Research On Remote Sensing Image Fusion Method Based On Structural Group Sparse Representation

Remote sensing image fusion adopts the different characteristics with multispectral remote sensing image and panchromatic remote sensing image. Image fusion in remote sensing gets a fused picture which has both the spectral characteristics of the multispectral image and the detail information of the panchromatic image.The fused picturecan be used better for remote sensing image classification, feature extraction and change detection, etc.In recent years, remote sensing image fusion based on sparse representation has achieved dramatic results.However, the classical sparse representation does not consider the similarity between image patches and patches, which causesperformance degradation in solving sparse coefficients and high computational complexity. To improve the accuracy and speed of sparse representation of remote sensing image fusion, this paper carried out exploratory research on the remote sensing image fusion algorithm based on the structural group sparse representation.The main researchcontents of this paper include:1 The current situation and development trend of the remote sensing image fusion was analyzed.Analysis of several commonly used remote sensing image fusion methods, including PCA transform, HPF transformation, GIHS, AWLP etc, was carried on. This chapter introduces the evaluation index of remote sensing image fusion, and had carried on the experimental analysis and comparison of traditional fusion method.These studies laid a solid foundation for the research of algorithm in this paper.2?A method of remote sensing image fusion based structural group sparse representation(SGSR)was proposed in this paper. Firstly, adaptive dictionary and group sparse coefficient of custom intensity component image and panchromatic image are calculated respectively via group sparse representation algorithm by composing of nonlocal patches with similar structures. Secondly,, using the absolute maximum fusion rule, substitution panchromatic image sparse coefficient can be obtained, and the high spatial resolution intensity image is reconstructed using the panchromatic image group dictionary and new sparse coefficients. Finally,The high resolution multispectral image is obtained via the general component substitution(GCOS) framework. The experimental results prove that theproposed method can improve the spatial resolution of the fused image, at the same time can better maintain the original multispectral image spectrum characteristics.Compared with the classical sparse representation algorithm(SR), the proposed SGSR algorithm has greatly reduced the time complexity of the dictionary learning.In classical sparse representationalgorithm,each image patches the basic unit ofsparse coding, failed to consider the similarity between image patch and patch.The SGSR algorithm throughthe SVD decomposition can get each structural group adaptive dictionaryand change the way of obtaining training dictionaryby multiple iterative optimization, greatly improving the calculation efficiency. By comparing the speed of training dictionary, SGSR is about ten times faster than SR algorithm. Moreover, atoms in adaptivedictionary are orthogonal to each other, more conducive to the stable and accurate for sparse coding.3?The remote sensing image fusion algorithm combining wavelet transform with structure group sparse representation(WTSGSR)is proposed. For improving thespectral informationkeeping ability of the SRSR algorithm, and wavelet transform has spectral information keeping ability. This algorithm uses wavelet transform to get the high-frequency andlow-frequency dataof luminance component and panchromatic images,and the low frequency datais the approximation of the original image, which wasanew low frequency imagereconstructed using the absolute maximum fusion rule get new sparse coefficient after according to the structure group sparse model to obtain the adaptive dictionary group and group sparse coefficient.The high-frequency data reflects thesignificant information of the original images, and most of its coefficients close or equal zero, which can be seen as “sparse” and can be directly fused. Through the experimental analysis ofthreedifferent types of satellite data, the algorithm can enhance the spatial resolution while the spectral characteristics can be retainedmaximumfrom raw multispectral remote sensing image.