| Tree species identification is an important work in forest resource management and forestry production construction.In China,the shortage of forest resources is becoming increasingly serious,and the accurate identification of tree species is the basis for protecting forest resources and studying the diversity of tree species,which is of great significance for strengthening the construction of ecological civilization.As tree leaves have the advantages of easy collection and long preservation time,they have become an important research object for tree species identification and classification.However,in the real experimental environment,due to the limited number of samples,it is often impossible to reach the expected results when using deep learning methods for classification and recognition of small samples.Based on this,this paper proposes to use generative adversarial network to expand the samples and apply the expanded datasets with the residual attention structure for accurate recognition of tree species categories.In this paper,10 tree leaf images of Populus tremula,Fraxinus excelsior,Fraxinus domestica,Acacia japonica,Lilac,Eucommia sylvestris,Torch tree,Sagebrush,Ashwagandha,and Red endosperm were used as research objects.First,the acquired images were subjected to size normalized,denoised by bilateral filtered,and histogram equalized to reduce the influence of objective factors such as light intensity and shooting angle on recognition results.Then,the dataset is expanded using a generative adversarial network.Based on the deep convolutional generative adversarial network(DCGAN)model,the residual conditional deep convolutional generative adversarial network(RC-DCGAN)is established,and the category labels are input into the generator to control the sample generation process;at the same time,the residual structure is embedded into the generator to improve the quality of the generated images,and finally the RC-DCGAN model is used to expand the dataset.Finally,the tree species are classified and recognized.On the one hand,the VGG-16 and Dense Net121 pre-trained models are used for migration training on the leaf image datasets;on the other hand,the discriminant network of RC-DCGAN is extracted for leaf image recognition,and in order to explore the effect of model depth on the recognition effect,the image features are improved on its basis by using the residual attention structure to extract the image features to enhance the important feature information learning and feature transfer between layers to establish the residual attention network(RSENet)model.The experimental results show that the leaf images generated by the RC-DCGAN model are better than those generated by the DCGAN model in terms of both image sharpness and image diversity,with FID and IS values of 44.68 and 2.38,respectively;the recognition accuracy of the VGG-16 and Dense Net121 pre-trained models on the test set is 96.20% and 93.20%,respectively;the recognition accuracy of the RC＿DCGAN discriminative network on the test set The recognition accuracy of RSENet network on the test set is 97.20%;on the same data set,the recognition accuracy obtained by Alex Net model,VGG-19 model,Res Net18 model,and Res Net50 model are 88.72%,90.50%,93.25%,and 95.46%,respectively,which are lower than the the recognition accuracies of the models in this paper.The method of combining generative adversarial network and residual attention has the highest recognition accuracy for the 10 tree species in this paper,which overcomes the problem that the generalization ability of the model decreases by using traditional image processing augmentation methods,indicating that the method of image augmentation using generative adversarial network is feasible and effective.The number of parameters and memory consumption,improved recognition accuracy,and faster convergence,which is a reference value for future related work. |
