Study Of Tree Leaf Recognition In Habitat Based On Deep Convolutional Neural Networks

Due to the fact that there are a variety of trees in the field and the surrounding environment is complex,it is of great significance to use advanced target detection technology to identify trees in natural habitats,which can improve the quality and efficiency of forest resources exploration.The leaves of trees contain a large amount of species information such as texture structure,shape characteristics,and color,which makes it feasible to identify tree species through leaves in complex environment.Compared with the flowers or fruits of trees,the survival time of leaves is longer,which is convenient for image sample collection and sample dataset construction.The rapid development of machine vision technology and deep learning field makes it possible to identify specific targets in complex environments.In the recognition field of deep learning,deep convolutional neural networks have extremely important contributions.This is mainly due to the strong fitting ability,representation ability and abstract feature extraction ability of deep convolutional neural networks.In this dissertation,ten kinds of tree leaves growing in northern part of China are taken as the research objects,which are collected as samples,and the generative adversarial network is improved through the combination of feature fusion and deep convolutional networks,so as to achieve the increase of sample numbers and then create the dataset of leaves.Then according to the image size,the number of species and other factors,the ability of the deep convolutional neural networks to recognize leaves is enhanced in a targeted manner.The improved deep convolutional neural networks are mainly through optimizing the structure and algorithm,forming a deep learning framework suitable for discriminating small leaves,and then exploring methods of identifying different tree leaves in complex environments.The main contents of this research are as follows:(1)Ten species of leaves of dicotyledonous plants growing in different periods,illumination and habitats(Syringa oblata,Tilia amurensis,Fraxinus mandshurica,Acer mono,Crataegus pinnatifida,Bischofia polycarpa,Euonymus fortune,Ulmus pumila,Prunus Cerasifera,Salix babylonica)are selected as the research objects,and then construct the tree leaf dataset under complex habitat conditions.Aiming at the problem of local feature collapse and the lack of diversity in the generated images by traditional generative adversarial network,a feature fusion deep convolutional generative adversarial network is proposed to generate leaf images.Convolution neural network is used instead of multilayer perceptron in this model.By adding small-sized convolution kernels to the generative network and the discriminative network to assist the model to learn the leaf characteristics,and the network structure of feature fusion is introduced into the discriminative network,which makes the model learn more abundant target details.In order to improve the fitting ability of leaves under unsupervised condition,the loss function of the least square method is used to replace the cross entropy loss function.Comparative experiments show that the proposed network model can effectively reduce the loss of the discriminative network and the loss of the generative network,and then generate more tree leaf images with different complex habitat backgrounds to achieve the purpose of increasing leaf samples.In addition,the serial numbers and naming rules of leaf images are formulated according to different categories,and then a tree leaf dataset is constructed to provide data support for the training of deep convolutional neural networks.(2)Aiming at the recognition of single species of tree leaves in a small-size image under complex habitats,and making the recognition model have better characterization ability and avoid increasing the parameter calculation,a network model based on multi-channel and multi-convolution kernel decomposition is proposed.The model is mainly designed for the structure of the feature extraction network.By using multiple branches and multiple convolution kernels in the deep feature extraction layer to perform multi-dimensional feature extraction on the leaf image can enhance the network model’s ability to distinguish similar abstract leaf features.Meanwhile,two improved methods of convolution kernel channel number decomposition and convolution kernel size decomposition are introduced to optimize the network structure,and are dedicated to reducing the calculation parameters of the model,making the model more suitable for leaf recognition in small-size images.Through experimental comparison,the three aspects of convolution kernel size,loss function and multiple convolution kernel decomposition show that the proposed method can effectively improve the identification accuracy of single species of tree leaves in natural environment.(3)Aiming at the recognition of multiple tree leaves in large-size images in complex habitats and ensuring the leaf features learned by the designed deep network model have good stability in deep network propagation,a model based on deep residual-convolutional connection is proposed.This model is mainly used to learn the subtle differences between similar leaves by constructing three different residual-convolutional modules.These three modules are named as residual module with parallel-equivalent branch,residual module with cross-level and equivalent branch and improved residual module with convolution branch.Among them,residual module with parallel-equivalent branch and residual module with cross-level and equivalent branch are used to increase the depth of the network;improved residual module with convolution branch is used to replace the pooling layer to adjust the size and the channel of feature maps.In addition,in order to make the recognition accuracy of the model does not be affected by the number of training samples,an enhanced normalization is proposed to process the input tree leaf features,thereby reducing the deviation caused by high-volume leaf images and small batch size.Through experimental comparison,it is shown that the proposed method can improve the identification accuracy of various tree leaves without relying on the number of training samples.In summary,the study of tree leaf recognition in habitat based on deep convolutional neural networks in this dissertation can accurately recognize a variety of small tree leaf targets in complex habitats.In addition,this article has certain research value and guiding significance in the identification and protection of tree species.