Research On Classification Of Tree Leaf Visual Feature Fusion Based On Deep Learning

Computer vision is a new discipline that has risen rapidly since the mid-1960s.With the development of computer vision technology and the improvement of technology maturity,the application field continues to expand,showing its superior advantages.And it is also concerned by researchers in the wood industry.Applying visual technology to tree classification and resource development frees forestry science can escape technology workers from highly repetitive identification work.Due to the problems of single feature selection,insufficient information and simple classifier in the previous tree leaf identification methods,the classification accuracy of tree leaves is limited.This paper combines deep learning with computer vision technology to effectively improve the accuracy and efficiency of tree leaf identification.At present,the classification model of the combination of tree leaf feature extraction and deep learning has become a research hot spot in the wood industry today.In order to verify the feasibility of the proposed method,a large number of tree leaf samples were obtained from a total of 9,500 leaf trees through Pl@ntNet Identify,leafsnap and other leaf databases.At the same time,adaptive dynamic local three-valued pattern features and gradient direction histogram feature extraction were performed on 9500 images,and feature fusion was performed using a zero-mean standardization method.The fused features were used as the classification basis,and deep belief networks were used to train,Identify and classify them to realize the rapid identification of tree leaves.The specific research content is as follows:1.Firstly,pre-process the blades and complete the preparations before the experiment,extract the adaptive dynamic local three-valued pattern features and gradient direction histogram features for the blades respectively,and analyze the principles of the two feature extraction methods and discuss the methods.Further study the core elements of deep belief network structure and learning algorithm theory,and use it as the theoretical basis of blade classification and recognition.2.Principal component analysis is performed on the two extracted features separately,and then the fusion is carried out using a zero-mean standardization method.This method can expand the space of the leaf detail information scattering,and suppy the information with each other to enhance the relevant features of the image and improve Recognition ability of target detection.Use the fused data as input,and a deep belief network based on Dropout is trained to obtain a leaf classification model.3.Test and analyze the factors that affect the performance of the deep belief network.Firstly,a comparative experiment is conducted on the selection of the number of network layers.The analysis shows that too few layers will result in a slower convergence rate,and too many layers will cause local minimum.The experiment proves that the four-layer network is the best model.At the same time,it also conducts an experimental analysis on the influence of the number of hidden layer units.Time is related to the number of hidden layer units.Through the above analysis,the optimal number of hidden layer units,the initial selection range of parameters and the selection value of Dropout are found,and the structure and optimal parameter value of this model are determined.4.In order to solve the problem of over-fitting the model,the Dropout method is used for network training to obtain the deep belief model based on Dropout.In the training and testing of 30 sets of 6000 tree leaf sets,the classification rate of the classification model combining leaf visual feature fusion and deep learning proposed in this paper can reach 95.33%,which is better than other algorithms in the comparison experiment.The test results prove that the fused features have a richer description of the picture details.In the experiment of artificially setting random irregular lighting,although the recognition rate is not ideal,it can be concluded that the fused features are more robust to lighting than single features.