Feature Extraction And Structural Optimization For Image Classification Based On CNN

Machine learning has developed rapidly in recent years.Deep learning has become a new area of machine learning and has been favored by a large number of scientific researchers.Convolutional Neural Networks(CNN)has been caused by its powerful feature learning ability by the majority of scientific researchers have paid great attention to it and obtained considerable development.Since the neural network was proposed by Hubel and Wiesel,CNN has come to the fore in many scientific fields and has become a dazzling new star.Especially in the category of pattern classification,its advantage lies in its ability to directly input the original image as input data and avoid it.In the earlier period,complex and huge features were extracted from the original image,which resulted in a more universal application.The first CNN was born in 1980 and realized by K.Fukushima.Subsequently,more researchers working in the field improved the neural network and applied more advanced algorithms to the model.This paper systematically studies the effects of convolutional neural network algorithms on the accuracy of grassland classification under remote sensing images and the improvement of convolutional neural network framework.The main research contents are as follows:(1)The paper studies the application of convolutional neural network model in grassland classification of remote sensing imagery,and proposes a grassland remote sensing classification network structure based on principal component analysis(PCA)bleaching convolution neural network.It can effectively use PCA bleaching for remote sensing image data.Reduce the correlation between data,speed up the learning rate of neural networks,and enhance the ability of feature learning.Based on this,the sampling pool is randomly pooled,which improves the generalization ability of network classification and improves the accuracy of grassland classification.(2)Based on(1),we deeply studied the algorithm of grassland classification for high remote sensing images,analyzed and studied the image features extracted by convolutional neural network,and found that some important features may be lost in the image are through the convolution layer and the pool,then a feature-basedintegration algorithm for deep convolutional neural network grassland classification algorithm.Firstly,the remote sensing image data is processed by PCA whitening to reduce the correlation between data and accelerate the speed of neural network learning.Secondly,the bi-linear integration of shallow features and deep features extracted from convolutional neural networks is performed.The new features after integration are more perfect and optimized.Finally,for the training of remote sensing data,due to the increase of effective information in new features,the feature expression ability is improved,and the purpose of improving the classification accuracy of grassland is achieved.Experiments show that the bilinear integration algorithm based on PCA whitening can effectively improve the accuracy of grassland classification.(3)Based on LeNet-5 model,an improved LeNet-5 Convolutional Neural Network(ILN-CNN)model is proposed to solve the problems of long time and low recognition rate of existing traffic sign recognition algorithms..Firstly,the original LeNet-5 convolutional neural network model is used to construct two relatively independent sub-convolutional networks with different convolution kernels to speed up feature extraction.Second,increase the number of convolution kernels in the subnetwork to enhance the ability of the network to distinguish different traffic signs.Finally,add the activation function ReLU,increase the Dropout layer,in order to speed up the convergence of the function,avoid CNN overfitting,and reduce the effect of the inter-neuronal adaptation.Experiments show that using the ILN-CNN model can effectively extract the characteristic information of traffic signs and then identify them efficiently.The recognition accuracy rate is 93.558%,which is 12.206%and 4.018% higher than that of BP and SVM,respectively.