Research On Remote Sensing Image Restoration Based On Full Convolutional Neural Network

Remote sensing technology is widely used in many fields such as geography,territorial science,and eco-environment,and is one of the important means for acquiring geographic information data.The remote sensing image is the image of the sensor during the orbital operation.It is mainly divided into panchromatic image,hyperspectral image,multispectral image and radar image.However,remote sensing images are affected by many factors such as sensor failure during the imaging process,and it is difficult to avoid the loss of local information.This problem seriously restricts the application of remote sensing images.Accurate reconstruction of lost remote sensing image information is of great significance to the analysis and application of remote sensing image data.At present,there are many algorithms for image restoration at home and abroad.Most of these algorithms need to establish a degenerate model and obtain sensor platform parameters.Under normal circumstances,some image degradation is difficult to find a corresponding model to express,and platform parameters are also difficult to obtain.Therefore,the establishment of an accurate and efficient fixed restoration framework is of great significance for the restoration of remote sensing images.In recent years,deep learning algorithms represented by the Convolution Neural Network(CNN)have shown powerful performance in the classification and recognition of remote sensing images.This paper builds the FCN restoration framework model on the basis of CNN,establishes the mapping relationship between the frameworks through sample training,and builds a spatial scale to enhance FCN to repair the effect of panchromatic band remote sensing imagery;based on FCN remote sensing image object segmentation,it searches on the same type of features.Similar pixels are used to achieve spectral repair of remote sensing images through the replacement of similar pixels.The main research contents of this article are as follows:(1)By training about 30,000 different remote sensing image samples,a full convolutional neural network repair model was established.Several images are added to add noise and a repair model is input.When any one of the output images is highly similar to its corresponding clear image sample,the training is stopped,the network parameters are fixed,and a repair model is generated.Selecting a remote sensing image to simulate the repair effect of the three information loss types contrasting the interpolation method and the Wiener filtering method,it is concluded that the method is superior to the above two methods in terms of average PSNR,SSIM and time,verifying the whole The effectiveness of the convolutional neural network repair model.(2)In order to improve the repair effect of FCN,Gaussian pyramid and Laplacian pyramid are used to construct the spatial scale to suppress image noise and fuse the images after restoration with different scales.Through experiments,it is proved that the PSNR and SSIM are improved by 1.403 and 0.149 respectively on the basis of the original repair model.Effectively improve the quality of remote sensing image restoration of FCN restoration framework,especially for the restoration of the texture of the panchromatic band of remote sensing imagery.(3)Mean-shift method is used to effectively improve the accuracy of segmentation of the same feature on the auxiliary image of the FCN,and the similarity is sought by observing the principle of correlation between the minimum mean value of the absolute value of the pixel values in each band and the pixel.Meta and cell replacement by location finding.Finding similar pixels on the same type of features reduces the probability of misclassification of "same-color foreign objects." And analyze the spectrum of the repaired image.The results show that the similar pixel replacement method based on FCN segmentation effectively restores the spectral information of remote sensing images.