Super-resolution reconstruction method for large-breadth optical remote sensing image

Abstract

The invention discloses a large-breadth optical remote sensing image super-resolution reconstruction method, and the method comprises the steps: cutting a large-breadth remote sensing image through employing a sliding window cutting strategy based on the overlapping of adjacent image blocks, and taking the image blocks obtained through cutting as the input of a super-resolution network; extracting prior information of a low-resolution image based on an E-SRADSGAN model of generative adversarial learning, and constructing a neural network model for super-resolution reconstruction; and after the reconstruction is completed, splicing the reconstructed image blocks by using an effective region splicing strategy to form a complete large-breadth remote sensing image. According to the method, on the basis of a sliding window cutting strategy for overlapping adjacent image blocks and a reconstruction image effective area splicing strategy, super-resolution reconstruction of a large-breadth visible light remote sensing image is achieved, and limitation of hardware computing power and a network design mode is effectively broken through.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and remote sensing image super-resolution reconstruction, and specifically to a method for super-resolution reconstruction of large-width optical remote sensing images. Background Art

[0002] Large-width remote sensing images have important military and civilian values due to their wide coverage and large amount of information, and play an irreplaceable role in fields such as target detection, ground object classification, natural disaster monitoring, and environmental change assessment. In the military field, large-width images can provide dynamic monitoring support for large scenes, such as monitoring the activities of targets such as aircraft takeoff and missile silo opening. Such dynamic targets usually only occupy a dozen to dozens of pixels in large-width images, and their information is prone to be blurred in low-resolution images, thus increasing the difficulty of recognition and analysis. In addition, large-width images are obtained through satellite imaging. Restricted by the resolution of imaging devices and external interference factors, the generated image quality often has problems such as blurred contours and insufficient texture details, severely restricting the precise capture of key military targets and the effective monitoring of dynamic changes.

[0003] To address the above problems, the super-resolution reconstruction technology for large-width remote sensing images has gradually become the focus of research. This technology can enhance texture details while improving the image resolution, and improve the ability to capture small targets and high-frequency information. However, large-width images usually contain billions of pixels. Limited by the existing hardware computing power and the computational complexity of super-resolution algorithms, traditional processing methods are difficult to directly reconstruct the entire image. In addition, since conventional super-resolution reconstruction algorithms need to perform block processing on large-width images, the edge regions of block stitching are prone to information loss due to the lack of sufficient adjacent pixels, further affecting the overall reconstruction effect of the image.

[0004] Therefore, designing an efficient super-resolution reconstruction method suitable for large-width remote sensing images, which can not only improve the reconstruction quality of low-resolution images, but also solve problems such as edge information loss, and maximize the application potential of large-width images in military and civilian fields, is an important research direction in the current technical field. Summary of the Invention

[0005] Aiming at the problems in the prior art, the purpose of the present invention is to provide a method for super-resolution reconstruction of large-width visible light remote sensing images. This method uses a hybrid processing method of the E-SRADSGAN super-resolution reconstruction model based on the generative adversarial network and the cropping and stitching strategy of large-width visible light remote sensing images for different tasks, making the edge details more prominent, thereby improving the quality of super-resolution reconstructed images and solving the problem that it is difficult for traditional methods to capture high-frequency information.

[0006] The present application achieves the above effects through the following technical solutions: A method for super-resolution reconstruction of a wide-field optical remote sensing image, the method comprising the following steps:

[0007] S1 Collect a wide-field visible light remote sensing image, and use a degradation link based on blur kernel and noise estimation to downsample the collected image, simulate the degradation process of a real remote sensing image to obtain a low-resolution image, and then use an edge information enhancement auxiliary network to enhance the edge information of the low-resolution image;

[0008] S2 Use a sliding window cropping strategy based on adjacent image block overlap to crop the wide-field visible light remote sensing image to obtain cropped low-resolution image blocks;

[0009] S3 Use the trained E-SRADSGAN model to perform super-resolution reconstruction on the cropped low-resolution remote sensing image blocks to obtain reconstructed image blocks;

[0010] S4 Use a reconstructed image block effective area stitching strategy to stitch the reconstructed image blocks, and finally perform edge processing to obtain a complete super-resolution reconstructed wide-field visible light remote sensing image.

[0011] Further, the S2 is specifically:

[0012] S11, obtain a real low-resolution wide-field visible light remote sensing image I real ;

[0013] S12, downsample the low-resolution wide-field visible light remote sensing image I real to obtain a high-resolution image I HR ;

[0014] S13, input the resolution wide-field visible light remote sensing image I real into a preset blur kernel extraction model to obtain blur kernel information k;

[0015] S14, set a variance threshold to crop noise blocks from the flat scene of the visible light remote sensing image I real to collect noise samples to obtain degraded noise n;

[0016] S15, after obtaining the blur kernel k and the noise n, convolve the blur kernel with the high-resolution image I HR and perform downsampling, and then superimpose the noise n to generate a low-resolution image I LR ;

[0017] S16, input the low-resolution image I LR into a preset edge information auxiliary enhancement network to obtain a low-resolution image I E-LR with enhanced edge features;

[0018] Further, S2 specifically includes: obtaining a wide-field-of-view visible light low-resolution remote sensing image I with enhanced edge features E-LR , and cropping the wide-field-of-view visible light remote sensing image based on a sliding window cropping strategy with overlapping adjacent image patches to obtain cropped low-resolution image patches.

