Satellite Remote Sensing Image Fusion Driving By Deep Convolutional Neural Network Super Reconstruction

Image super-resolution reconstruction technology is widely used in remote sensing image processing,medical imaging,video surveillance,traffic violation monitoring and other fields,and has strong application value.For the research of image super-resolution reconstruction methods,the development trend is from interpolation-based,reconstruction-based,learning-based algorithms to deep learning based in recent years,based on deep learning super-resolution reconstruction such as SRCNN,FSRCNN,VDSR,DRCN and other algorithms.To some extent,it can improve the quality of images after super-resolution reconstruction,but there are still some shortcomings,such as too few convolution layers,and can not fully learn the detailed features between low-resolution images and high-resolution images.At the same time,in the fusion process of traditional satellite remote sensing images,the multi-spectral images are mostly enlarged to the same size as the full-color images by simple interpolation.The focus of the fusion is the spatial detail information and low resolution of the panchromatic band.Merging of multispectral spectral information.Therefore,there is a case where the low-resolution spatial information is not fully utilized in the interpolation process.Based on the previous studies,this paper improves the problem of information loss and the insufficiency of image super-segment reconstruction in the process of traditional satellite remote sensing image fusion,and improves the super-resolution reconstruction method based on multi-scale convolutional neural network.While retaining spectral information,the spatial information of low-resolution multi-spectral images is enhanced.Secondly,the multi-spectral and panchromatic image of spatial information enhancement is fused by Schmidt orthogonal transform to obtain a fusion result with high spatial resolution and good spectrum.Finally,the fusion model and the four classical conventional fusion algorithms are compared by qualitative and qualitative quality evaluation indicators.The method has achieved good results in subjective and objective quality evaluation.The main research contents of this paper are as follows:(1)Review image super-resolution reconstruction technology,domestic and foreign researchers summarize the principle,algorithm and research significance of super resolution reconstruction technology.Super resolution reconstruction technology includes interpolation-based,reconstruction-based and learning based algorithms.The image preprocessing steps and several common methods of image fusion are introduced.The image quality evaluation indicators are introduced to evaluate the image quality after over segment reconstruction.(2)Combining convolutional neural network model in deep learning,convolutional neural network(CNN)structure and algorithm,a convolutional neural network super-resolution reconstruction method combining convolutional neural network and super-resolution reconstruction technology is introduced.(SRCNN).Secondly,an improved multi-scale convolutional neural network super-resolution reconstruction algorithm is proposed.The multi-scale convolutional neural network can autonomously learn the inherent logic features of high-resolution images and low-resolution images.The Set5,Set14,BSD100,and Urban100 data sets were used as experimental data and test data.The quantitative and qualitative comparisons were made with the Bicubic,SRCNN,SelfEx,VDSR,and DRCN methods with better reconstruction effects.The objective signal was selected as the peak signal-to-noise ratio(PSNR).And structural similarity(SSIM).The scaling scale factor is(×2,×3 and ×4)to verify the reliability and effectiveness of the improved method.The experimental results show that the improved method is better than the traditional reconstruction method.(3)The improved multi-scale convolutional neural network super-resolution reconstruction method(MSCNN)is combined with the GS fusion method to form a fusion model.Firstly,the multi-spectral image is super-resolution reconstructed by the improved multi-scale convolutional neural network super-resolution algorithm to obtain the image with enhanced resolution.Then,the Schmitt orthogonal transform(GS)is used to fuse the spatial information enhanced image with the full-color image to obtain a fusion result with high spatial resolution and high spectral resolution.At the same time,the contrast fusion experiment uses Brovey transform,Gram-Schmidt transform,NNDiffuse Pan Sharpening,PC Spectral sharpening and other methods.The fusion results are compared by subjective objective quality evaluation method.The experimental results show that the fusion effect of the proposed method and GS fusion model is achieved.Better than the four commonly used methods.