| Optical remote sensing images are one of the important carriers for acquiring and reflecting actual surface information.The spatial,spectral and temporal resolution characteristics of optical remote sensing images are important indicators to measure their image quality and application potential.On the one hand,due to the shortage of imaging equipment and the limitation of imaging conditions,remote sensing images will be affected by different degradation factors during the acquisition process,resulting in the inevitable distortion or blurring of the obtained images;on the other hand,due to the limitation of the hardware design of satellite sensors,the spatial,spectral and temporal resolutions of single-source remote sensing images are mutually constrained and cannot have both.In order to eliminate the adverse effects caused by degradation factors and overcome the resolution constraints of single-source remote sensing images,on the basis of the existing imaging system conditions and current observation images,super-resolution reconstruction technology combines the prior knowledge of remote sensing images and digital image processing technology and attempts to reconstruct high-resolution remote sensing images with clearer texture information and richer spectral bands from one or more low-resolution remote sensing images,so as to provide higher quality data resources for the wide application of remote sensing images.Therefore,in this sense,super-resolution reconstruction technology is still one of the promising research directions in the field of image restoration.Compared with natural images,remote sensing images have more abundant spectral bands and more complex features of ground objects.If the previous mature super-resolution reconstruction methods of natural images are directly applied to the reconstruction tasks of remote sensing images,the reconstruction effect will often be poor and low efficiency.In view of this,this paper conducts research on super-resolution reconstruction methods for improving the quality of optical remote sensing images,mainly focusing on the super-resolution reconstruction model and algorithm development of a single remote sensing image based on the convolution sparse representation theory,and the super-resolution reconstruction model and algorithm development of hyperspectral remote sensing images based on sparse low-rank representation theory.The main research work of this paper is as follows:1、A new convolutional sparse representation theory is introduced to study the super-resolution reconstruction method of a single remote sensing image.The model framework for the traditional sparse representation methods has been developed relatively maturely,and the scalability of the algorithm is limited.In most studies,the evaluation of the overall performance of the super-resolution algorithm framework is only based on individual traditional super-resolution reconstruction methods,which has some problems,such as one-sidedness and deficiency.This paper systematically analyzes the research background and existing difficulties of the current optical remote sensing image quality improvement,and comprehensively summarizes and generalize the development process,research status,model advantages and disadvantages of two post-processing technologies to improve the optical remote sensing image quality,namely,the super-resolution reconstruction technology of single remote sensing image and the spatial-spectral fusion technology of multispectral image and hyperspectral image.Considering that the convolutional sparse representation method has shown strong reconstruction performance in the super-resolution task of natural images,this paper compares and expounds the theoretical framework of the traditional sparse representation and the theoretical framework of the convolutional sparse representation in detail and discusses the mathematical principle and model structure of these two signal sparse representation theories in the super-resolution reconstruction task of remote sensing images.The dictionary forms and dictionary learning methods of the two sparse representation theories are discussed,the boundary conditions of the model solutions for solving the ill-conditioned inverse problem in the super-resolution reconstruction of remote sensing images under the two sparse representation theories are described,and their commonly used iterative optimization algorithms are also introduced in detail,which is helpful to understand the iterative essence and implementation form of two sparse representation theories in the task of super-resolution reconstruction more clearly.2、A hybrid non-local self-similarity constrained convolutional sparse coding method for single remote sensing image super-resolution reconstruction is proposed.Aiming at the traditional super-resolution reconstruction method of single remote sensing image based on image patches,the consistency between overlapping patches is often ignored,resulting in poor reconstruction effect or block effect.In this paper,the input remote sensing image is extracted as a smooth component and texture components.The smooth component and texture component of the image are reconstructed by the non-parametric Bayesian method and the non-local self-similarity constrained convolutional sparse coding method,respectively.(1)In the reconstruction of smooth components,the non-parametric Bayesian method has the advantage of using more priori information of the underlying structure of the image itself to perform non-parametric derivation of model parameters through probability sampling,which can more fully restore the residual high-frequency information in the smooth components.(2)In the reconstruction of texture components,a non-local self-similarity constrained convolutional sparse coding method is proposed.Its algorithm idea is to combine the non-local self-similarity structure of each feature map and convert the reconstruction problem of texture components into an iterative threshold optimization problem that minimizes the convolutional sparse coding noise of the feature map,so as to obtain the feature map that is closer to the real value,and finally reconstructs the texture components through the convolution operation.In addition,in the process of searching the non-local self-similar structure of the feature map,the correlation coefficient is used as the structure information to classify the image patches in the search space,avoiding unnecessary weight calculation.The experimental results show that the texture and edge structure of the reconstructed image obtained by this method are clearer and have good anti-noise performance.3、A multi-scale semi-coupled convolution sparse coding method for single remote sensing image super-resolution reconstruction is proposed.Aiming at the traditional convolutional sparse coding super-resolution reconstruction method only introduce the linear projection relationship in the feature space conversion and fail to consider the local detail information in the learning of the feature maps,resulting in the unsatisfactory edge and detail of the reconstructed image and so on.(1)This paper firstly uses smooth gradient to decompose the input image at multiple scales and extracts its smooth components and texture components at multiple scales.The final smooth component is reconstructed by bicubic interpolation,and semi-coupled convolutional sparse coding method is performed on the texture components at each scale;(2)On this basis,a nonlinear convolution operator is used as the projection function between the high-resolution feature map and the low-resolution feature map of texture components at each scale,and a non-local self-similarity structure is introduced into the feature map learning for constraint optimization,which can better reconstruct the texture components at each scale.(3)Finally,the reconstructed smooth components and the reconstructed texture components at each scale are superimposed to obtain the final reconstructed image.The experimental results show that the reconstructed image obtained by this method can recover more detailed information,and the nonlinear convolution operator improves the flexibility of the model and makes the reconstruction effect better.The method also has certain anti-noise performance.4、A super-resolution reconstruction method of hyperspectral remote sensing images combining spatial-spectral double-dictionary optimization and structural sparse low-rank representation constraints is proposed.Aiming at the problem that most of the existing super-resolution reconstruction methods of hyperspectral remote sensing images only consider a single spectral dictionary or spatial dictionary,which can only reflect the unilateral features of the potential hyperspectral remote sensing images,this paper combines the spectral global correlation and spatial non-local similarity of hyperspectral images,and innovative researches are carried out from two levels of spectral domain and spatial domain:(1)In the spectral domain,a spectral dictionary characterizing generalized spectral features is learned from input low-resolution hyperspectral images,and a shape-adaptive superpixel strategy is used to over-segment multispectral images with high spatial resolution,and a local structure low-rank prior regularization term is applied to the sparse coefficients of the spectral pixels within the same superpixel,so as to establish a fusion model of the spectral domain.(2)In the spatial domain,this paper uses the residual high-frequency information that cannot be represented by the spectral dictionary to perform spatial dictionary learning and applies the regularization term composed of the structured sparse constraints to the low-rank subspace of the spatial sparse coefficients,so as to establish a fusion model of the spatial domain.The method in this paper formulates two fusion models in the spectral domain and the spatial domain as an iterative optimization problem of variables and solves these fusion variables efficiently by the Alternating Direction Methods of Multipliers(ADMM).The experimental results show that the potential hyperspectral remote sensing images obtained by this method have more precise spatial details and less spectral distortion and have efficient fusion performance. |
PDF Full Text Request