Research Into Pan-sharpening Methods Based On Bayesian Framework

Due to the constraints of physical conditions,general satellites will provide high-resolution panchromatic images and low-resolution multispectral images.However,in real life,it is necessary to combine high-resolution and spatial resolution images for object recognition,detection and other tasks.Pan-sharpening technology,which integrates two different qualities,has been constantly studied by many scholars.The previous Pan-sharpening methods can be classified into model driven methods and data driven methods:model driven methods such as CS method,MRA method,variation method,etc.model data based on the human experience and observation,ignoring the mining of the inherent laws of the data itself;while data-driven methods such as dictionary learning and neural network do not use explicit functions to constraint the data distribution,but through the training of the data to obtain the fitting function,it has poor generalized ability and is difficult to extend to unfamiliar data,it can be successful because of the full use of supervision information and can also fail because of the inaccurate supervision content.In this paper,we complete the Pan-sharpening task by maximum a posterior estimation based on the basic framework of Bayesian,the likelihood function and prior knowledge are set as complementary to each other,so as to adapt to the changing characteristics and avoid the influence of noise.For the likelihood function,it is used to find the invariant descriptors from the variable data.In the Pan-sharpening task,it is specifically composed of the spectral fidelity term and the structural enhancement term.The content of the spectrum is obtained by assuming that after non-linear downsampling operator the high-resolution multi-spectral image have the consistency with the low-resolution multi-spectral image.In view of the large structural distortion caused by the global structural assumption,the article construct the relationship between highresolution multispectral image and panchromatic image by using local linear regression model.According to the prior knowledge,we should try our best to capture the variability between images and even between image blocks.Although the hyperLaplacian prior can well reflect the heavy tail statistical characteristics of an image’s gradient,the differences in the properties of each local region make the generalized Gaussian function better to describe.Prior depends on the efficiency of learning,but in the absence of real high-resolution multispectral images,this paper proposes to capture the edge prior based on the existing likelihood results,and obtain the texture prior with the help of conditional random field transformation,so the prior distribution is deduced by self-supervision.A lot of experimental results show that the algorithm proposed in this paper has excellent performance and better generalization performance than the deep learning method.