Research On Target Classification Of Remote Sensing Image Based On Deep Learning

With the continuous development of satellite remote sensing technology and computer technology,computer vision technology has a very wide range of applications in the field of remote sensing images.Remote sensing images have the characteristics of complex imaging,wide variety of information,and very fast update rate.Therefore,how to use computer vision technology to quickly and accurately extract useful information from remote sensing images has become the focus of research.This paper makes use of the deep learning method that has made great achievements in image processing in recent years and semantic segmentation technology to realize the classification of remote sensing images.The research contents mainly include the following aspects:(1)A new activation function TReLU is proposed.In recent years,deep learning achieve more and more results in various research areas,and they can't be separated from the development of the activation function.However,the existing activation functions Tanh,ReLU and PReLU are exposed to more and more problems with the depth of the study,such as the existence of "neuronal death" and the bias shift,and they are not robust to noise.In view of these problems,combined with the advantages of Tanh and PReLU,the TRe LU activation function is proposed to preserve the negative half-axis activation value,the mean shifts toward zero and the soft saturation is robust to noise.The experimental results show that TReLU has the best effect on three different data sets,and is robust to different optimization methods,so TRe LU has certain practical value.(2)The convolutional neural network model structure has been improved.Aiming at the problem that the parameters of convolutional neural network model are too expensive to consume and the complex model structure consume a lot of time,proposed a method of network parallel connection and network sharing in series,which reduces the parameters and the training time by reducing the size of convolution kernels and increases the nonlinear activation,at the same time ensure the accuracy of the model.In order to further reduce the parameters,the standard normalized pooling layer is proposed to replace the full connection layer,and the improved model allows multi-size image input.Experimental results show that the proposed method can effectively improve the generalization ability of the model and improve the efficiency of the algorithm while reducing the training time.(3)Remote sensing image deblurring method was studied by using deep learning.In order to get rid of the limitation of existing methods that can only restrain motion blur or noise interference of remote sensing images,the existing denoising convolution neural network is improved.The eigenvalues of convolution layer,and the eigenvalues of transposed convolution layer,and the eigenvalues of up-sampling and convolution-recovering are combined to form the deconvolution eigenvalue to deblur the remote sensing image.Experimental results show that the proposed improved network can reconstruct the fuzzy remote sensing image effectively and achieve the best results compared with the existing methods.At the same time,it is robust to different noises.(4)Using deep learning to study the remote sensing image target classification.By analyzing the existing image segmentation methods FCN,SegNet,and U-Net,an improved target classification model is proposed.The pool indexing method in SegNet is used to achieve upsampling of the image.The pool indexing and the eigenvalues of the convolutional convolution layer and the eigenvalues of the transposition convolution layer are combined to form feature groups and collectively restore the class pixel information of the image.The experiment is compared with FCN,SegNet and U-Net methods.The proposed method achieves the best target classification effect.In conclusion,this paper has conducted some research on deep learning in the field of target classification of remote sensing images.The proposed method has obtained certain experimental results,proved the effectivity of the proposed method,and demonstrated that the research results have a certain theoretical significance and the practical application value.