Image Super-resolution Restoration And Its Application In Classification

High-quality high-resolution images in today's society are in urgent need of various areas.In order to obtain high-resolution high-quality images in recent years,image super-resolution reconstruction technology has been extensively studied.Image super-resolution technique has reached a peak,and various methods have been endless.However,compared with the natural image processing,remote sensing image super-resolution technology there are still many difficulties.In particular,SAR is affected by factors such as hardware system,non-ideal platform,poor imaging conditions and system speckle noise.The SAR image resolution is low,which is difficult to meet the requirements of scientific research and application.So can the SAR image with the characteristics of SAR image to achieve super-resolution? And can the super-resolution technology promote and help other problems in the field of computer vision? Based on these two problems,this is analyzes the existing super-resolution method,and studies the super-resolution technique of SAR image.Then,the application of super-resolution method in image classification is explored.Finally,an image classification method is proposed.First,a super-resolution algorithm for SAR images based on joint optimization is implemented.Combined with the characteristics of SAR image structure,we can learn multiple dictionaries for different image blocks,and use EM algorithm to optimize multiple dictionaries by multiple iterations,so that the trained dictionaries have more accurate super-resolution of low resolution image blocks recovery.Each low-resolution image block algorithm for the test will adaptively select the most suitable dictionary for each low resolution image for super-resolution reconstruction.Finally,high resolution SAR images are obtained by aggregating high resolution image blocks.Second,the image super-resolution method is explored to promote the image classification problem.The influence of super-resolution method based on compressed sensing on image classification is explored by a series of experiments.The strategy of this paper is to super-resolution the image as a preprocessing of the image and then classify it.Three sets of experiments were made on the remote sensing data set and the optical data set.Respectively,the first group is to classify low-resolution images after downsampling.The second group is super-resolution by BICUBIC method and then classified.And the last group is based on compressed sensing super-resolution method for super-resolution and then classified.Experiments show that the super-resolution has a good effect on the classification problems,and can improve the accuracy of classification.Thirdly,based on the work of the previous chapter,a multi-scale classification algorithm based on deep self-coding is proposed.By using the super-resolution method to generate two-scale resolution images,combined with the deep learning,two deep auto-coder networks are constructed for low-resolution images and high-resolution images.By using the correlation analysis,the two networks are trained to maximize the correlation between the two networks.But only low-resolution images in the test,the training of the network according to the characteristics of low-resolution images to infer the characteristics of the corresponding high-resolution images,combined with the characteristics of two resolution images and then classification.Experiments show that our method has a good effect on multiple data sets.