco leaves at last by comparing theclassification accuracy combined with grading speed and convenience of spectralacquisition.2. The parameters optimization of SVM. The linear kernel function and the thegrid optimization method are employed in SVM by Comparing the grading accuraciesof the RBF and linear kernel functions and PSO, GA, and the grid method ofparameter optimization methods.3. Research on flue-cured tobacco grading model. A SVM multi-classclassification can be reconstructed by building some SVM two class classifiers. Thegrading result is obtained based on the votes of each two-classifier. Obversely thereare two kinds constructing methods: the cascade (first grouping then grading) and theparallel classification models. Cascade model needs less two-classifiers which means less grading time (0.114819s for2nm spectrum) but exists accumulative error(accuracy is89.8%for2nm). Parallel model requires more two-classifiers and a littlelong time (0.396101s); however has higher accuracy (93.88%).4. Selecting spectral features based on Clustering. The clustering method isapplied to select the spectral features. The intra-class parameters γ1and the inter-classparameters γ2are chosen to move some spectrum which have bad influence on theclassify. The number of spectrum reduces30%(from451to312) for the cascademodel with the grading accuracy87.78%,5. Selecting spectral features based on the BPSO and selected probability. Inorder to reduce the number of spectrum, the BPSO is adopted after clusteringselection. The improved selecting method is proposed by comparing the best spectralfeatures (in1000times) and the selected probability of each spectrum. The lastspectrum features are constructed by the spectra with larger then60%selectedprobability and the best ones. The number decrease48%from91to47, the gradingtime is0.040926s, the accuracy is88.75%for parallel model with10nm samplinginterval.