Study On Classification And Recognition Method Of Bamboo Subfamily(Bambusoideae) Plant Based On Composite Features Of Bamboo Leaf Images

China’s bamboo forest resources are extremely rich.Bamboo plants have high morphological similarity.Their flowering and fruiting cycles are long,and flowering is usually accompanied by the death of bamboo forests.Thus,the traditional angiosperm method with flower and fruit morphological structure and DNA molecular methods is difficult to fully apply to the the classification of bamboo plants.With the development of smart forestry,"digital bamboo recognition" in the new form is one of the development directions of modern bamboo taxonomy.Exploring new ways to classify the bamboo subfamily through image processing technology of multi-feature fusion can reduce the difficulty of artificial identification,promote the scientific unification of bamboo classification,realize the mechanized identification of bamboo plants.Additionally,it is important for studying the quantidication of characteristic parameters of bamboo species,the phylogenetic trees of the bamboo subfamily(Bambusoideae)and the interspecific relationships.A total of 70 species of bamboo plants in 17 genera were collected as samples from the "Century Bamboo Garden" in Changning County,Sichuan Province and the bamboo forest wetland in Qingshen County.A digital scanner was used to obtain images of bamboo leaves of normal size,and a digital camera was used to obtain images of large bamboo leaves.A total of 3,500 images were selected to establish a BZY bamboo image database.From the BZY database,1760 images of 44 species healthy and mature bamboo leaves from 15 genera were selected(including 12 types of scattered bamboo,18 types of clumped bamboo and 14 types of mixed bamboo).Each bamboo species,30 leaf images were selected for the training set and 10 images for the test set.Matlab2018 b software was used to pre-normalize,grayscale,median filter,binarize,morphological filter and edge detection preprocessing.We extracted 16-dimensional shape features such as aspect ratio,area,shape parameters,Hu invariant moment,etc;extract 10-dimensional texture features such as energy,contrast,and entropy based on gray-scale histogram and gray-level co-occurrence matrix;extract 80-dimensional SP-Forioer features using improved fan-shaped projection Fourier descriptor(SP-FD);extract 400,000-dimensional features using 40 Gabor fulters,and output 300-dimensional Gabor features through PCA dimensionality reduction.Finally,the training set data is used to train the extreme learning machine,support vector machine,probabilistic neural network and GA-BP neural network classifier to obtain the optimal training model,and the test set data is tested to obtain the final classification recognition rate.Based on the composite features of bamboo leaf images,this study explores a new way to classify plants of the bamboo subfamily.The main results of this study are as follows:(1)The pretreatment process for the leaves of 44 types of bamboo images was done by Matlab2018 b programming software.Based on the leaf morphology extraction,we extracted the shape features and texture feature vector values of the bamboo leaves.The results showed that these eigenvalues all followed the Weibull distribution,reflecting the differences in bamboo leaf age,variability,and location.(2)The leaf image feature values extracted from the computer can reflect the characteristics of real bamboo leaves numerically,and the description of small features such as leaf vein texture roughness,leaf shape width and narrowness is more precise and scientific,so it can be used as a quantitative reference basis for bamboo taxonomy.(3)Multi-feature fusion of shape and texture can effectively improve the accuracy of bamboo leaf classification and recognition,eespecially in the classification of different underground stem types,the same genus or different genus and the overall bamboo species classification.The composite features have improved the classification recognition accuracy of 20.00%,15.56%,7.86% and 12.5% on scattered bamboo,clumped bamboo,mixed bamboo and comprehensive classification.For single and composite features,using the Extreme Learning Machine(ELM),the highest recognition accuracy rates for 12 types of scattered bamboo,18 types of clumped bamboo,14 types of mixed bamboo and comprehensive classification are 72.50%,75.56%,80.14% and 85.00%.(4)When dealing with small sample problems,the classification accuracy increased.But the number of bamboo species within the bamboo genus increases,the classification accuracy decreases,but it was still more than 80%,and also has had a high recognition accuracy rate for variants.In the identification of bamboo varieties,achieving 100%classification accuracy.In the identification of different genera,Dendrocalamopsis,Shibataea and Pseudosasa the highest bamboo species recognition rate reacheds 100%;for Dendrocalamus and Pleioblastus,the highest classification accuracy rate can reach was more than 95%;for both Phyllostachys and Bambusa,and the highest classification accuracy rates are were 85.56% and 84.44%,respectively.Among the 15 genus bamboo species,each genus selects a total of 15 types of bamboo species were selected from each genus,and the classification accuracy can could reach 81.33%.(5)In this study,extreme learning machine,support vector machine,probabilistic neural network and genetic algorithm optimized BP neural network classifier were used to identify bamboo leaf images respectively,and the highest classification rate reached72.45%.The extreme learning machine has a good generalization ability for bamboo leaf image classification and recognition.However,due to the interval and complexity of the leaf characteristics of the same species or even the same bamboo species,the performance of the classifier still needs to be improved and improved.In summary,due to the volatility and high similarity of the bamboo leaf morphology itself,it is difficult to classify and recognize bamboos using bamboo leaf images for classification and recognition is a complicated and difficult task..The final classification recognition rate obtained in this study is higher when the number of samples is small while it,but the accuracy is not high when the sample size is high,which is related to the extraction of feature dimensions,bamboo leaf similarity and classifier performance.Therefore,it is a long-term work to use image processing techniques to extract the morphological features of the bamboo subfamily and improve the performance of the classifier.This method can provide a practical and scientific basis for the classification and recognition of the bamboo subfamily.