CN-122023130-A - Optical remote sensing image super-resolution reconstruction method based on generation countermeasure network

CN122023130ACN 122023130 ACN122023130 ACN 122023130ACN-122023130-A

Abstract

The invention discloses an optical remote sensing image super-resolution reconstruction method based on a generated countermeasure network, which comprises the steps of S1, collecting a multi-source optical remote sensing data set, S2, preprocessing and storing an original image, S3, constructing a generated countermeasure network model, wherein the generated countermeasure network model comprises a generator (a convolution, batch normalization, a multi-scale feature extraction module, a dense block, a residual block and an up-sampling output reconstruction image) and a discriminator (a down-sampling and up-sampling output two classification result), optimizing through countermeasure training, S4, optimizing super-parameters (such as a learning rate and a batch size), verifying PSNR and SSIM precision, and S5, outputting the reconstruction image. And adopting block prediction stitching for large-size images. The method integrates the high-resolution series and the google image, improves the resolution and detail recovery of the remote sensing image, enhances the adaptability and generalization capability of multi-source data, obviously improves the fuzzy problem of the traditional method, and improves the interpretation precision and the application value.

Inventors

YANG ZHIGAO
LI YAN
ZHANG LEI
ZHANG QINGFANG
YANG LANYAO
Xiang Zequn

Assignees

中科卫星(安徽)数据科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251211

Claims (10)

1. The optical remote sensing image super-resolution reconstruction method based on the generation countermeasure network is characterized by comprising the following steps of: S1, collecting and integrating remote sensing images from different satellites, in different periods and under different illumination conditions, and constructing a multi-source optical remote sensing data set; s2, preprocessing a sample of the multi-source optical remote sensing data set and storing the sample as an original image; S3, constructing a remote sensing image super-resolution model based on a GAN (generation countermeasure network) architecture, wherein the remote sensing image super-resolution model comprises a generator and a discriminator, the input of the generator is the original image, the generator is responsible for generating a reconstructed image according to the input, the input of the discriminator is the reconstructed image and a real image, the discriminator is used for evaluating the difference between the reconstructed image and the real image, and the generator and the discriminator continuously optimize the output quality of the generator through a countermeasure training mechanism; S4, optimizing the super parameters of the remote sensing image super-resolution model and verifying the accuracy of the remote sensing image super-resolution model in generating the reconstructed image; S5, outputting the reconstructed image meeting the precision requirement.
2. The method of claim 1, wherein the source of the remote sensing image comprises high-resolution satellites and google images.
3. The method of claim 1, wherein the preprocessing comprises unifying the remote sensing image size, data format conversion, radiation normalization, spatial registration, and image cropping.
4. The method for reconstructing the super-resolution of the optical remote sensing image according to claim 1, wherein after the original image enters the generator, the original image is firstly subjected to preliminary feature extraction through a convolution layer, batch normalization and an activation function to obtain a feature image, then the feature image enters a main structure, the main structure comprises a multi-scale feature extraction module, a dense block and a residual block, the multi-scale module captures multi-scale information by using convolution kernels with different sizes, the dense block enhances feature propagation and reuse, the residual block helps the super-resolution model of the remote sensing image to converge faster and retain original information, then the feature image is subjected to an up-sampling stage, the feature image is gradually amplified in the up-sampling stage, the feature image is refined by adding the convolution layer after each step, and finally the reconstructed image is output through the convolution layer without the activation function.
5. The method of claim 1, wherein the arbiter performs a downsampling phase to reduce the spatial size of the image, and then the arbiter enters an upsampling phase to recover the spatial size of the image, and finally outputs a classification result, and the generator determines whether the input image is the reconstructed image or the real image.
6. The method of claim 1, wherein the super-parameters include learning rate, batch size, and loss function weight.
7. The method for reconstructing the super-resolution of the optical remote sensing image according to claim 1, wherein the optimization mode comprises searching for the optimal configuration by adopting grid search and bayesian optimization, and simultaneously evaluating the convergence speed and the stability of the model by combining the verification set performance.
8. The method for reconstructing an optical remote sensing image according to claim 1, wherein the verification of the accuracy includes calculating PSNR (peak signal to noise ratio) as follows: Wherein, the Is the maximum possible value of the pixel value, Is the mean square error, I and K represent the original image and the reconstructed image, respectively, m and n are the height and width of the original image and the reconstructed image, respectively, and I and j are the indices of pixel positions, respectively.
9. The method for reconstructing an optical remote sensing image according to claim 1, wherein the verification of the precision further comprises calculating SSIM (structural similarity), the formula is as follows: Wherein x and y represent small block areas of the original image and the reconstructed image, respectively, and the size of the small block areas is 8×8 or 11×11 pixels; 、 Respectively representing the average brightness of the small areas; 、 Is the variance of the patch area, representing the contrast; Is the covariance between two of the patch areas, used to measure the structural similarity between them; 、 Is two stability constants to avoid the case where the denominator is 0.
10. The method for reconstructing an optical remote sensing image according to claim 1, wherein a remote sensing image with a resolution of 5m×5m or more is subjected to block prediction, the remote sensing image with a resolution of 5m×5m or more is cut into an image input network with a size of 256×256 pixels or 512×512 pixels for prediction, and then the prediction results are spliced into a final result image according to the cutting order.

