Research On Super-resolution Reconstruction Method Of Optical Remote Sensing Image

High-resolution and high-quality optical remote sensing images have important application values in military and civilian fields such as remote sensing mapping,reconnaissance and surveillance,public safety monitoring and identification,and medical diagnosis.In general,the resolution of the image can be improved by means of improving the hardware performance and image processing.The hardware upgrading method is constrained by the development bottleneck of the level of technology,development difficulty,and cycle cost.Therefore,domestic and foreign scholars in image super-resolution reconstruction processing A lot of research work has been carried out.From the existing literature,the existing overreconstruction methods have achieved ideal results in the processing and application of natural images.However,there is still a certain gap between the texture detail enhancement and resolution enhancement capabilities of remote sensing images.In this paper,aiming at this problem,aiming at greatly improving the target resolution and texture details of remote sensing images,image superresolution reconstruction and evaluation methods based on single-frame and multiframe sequential images are studied and corresponding experimental verification is carried out.Main research works are shown as follows.(1)Analyze the basic principles of super-resolution reconstruction method of single frame image based on machine learning,and study three typical methods of SRCNN based on convolutional neural network,VDSR based on deep convolutional neural network,and SRGAN based on generational confrontation network,then conduct the experiment of super-resolution reconstruction of single frame image to compare and analyze the three methods.(2)Master the traditional multi-frame image super-resolution reconstruction method,based on this,consider deterioration of image quality due to remote sensing image degradation characteristics from actual imaging process,such as Motion blur,noise,deformation,combining high-precision sub-pixel motion estimation algorithms,several key technologies have been studied as follow.For obtained high-precision initial estimates,a bilateral filter interpolation algorithm is proposed which can maintain edge features well.Propose motion estimation method based on optical flow method.Introduce the way to remove abnormal frame.At last,propose improved multi-frame image super-resolution reconstruction method,and complete verification analysis by experiments.(3)Establishment of efficient evaluation system for SR algorithm has always been a difficult and challenging research area in SR research.To solve this problem,firstly,Review and analyze the current status of evaluation methods in the field of image super-resolution,based on this,propose several basic principles and methods for evaluating quality of reconstructed image.At last,suggest a method for evaluating resolution increase multiple for super-resolution reconstructed images and enhancement performance for image by algorithm.