Research On Sub-pixel Target Recognition Of Remote Sensing Image Based On Super Resolution Reconstruction

In the research of optical remote sensing images,spatial resolution is one of the important indicators to measure the image quality.Higher spatial resolution means more abundant data and more useful information.At present,the acquisition and application of remote sensing images have become more convenient with the continuous development of remote sensing technology.However,due to the limitation of spatial resolution on remote sensing images,low-resolution images cannot meet the needs of further processing and application of remote sensing images.Then,under the condition of the original hardware system,the use of software technology to obtain higher resolution remote sensing images has developed rapidly.In response to the topic of the thesis,this article mainly develops from the following aspects:(1)For super-resolution reconstruction based on compressed sensing,analyze the sparse representation of the signal in compressed sensing,the construction of the measurement matrix and the reconstruction algorithm of the signal,compare the advantages and disadvantages of the traditional method to achieve super-resolution reconstruction of compressed sensing,and determine the discrete wavelet Transform to sparse representation method,part of Hadamard matrix and Gaussian random matrix are measurement matrix,orthogonal matching is reconstruction algorithm for compressed sensing super-resolution reconstruction,and analysis of objective evaluation indicators.This algorithm reconstructs higher quality remote sensing than traditional algorithm image.(2)For super-resolution reconstruction based on convex set projection,the steps of constructing reference frames,motion estimation and correction of pixels to the threshold range are analyzed,and a POCS super-resolution reconstruction algorithm based on threshold correction and improved PSF fusion is proposed.It is used in worldview-3 optical remote sensing image.After analyzing the reconstructed image with objective evaluation indexes,compared with the traditional algorithm,the algorithm in this paper can effectively improve the image resolution while ensuring the image quality.It laid a solid foundation for the further use of remote sensing images.(3)After completing the related research on super-resolution reconstruction based on remote sensing images,combined with the objective evaluation criteria of MSE,PSNR,SSIM,information entropy,and average gradient of the remote sensing images,the super-resolution reconstructed remote sensing images and bi-cubic difference Values,traditional POCS algorithm,SCSR and other algorithms are compared with relevant indicators,and the relevant indicators obtained by the algorithm in this paper are superior to the traditional algorithm,and then the relevant algorithms in this paper can effectively reconstruct high-quality remote sensing images.(4)After completing the super-resolution reconstruction of the remote sensing image and the comparison between the reconstruction results of the algorithm and the reconstruction results of the traditional algorithm on the objective evaluation indicators,combined with the currently popular YOLO-based network model for DarkNet-53 target detection method,In-depth study of the algorithm reconstructed image in this paper.Select the high score No.2 data,construct a training set based on the target as an aircraft,detect the aircraft in the reconstructed high score No.2 remote sensing image,analyze the accuracy of the target detection before and after the reconstruction,and obtain the remote sensing image obtained by the reconstruction algorithm in this paper Compared with the unreconstructed low-resolution images,the accuracy of target detection is significantly improved.This shows from another aspect that the quality of the reconstructed images obtained by this algorithm is superior to the traditional algorithms.So far,the research on super-resolution reconstruction based on remote sensing images and corresponding target recognition has been basically completed in this paper.