Research On Super-Resolution Reconstruction Algorithm Of Optical Remote Sensing Image Based On Example Prior Learning

With the wide application of optical remote sensing images in military reconnaissance and national economic society,it is extremely important to obtain high-resolution optical remote sensing images with high-definition detail information for subsequent image processing,analysis and application.However,due to the resolution limitation of the optical imaging system and the image degradation factors which are difficult to avoid in the imaging,transmission and processing of the images,the resolution of the optical remote sensing images obtained by the final imaging will be reduced.Image super-resolution reconstruction technology is a method to improve the spatial resolution of the image by the correlated processing without changing the imaging system equipment.In-depth study of super-resolution reconstruction of optical remote sensing image can make full use of the existing optical remote sensing imaging system and the optical remote sensing images,thus has very important theoretical research significance and extensive application value.The problem of image super-resolution reconstruction is a highly ill-posed problem.In order to solve this problem,a prior knowledge is required to regularize the solution space to obtain the unique optimal solution.For the super-resolution reconstruction of single image,the prior information of the image can be obtained by learning from the external training examples or by using the internal self-examples.Therefore,the main work of this paper is as follows:1)In order to solve the asymmetric problem between the testing and training phases in the image super-resolution reconstruction method based on sparse representation,this paper proposes a super-resolution reconstruction algorithm for optical remote sensing image based on joint sparse mapping learning.Sparse representation provides effective prior information for image super-resolution reconstruction.The image super-resolution reconstruction algorithm based on sparse representation considers that the sparse representation of high-and low-resolution image patch over the corresponding high-and low-resolution dictionary is the same.In the training phase,a large number of external training image examples are used to jointly learn high-and low-resolution dictionaries,and in the testing phase,the sparse representations of low-resolution image patches are obtained over the learned lowresolution dictionary.However,there is no guarantee that the sparse representation of the low-resolution image patch is consistent with that of the corresponding high-resolution image patch which is unknown to be solved.Therefore,the solution is not optimal.In order to improve this problem,the proposed algorithm takes advantage of feed-forward neural network and combines the sparse representation learning and dual dictionaries learning into a framework for jointly learning in the image super-resolution reconstruction.The training process is optimized by using auxiliary coordinates to obtain the corresponding parameters,which can be applied directly to obtain the super-resolution reconstruction results.Experimental results show that the proposed algorithm is not only visually superior to other image super-resolution reconstruction algorithms based on sparse representation learning,but also has better objective evaluation indexes.Moreover,the fast testing speed can nearly realize real-time processing.2)From the perspective of image restoration,this paper presents a super-resolution reconstruction algorithm for optical remote sensing image based on non-local self-similarity prior learning,which takes advantage of the image prior model to regularize the solution space to solve the image super-resolution reconstruction problem.Generally,an image patch can be considered as the samples of multivariate vectors,and the image is non-Gaussian.Therefore,the image patch can be modeled with Gaussian mixture model,which has been successfully used in various image restoration problems.According to the statistical characteristics of the image,the image contains a large number of non-local image patches redundancy.But in most of the image restoration methods based on non-local self-similarity,the similar patches are searched from the degraded image,which will have a bad effect on the results of image processing.Therefore,the proposed algorithm learns the prior model by taking advantage of Gaussian mixture model and non-local self-similarity from a set of high-resolution and high-quality images.In the testing phase,each image patch is searched for its non-local self-similarity patches for an image to form an image patch group.Then,the best matching Gaussian component is selected from the prior model obtained in training phase based on maximum posterior probability.And the component can provide the dictionary and regularization parameters for image super-resolution reconstruction.Experimental results show that the prior information based on image patch group learning from high-resolution training images can contribute to obtain better visual effects and objective evaluation indexes for image super-resolution reconstruction.3)Deep network learning can obtain more sophisticated and high-level features from image data through multi-layer network.Inspired by the relationship between deep learning and dictionary learning,this paper proposes a novel super-resolution reconstruction algorithm for optical remote sensing image based on multi-layer analytic-synthesis dictionary learning,which takes advantage of deep dictionary learning to represent the image more effectively.By using the framework of the analytic-synthesis dictionary model and multi-layer sparse model,the nonlinear mapping relationship between low-and high-resolution examples which are extracted from a large number of external training images is learned directly,in which multi-layer analytic dictionaries of different sizes are used to extract higher-level features from low-resolution image examples and a synthesis dictionary in the last layer is used to optimize the regression process.Similar to back propagation in deep learning,the dictionaries are hierarchically updated based on the error minimization using the backward projection method to obtain the optimal multi-layer dictionaries.In the testing phase,the trained multi-layer analytic-synthesis dictionaries are directly applied to the testing lowresolution image for super-resolution reconstruction,which could achieve better reconstruction results with both visual and objective evaluation indexes.4)For optical remote sensing images with small samples,this paper proposes a superresolution reconstruction algorithm for optical remote sensing image based on structuralcorrelated self-examples,which does not rely on any external training image database and make any assumptions about the image space.Because of the data redundancy in the image,similar image patches can be found for each patch from the image itself and different scales of the image.These non-local data redundancy can also provide powerful image prior knowledge for image processing.In this paper,the input low-resolution image is used to perform down-sampling with a small scale and corresponding up-sampling operations to build double image pyramids with high-and low-resolution image pairs,from which the internal training examples are extracted.The proposed algorithm is achieved by taking advantage of sparse representation to replace nearest neighbor search to find the structural-correlated patches with the testing patch,and learns the direct mapping relation between high-and lowresolution structural-correlated image patches.The super-resolution reconstruction process is also realized layer by layer with a small scale.Experimental results show that the superresolution reconstruction algorithm based on direct mapping of structural-correlated image patches is superior to other methods based on internal examples learning,especially for point target and line edge of the target.