Study On Image Feature Selecting And Grading Of Tobacco Leaf Based On BPSO

The correct classification of tobacco affects tremendously its economic valueand the quality of the cigarette. Artificial grading used in the actual process results inthe recognition rate influenced by subjective factors. The study on tobaccointelligent classification, therefore, is very necessary. All the visual features relatedwith the tobacco grading are extracted from the tobacco image in this dissertation.To reduce the correlation and improve the grading speed, the self-adapting selectionmethods based on the binary particle swarm optimization (BPSO) are proposed toreduce the features. The features are self-adapting selected respectively by RBF,SVM, and Adaboost combined BPSO. And RBF, SVM, and Adaboost gradingmodels are reconstructed based on those three selected features respectively. Theexperimental results show that the feature selection can improve not only thegrading speed but also the correct recognition rate.. The main studies are as follow:First, tobacco image capturing, preprocessing and feature extracting. A systemis build to obtain the tobacco images. The filter, edge extracting, backgroundmoving are carried out. Then8geometry,12color,4texture and15vein totally39visual features are extracted from the transmission images.Second, tobacco grading models reconstructing. RBF neural network, SVM andAdaboost are reconstructed and trained based39features. The correct grading rateof training samples and test samples are89.6%and85.5%,89.3%and86.6%,96.3%and93.2%respectively.Last, Feature selecting and grading. Visual features are self-adapting selectedby BPSO combines RBF neural network, SVM or Adaboost respectively.16,21and21features are selected from all39by these three selection methods respectively.9grading models are reconstructed based on these3kinds feature combinations. Allgrading results show that:(1) these selecting methods based on BPSO can removesome features which have large correlation or less influence to grade effectively.(2)the correct recognition rate is affected not only by selected algorithm but also the grading method. The best rate can be obtained when they are matched.(3) Althoughthe selected features are different by different method, all of them can improve thegrading speed and correct rate for different classification models. This means thatthe selected features have some generalization performance.