Research On Leaf Image Recognition Based On CapsNet Neural Network

In recent years,more and more attention has been paid to the protection of natural environment and ecological resources.As an important part of natural resources,forest resources have a wide range of trees.Efficient information management of these trees is one of the problems that people need to solve.Rapid and accurate classification and recognition of trees is a necessary prerequisite for information management of forest resources,and leaves have certain texture shape and vein characteristics,so they are often selected as an important symbol to distinguish tree species.The traditional classification and recognition of leaves is mainly based on image processing technology.The task of classification and recognition depends on the selected features,so the generalization ability is not strong and the recognition efficiency is low.With the continuous development of deep learning,compared with the traditional plant leaf classification and recognition algorithm,convolutional neural network shows great advantages,which can quickly and effectively extract and analyze the complex deep level features in plant leaf image.However,when using traditional convolution neural network to recognize plant leaves,on the one hand,it needs a lot of training data to learn leaf features,on the other hand,the existence of multiple pooling layers leads to the loss of valuable information,which increases the difficulty of recognition.CapsNet has capsule unit,which can transform feature information into vector,and enhance the information expression ability of the model.It uses less training data and has higher accuracy.In this paper,the structure of capsule network,network parameters and image recognition are used to study the leaf recognition.The specific research includes the following aspects:(1)Capsule network innovatively puts forward vectorized capsule,which can express the object’s attitude information better than the scalar of traditional neural network,and can learn more robust features.In order to verify the classification effect of capsule network on the leaf data set,this paper first establishes the leaf data set,and uses data enhancement and generation technology to rotate,translate and transform the data and image in order to increase the number of data sets and improve the generalization ability of the model,noise and other operations are added.(2)CapsNet was trained by data set,and the classic convolutional neural networks such as AlexNet,VGG and Inception V3 were used as the control group.The experimental results show that the recognition accuracy of CapsNet is higher,the convergence speed of the model is faster,and the training data used to achieve the highest accuracy is less.In addition,the influence of dynamic routing iterations on the accuracy of the model is tested,and the test results show that the best number of dynamic routing iterations is 3.(3)Because the original CapsNet can only recognize images with lower resolution,in order to improve the recognition accuracy and realize the recognition of images with larger resolution,this paper improves CapsNet.Firstly,the structure of the capsule network is optimized.Considering the training efficiency and accuracy,a network structure combining CNN and CapsNet is obtained.Aiming at the problem that the learning rate can not be adjusted dynamically,this paper uses the optimization method of convolutional neural network for reference,introduces the Adam algorithm,and improves the compression function in the algorithm,and discusses the advantages and disadvantages of different improvement schemes through experiments.The results show that the improved scheme can improve the test accuracy of CapsNet to 98.63%.