关键词: 卷积神经网络, 树种识别, 生成对抗网络, 残差结构

Abstract: Objective When using convolutional neural network models for automatic image recognition, a large number of training samples are usually required to prevent model overfitting. In this study, we propose a tree species recognition method that fuses generative adversarial networks with convolutional neural networks to improve the accuracy of tree species recognition and expand the image samples based on the original leaf images to ensure the training quality.Method Under the framework of Pytorch, images of the leaves of 10 common tree species (aspen, ash maple, house elm, locust, clove, eucommia, torch tree, mountain wattle, ash willow, red endosperm) were collected as research objects. First, the images were pre-processed using mean filter denoising and size normalization. Second, the datasets are expanded with images generated by generative adversarial network, in which the residual conditional deep convolutional generative adversarial network (RC-DCGAN) model is built based on and improved by the deep convolutional generative adversarial network (DCGAN) model, which takes random noise and category labels as the input of the generator to control the sample generation process; the residual structure is embedded in the generator so that the generative model learn more feature information to improve the quality of the generated images. Then, the original and augmented images are used as the training set of the convolutional neural network (CNN). On the one hand, 11 400 images were expanded using the RC-DCGAN model and traditional image expansion methods such as rotation, mirroring, and changing contrast, respectively. On the other hand, the original and generated images, the original and traditional expanded images were inputted to the CNN for training, and 50 images in each category of the original images were randomly selected to test the model to verify the feasibility of the generative adversarial networks to improve the recognition accuracy. Finally, the CNN classification model suitable for the experimental requirements are determined and compared with the recognition effect of AlexNet model, VGG-16 model, VGG-19 model and ResNet18 model to test the feasibility of the method in this study.Result The RC-DCGAN model generates higher quality images than the DCGAN model, which fit the real image; the training accuracy is 99.03% and the average validation recognition accuracy is 97.20% using the generative adversarial network augmented image method with the ResNet30 tree recognition model; while the recognition rate of the traditional image augmentation method is 95.50% under the same tree recognition model; under the same dataset, the recognition rates obtained by AlexNet model, VGG-16 model, VGG-19 model, and ResNet18 model were 86.52%, 87.57%, 91.43%, and 93.25%, respectively, which were lower than that of the models in this study.Conclusion The method of combining generative adversarial network and convolutional neural network has the highest recognition accuracy for the leaf images of 10 tree species in this study, which overcomes the problem of decreasing the generalization ability of the model by using the traditional image processing augmentation method, and shows that the method of image amplification by generating adversarial network is feasible and effective, which can provide reference significance for related research work.

Key words: convolutional neural network, tree species identification, generative adversarial network, residual structure