Study On The Algorithms Of Several Feature Selection In Tobacco Leaf Grading

Tobacco leaf plays an important role in the area of China’s agricultural. The quality of tobacco leaf affects the economic income of tobacco industry and the health of somkers. At present, the tobacco grading is mainly rely on the manual, this subjectively grading method can cause correct rate and unnecessary disputes. Therefore, the intelligent grading method of tobacco leaf is imminent.The correct grading rate and the grading speed of tobacco are directly related to the particle application of the intelligent grading system. Under the premise of ensuring the correct grading rate, the grading speed is not only related to the grading model, but also has a great relation with the number of features. In this paper, study the feature selection. The article research as follows:1. Tobacco image capturing、preprocessing and feature selection. Use the build system to collect the 2013(462 slice tobacco leaf)、2014(1172 slice tobacco leaf) and 2015(1429 slice tobacco leaf) transmission image of tobacco. Firstly, background segmentation, denoising and preprocessing for the transmission image; Secondly, select the morphological feature, color feature, text feature, vein features total 39 features. In order to improve the grading speed use the clustering algorithm to select features and use the artificial method to remove the large correlation feature, by use the two method the number of features reduced from 39 to 24.2. Establishment the grading model of flue cured tobacco leaf. Respectively established SRC、DSRC、SVM、RF grading model, the grading accuracy and time for 2013、2014、2015 test samples of tobacco leaves are:90.1% 、92.4%、80.2%;92.8% 、93.6%、80.7%;88.9% 、92.1%、76.8%;92.38% 、96.6%、83.14%;so the feature selection of this paper use random forest as the grading model.DSRC is an improved model of SRC, the idea is based on the density to update the dictionary, by guarantee the correct grading rate, the grading speed is improved.3. Depth feature selection. In order to improve the grading speed, this paper use depth feature selection to reduce the number of features. First, the grading model is established to judge the importance of feature. According to the importance of each feature, the 24 feature is reduced to 20.Then use the improved particle swarm algorithm、ant colony algorithm、genetic algorithm and the probability of feature selected algorithm to select the optimal features. Under a certain grading accuracy, the number of features is reduced to 14、16、15、13. The grading rate and the grading speed for the 4 method to grade to tobacco is : 82.70%、0.083s;82.49%、0.090s;82.59%、0.088s;82.27%、0.076 s.In this paper, the features selected by an improved particle swarm as the optical combination of features.