Researches On Classification Method Of Chinese Medicine Plant Leaves Based On Digital Image

Objective In order to overcome the limitations of artificial classification methods to identify the Chinese medicinal plants, the paper researches on an automatic classification method based on leaf images of medicinal plants.Methods 1. It collects 12 kinds of Chinese medicinal plant leaf images and does the operation of denoising, sharpening, binarization, and etc. It uses the method of threshold segmentation to remove leaf image background, and filling the hole for binarization image by morphological processing. Finally, the leaf image edge is extracted by the edge tracking algorithm. 2. From the point of multi-feature extraction, the 10 shape features and 5 texture features are extracted. 1) Aiming at the binary images of the leaves, it computes three relative geometrical features(eccentricity, rectangularity and circularity), and using the classical algorithm of Hu invariant moments to extract seven moment invariants. 2) For the gray level images of leaves, it uses the GLCM algorithm to extract five texture features, respectively, such as energy, entropy, contrast, correlation, and inverse difference moment. 3. The SVM, BP neural network, PNN and KNN classifiers are used for the classification of Chinese medicinal plant leaf image. Through the contrast experiment, the classification method which has higher recognition rate is chosen.Results The experiment uses four kinds of classifiers to classify the Chinese medicinal plant leaf images respectively. Among them, the correct rate of the SVM classifier is 93.33%; the correct rate of the BP neural network classifier is 87.5%; and the correct rate of the PNN classifier is 91.67%; the correct rate of the KNN classifier is 91.67%.Conclusion 1. It is feasible to use the leaf image to carry on the automatic classification of Chinese medicinal plant leaf, and has higher accuracy; 2. In terms of the image recognition of Chinese medicinal plant leaves, the classification method of the SVM is superior to the BP neural network, PNN and KNN in this experiment.