Hyperspectral real downsampling fuzzy kernel estimation method

Abstract

The invention belongs to the technical field of image processing, and discloses a hyperspectral real downsampling fuzzy kernel estimation method, which comprises the steps of performing four-time downsampling on a hyperspectral image to be processed; traversing and calculating the spatial spectrum similarity of the image blocks in the obtained hyperspectral images with the two scales by using a search box with the same size; adjusting the traversal step length according to the similarity to complete search; constructing a hyperspectral down-sampling real fuzzy kernel generator and a discriminator; after the cross training of the generator and the discriminator is completed, simulating and generating a low-resolution image of the hyperspectral image which is sampled under the real kernel by the generator. According to the method, the real down-sampling blurring kernel of the hyperspectral image can be simulated, and a real hyperspectral image super-division or fusion data set is generated to break through the bottleneck problem of a data source; a training set for sampling blurring kernel estimation under a real hyperspectrum can be extracted from common hyperspectral data by utilizing co-occurrence in a real image.

Description

technical field [0001] The invention belongs to the technical field of image processing, and in particular relates to a hyperspectral real downsampling blur kernel estimation method. Background technique [0002] At present, compared with ordinary RGB images, hyperspectral images contain rich spectral information, but due to the limitations of radiation transmission process and sensor technology, the spatial resolution of hyperspectral images taken is low. The low spatial resolution cannot obtain the details of ground objects well, which greatly limits the performance of subsequent ground object classification and target detection of hyperspectral images. Therefore, improving the spatial resolution of hyperspectral images is a necessary prerequisite for improving the subsequent processing performance of remote sensing images. [0003] At present, the fusion of hyperspectral images with panchromatic or RGB images is a common method to improve the spatial resolution of hypers...

AI Technical Summary

Problems solved by technology

[0006] (1) For classical algorithms, most of them cannot satisfy the spatial and spectral balance at the same time, resulting in information loss

[0007] (2) For the existing deep learning methods, most of the training and testing are completed under the ideal downsampling data set generated according to the wald's protocol. This kind of fuzzy kernel is difficult to accurately simulate the degradation process

[0008] (3) The problem of super-resolution of real images based on deep learning is how to introduce an accurate degraded model to ensure that the downsampled degraded image has the same domain properties as the original image

If two resolution images of the same target area are obtained directly through the remote sensor, in order to reduce the influence of the external environment, the two sensors need to be shot at a similar time, and the subsequent pixel alignment of the image needs to be done, which will consume a lot of time resource

Images