Description

Optical remote sensing image super-resolution reconstruction method based on generation countermeasure network Technical Field The invention relates to the technical field of remote sensing image processing, in particular to an optical remote sensing image super-resolution reconstruction method based on a generated countermeasure network. Background The main purpose of the super-resolution reconstruction of the remote sensing image is to improve the spatial resolution of the original low-resolution remote sensing image, thereby acquiring more detailed and accurate surface information, effectively improving the ground feature recognition precision, enhancing the change detection capability, improving the visual effect and supporting the fine management and decision making. The traditional super-resolution reconstruction method comprises a frequency domain method, an interpolation method, a back iteration projection method, a convex set projection method and a maximum posterior probability method, and because the algorithm structures are relatively simple, the method is difficult to have good performance when facing complex scene tasks, the super-resolution effect of the method is difficult to achieve satisfaction degree of people, the resolution is not high enough, and the contained detail information is not rich enough. As the deep learning technology is widely applied to the field of remote sensing image processing, the deep learning technology can be classified into supervised, semi-supervised and unsupervised according to a training mode, and more deep learning models are applied to remote sensing image super-resolution reconstruction. Model training is carried out through large-scale remote sensing image data by combining a deep learning model, and a better remote sensing image super-resolution reconstruction effect can be obtained generally. Disclosure of Invention Based on the analysis, the invention provides a super-resolution reconstruction method of an optical remote sensing image based on generation of an countermeasure network, which is characterized by comprising the following steps of: S1, collecting and integrating remote sensing images from different satellites, in different periods and under different illumination conditions, and constructing a multi-source optical remote sensing data set; s2, preprocessing a sample of the multi-source optical remote sensing data set and storing the sample as an original image; S3, constructing a remote sensing image super-resolution model based on a GAN (generation countermeasure network) architecture, wherein the remote sensing image super-resolution model comprises a generator and a discriminator, the input of the generator is the original image, the generator is responsible for generating a reconstructed image according to the input, the input of the discriminator is the reconstructed image and a real image, the discriminator is used for evaluating the difference between the reconstructed image and the real image, and the generator and the discriminator continuously optimize the output quality of the generator through a countermeasure training mechanism; S4, optimizing the super parameters of the remote sensing image super-resolution model and verifying the accuracy of the remote sensing image super-resolution model in generating the reconstructed image; S5, outputting the reconstructed image meeting the precision requirement. Preferably, the sources of the remote sensing images include high-resolution satellites and google images. Preferably, the preprocessing includes unifying the remote sensing image size, data format conversion, radiation normalization, spatial registration, and image cropping. Preferably, after the original image enters the generator, the primary feature extraction is performed through a convolution layer, batch normalization and an activation function to obtain a feature image, then the feature image enters a main structure, the main structure comprises a multi-scale feature extraction module, a dense block and a residual block, the multi-scale module captures multi-scale information by using convolution kernels with different sizes, the dense block enhances feature propagation and reuse, the residual block helps the super-resolution model of the remote sensing image to converge faster and retain the original information, then the feature image is subjected to an up-sampling stage, the feature image is gradually amplified in the up-sampling stage, the feature image is refined by adding the convolution layer after each step, and finally the reconstructed image is output through a convolution layer without the activation function. Preferably, the arbiter performs a downsampling phase first, during which the spatial size of the image is reduced, after which the arbiter enters an upsampling phase, through which the spatial size of the image is restored, and finally outputs a classification result, and the generator determines whether t