The Study Of Tree Species Identification Based On Deeping Learning Method

Conventional image-based tree species identification studies,including plant leaf identification,flower identification,bark texture identification,and wood texture identification,are all based on a single organ of the plant,and conventional machine learning methods based on manual extraction of features has an good identification effect on plant images with single pure background,but the identification effect of plants image with the complex background in the natural environment is poor.This paper proposes a tree species identification method based on deep learning and overall tree image,which realizesthe automatic recognition of tree images with complex background in natural scenes.The main contents of this paper are as follows:(1)This paper uses a combination of autonomous photography and internet crawling to build an overall tree image dataset-TreesNet,which contains a total of 1593 RGB images of 10 tree species.At the same time,in order to reduce the common over-fitting problem in machine learning,the original data is expanded by using image data augmentation such as flipping,cropping,and contrast change.The number of expanded images is 10 times that of the original data set.(2)Based on the model of AlexNet,VggNet-16,Inception-V3,and ResNet-50,the original tree species dataset and the dataset with data augmentation are used for training,and two training modes of train from scratch and transfer learning are adopted respectively.Compare the accuracy of the models obtained from transfer learning and train from scratch.(3)Based on the idea of ensemble learning,the relative majority voting method and the weighted average method are used to establish an ensemble model for the four models trained in the above direct training and transfer learning.(4)The model isfinally packaged and transplanted to the mobile terminal through Web App technology,and a tree species identification application is developed.The experimental results show that data augmentation can effectively solve the over-fitting problem and the training model has better generalization ability and higher recognition rate.At the same time,the image recognition accuracy of the tree species in the complex background with the ensemble of transfer learning method proposed in this paper reaches 99.15%,which is much better than conventional machine learning method.