Research On Tree Species Recognition Based On Optimized ResNet50 Neural Network

As a major component of the earth's terrestrial ecosystem,forests play a vital role in maintaining ecological balance.China is a country devoid of trees and forests where identifying tree species correctly holds the key to the protection of forests.However,the traditional method of feature extraction by mere hands is low-efficient,inaccurate,complex and demanding with regard to the expertise of the operator.In light of this fact,this paper proposes an automatic method that identifies the images of bark textures based on convolutional neural network.Revolving around this aspect,this paper conducts the following researches.Traditional identification methods are limited.In order to deal with this problem,this paper studies the automatic identifications and classifications of up to 12800 images of bark textures on the basis of one shallow convolution neural network and two deep convolution neural networks.These data derive from such common trees in Northeastern China as Fraxinus mandshurica Rupr and Ulmus pumila L and Betula platyphylla Suk and Picea koraiensis Nakai and Quercus mongolica Fisch.ex Ledeb are divided into training sets and test sets in the ratio of 7:3.(1)The shallow neural network is mainly composed of two convolutions,the largest pooling layer and the output layer.The identification accuracy of the test set is 84.51%.(2)Shallow CNN is not fully capable of extraction according to features.Therefore,the classic ResNet50 network model is harnessed to improve the identification accuracy to 89.1%.However,it cannot be fully applied to textures deficient in clear-cut features in this research.(3)In order to address the problem that the classical network is likely to confuse trees with similar textures,the research adds cross-scale fusion connection layer between residual blocks and changes the average pooling of output modules to bilinear pooling operation on the basis of the classic ResNet50 network.In this way,the emphasis weight can be given to the high response region of shallow layer,and the features of the corresponding deep layer in the same position can be activated.This is conducive to extracting texture information better and improving the ability of distinguishing similar features.Bilinear pooling can reemphasize texture features,reuse low-level features,and enhance the accuracy of the network.The identification accuracy of the optimized ResNet50 network is 96.13%.Through the comparison and analysis of the confusion matrix,ROC curve and AUC in the above three convolution neural network identification process,the improved ResNet50 network in this paper is proved to be accurate and effective.The automatic method that identifies the images of bark textures based on convolutional neural network is feasible.Deep CNN is more effective than shallow CNN.The optimized ResNet50 network proposed in this paper improves the performance and accuracy of network identification,which can provide some insights for intelligent identification of tree species and offer new perspectives for protecting forests in a reasonable manner.