Research And Implementation Of Leaf Pattern Classification Method For Broad-leaved Tree Species

Different characteristics of the texture and contours of hardwood leaves can be used for tree species classification.Because of the huge number of tree species,there are a large number of tree species related data reserves for classification.Generally non-professionals do not possess these skills and can only rely on computers to store large amounts of image data.The popularization of high-speed Internet technology provides a good research and use environment for computer tree leaf identification.Firstly,a total of 360 pieces of common broad-leaved leaves in Northeast China were collected,including hair cherry,honeysuckle Lonicerae,lilac,and aster,and the images were collected after cleaning and placed in a single background to generate 360 leaf images.Using OpenCV image processing to develop a program package,using C++ language program,image preprocessing is performed on acquired images to obtain images that are conducive to acquiring features,and the images are divided into test and training sets.Extract leaf area and circumcircle area ratio of each image,aspect ratio,ratio of area to contour perimeter,area to length ratio,area to width ratio,entropy of pixels in image,leaf vein texture entropy,contour of leaf edge The entropy of the total of eight kinds of characteristic values,and these characteristic data are stored in the data file.Using the training set feature data to train in the linear classifier,the trained model is tested using the test set data,and finally the recognition rate is 87%.In order to improve the recognition rate,the convolutional neural network is used for image feature research.The structural order of the convolutional neural network is one input layer,two convolution layers,and one pooling layer.Finally,it is fully linked to a layer of excitation layers composed of RELU functions to implement classification.The training data set and the test data set image are directly put into the network model.The network obtains the optimal state through the training set image training adjustment weights.The training method is the stochastic gradient descent algorithm.Using the convolutional neural network,the final result is that the average recognition rate is 94%,and the highest one is 98%.The experimental results show that the use of convolutional neural network deep learning model has high classification and recognition rate for broad-leaved tree leaf images.Using the feature map extracted by the convolution kernel,more details of the image can be obtained,and the design and extraction of feature values are not required.The disadvantage is that a large number of broad-leaved tree leaf images are required for network training,but with the development of computer technology,data-driven models have more development significance than models classified by feature values.At the same time,the model after training in this experiment provides some theoretical support and practical models for the classification and research of tree species.