Research On Plant Leaf Image Classification Based On Stacked Auto-Encoder Network

Plants are one of the most basic components of life on the earth.Protecting plant species diversity is essential for maintaining the balance of the earth's ecosystem.The premise of plant protection is to classify and recognize plants accurately,so as to build a perfect plant database system.In order to solve the problem that the accuracy of plant leaf classification based on traditional machine learning method is low,the stacked auto-encoder is used to automatically learn the high-level representation features of data from the original data.The image classification of plant leaf based on stacked auto-encoder is studied,and a network model with good performance in image classification of plant leaf is constructed to improve the classification of plant leaf.The main contents of this paper are as follows:(1)In order to solve the problem of over-fitting of stacked auto-encoder,a layer of denoising auto-encoder is connected in the input layer,and the training data is damaged by random zero placement.By learning the features of damaged data,more robust features can be obtained.The difference between training data and test data can reduce the possibility of over-fitting in training.(2)To solve the problem that sparse auto-encoder can not restrict each input neuron sparsely,k sparse method is used in sparse auto-encoder.Only k neurons with the highest activation value are retained in the hidden layer,and the remaining activated neurons are returned to zero.This method realizes the sparse restriction of each input neuron in the hidden layer,makes the sparse data mismatch between the training stage and the testing stage,and concentrates more on the feature representation of the key information of the blade image,which improves the classification accuracy.(3)Aiming at the problem of long training time caused by multi-layer structure and a large number of neurons of stacked auto-encoder,adding batch normalization to input data of sparse auto-encoder in each layer to solve the problem of internal covariant offset can accelerate the training speed of network model and reduce thetraining time.(4)Design the network model analogy experiment to verify that the improved stacked sparse denoising auto-encoder improves the classification accuracy and robustness in plant classification.