[0019] Even further, S2 specifically includes:

[0020] S21, obtaining the height and width information H×W of the wide-field-of-view visible light low-resolution remote sensing image I with enhanced edge features E-LR ;

[0021] S22 then initializes new regions by padding P pixels in four directions of I E-LR according to the height and width information of I; E-LR

[0022] S23, calculating the cropping step size of the sliding window according to the height and width information and the padded pixel P;

[0023] S24, placing I E-LR at the center of the expanded new region to complete image preprocessing and obtain a new image I E-LR ′ to be cropped;

[0024] S25, cropping I E-LR ′ using the sliding window according to the cropping step size, and saving the cropped image patches in a unified format.

[0025] Further, S3 specifically includes: sequentially taking all the image patches as inputs to the super-resolution reconstruction network E-SRADSGAN according to the cropping order, and saving the super-resolution reconstructed images in a specified folder.

[0026] Further, S4 specifically includes: obtaining the super-resolution reconstructed image patches from the specified folder, and stitching all the image patches according to the stitching strategy of the effective regions of the reconstructed image patches to achieve super-resolution reconstruction of the wide-field-of-view visible light remote sensing image.

[0027] Even further, S4 specifically includes:

[0028] S41, reading the super-resolution reconstructed graphic patches from the specified folder;

[0029] S42, stitching the super-resolution reconstructed image patches according to the cropping order, discarding the padded regions around the image patches, and only taking the middle effective regions for stitching;

[0030] S43. Remove the extended outer frame area from the above-mentioned stitched large-width visible light remote sensing image to obtain the super-resolution reconstruction result of the complete large-width visible light remote sensing image.

[0031] The beneficial effects of the present invention are as follows:

[0032] 1. Before the super-resolution generator network processes, the high-frequency information in the low-resolution image is enhanced in advance through the edge information assistance network. At the same time, an edge attention mechanism is introduced into the super-resolution network, combined with the edge information loss function. Through the above improvements, the edge details are more prominent, thereby improving the quality of the super-resolution reconstructed image and solving the problem that it is difficult for traditional methods to capture high-frequency information.

[0033] 2. A sliding window cropping strategy based on adjacent image block overlap and a stitching strategy for the effective area of the reconstructed image blocks are designed, which solves the problems of insufficient edge pixel information during convolution processing and false edge information and edge blurring caused by filling strategies (such as zero filling or copying edge filling), realizes the efficient super-resolution reconstruction of the entire large-width remote sensing image, and breaks through the limitation that traditional methods can only process small-size images.

[0034] 3. Generate low-resolution visible light remote sensing images based on the real degradation link, so that the super-resolution reconstruction network can more accurately learn the image restoration process; comprehensively apply the edge information enhancement method, the cropping and stitching strategy and the E-SRADSGAN network to construct a super-resolution reconstruction method suitable for large-width visible light remote sensing images, effectively improving the reconstruction efficiency and accuracy of large-width remote sensing images. Description of the Drawings

[0035] Figure 1 is the large-width visible light remote sensing image;

[0036] Figure 2 is the structural schematic diagram of the blur kernel extraction model;

[0037] Figure 3 is the schematic diagram of the super-resolution reconstruction strategy for large-width visible light remote sensing images;

[0038] Figure 4 is the overall scheme process demonstration diagram;

[0039] Figure 5 is the comparison diagram of the results before and after the improvement of the cropping and stitching strategy;

[0040] Figure 6 is the gray value curve diagram of randomly extracting 100 pixel points from each of the traditional stitching result image and the improved strategy stitching result image. Detailed Embodiments

[0041] For the convenience of those skilled in the art, the present invention will be further described below in conjunction with embodiments and drawings. The content mentioned in the embodiments does not limit the present invention.

[0042] A method for super-resolution reconstruction of large-width optical remote sensing images in this application is implemented based on three modeling methods: the E-SRADSGAN large model, the KernelGAN small model, and the EESN small model. The present invention makes improvements, expansions, splits, recombinations, and reuses on the basis of these three technologies. There are many common points and many different improvement measures in the implementation of the specific technical solutions compared with the original technology. It is specifically stated hereby. Through a sliding window cropping strategy based on overlapping adjacent image patches, preprocessing before super-resolution reconstruction is achieved, and then the super-resolution reconstruction network performs super-resolution reconstruction on each image patch one by one according to the cropping order; subsequently, an effective region stitching strategy for the reconstructed image patches is designed to stitch the reconstructed image patches, effectively reducing the loss of edge pixels of the image patches caused by super-resolution reconstruction. Finally, the extended edges are removed to complete the super-resolution reconstruction of the entire large-width visible light remote sensing image.

[0043] First, a brief introduction to the E-SRADSGAN large model, the KernelGAN small model, and the EESN small model is given:

[0044] The E-SRADSGAN large model. The original paper was by Fanen Meng 1 ; Sensen Wu 1 ; Yadong Li 2 ; Zhe Zhang 1 ; Tian Feng 3 ; Renyi Liu 1 ; Zhenhong Du 1 . Single Remote Sensing Image Super-Resolution via a Generative Adversarial Network With Stratified Dense Sampling and Chain Training[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, Vol.62:1-22. The model structure can be seen in Figure 4。The input of this large model is a high-resolution remote sensing image of 216×216, and the output is an image of 216×216 after super-resolution reconstruction. