Multi-spectral Remote Sensing Images Super-resolution Based On Convolutional Neural Network

With the development of hardware,people have higher and higher requirements on the resolution of remote sensing data.However,in the actual application environment,limited by the cost of image acquisition equipment,including the limitations of the weight,volume,power consumption and other conditions of the remote sensing satellite or the technical bottleneck of the imaging modality itself,it is not possible to obtain large-size high-definition images every time.Remote-sensing image super-resolution reconstruction performs resolution reconstruction on low-resolution images,improves the resolution of the original image within a certain range,and obtains a clearer reconstructed image.The traditional super-resolution reconstruction method has the problems of high-frequency image loss,blurry edges,and large amount of calculation.This paper focuses on the characteristics of deep learning that can automatically extract multi-layer features and process multi-dimensional data in large quantities,and applies convolutional neural networks to Remote sensing image super-resolution reconstruction,taking into account the texture and spectral characteristics of the remote sensing image,constructed a corresponding super-resolution network.(1)Considering the texture characteristics of remote sensing images,an end-to-end deep learning model is used to super-resolution the remote sensing true color images.Input the low-resolution image and the corresponding reference image,use the improved feature extraction network to extract the features of the image block,use ResNeXt50 instead of VGG19,and compare the texture similarity between the low-resolution feature image and the reference feature image The texture is transferred in the image to construct a reconstructed image with rich texture details.Experiments show that the improved SRNTT has better visual effect and higher image quality evaluation index compared with the original method and other advanced super-resolution methods.(2)Considering the spectral characteristics of remote sensing images,using the similarity between different bands,input multi-resolution images and output high-resolution images.By using two convolutional neural networks to jointly learn the mapping from all input bands to the 10-meter super-resolution output band,combined with the asymmetric convolution in ACNet,the original 3 × 3 convolution kernels are replaced by sizes of 3 × 3,1 × 3,and 3 × 1,and then batch normalization is performed after each of the three layers to become a branch,and the output of the three branches is integrated as the output of the ACB,and the skeleton parameters are strengthened to improve the reconstruction effect,at the same time,improve the resolution of all bands,and better preserve the spectral characteristics.Experimental results proved that the asymmetric convolution improved method in this paper has better visual effect and higher image quality evaluation index than the original method.This thesis has 38 figures,8 tables and 102 references.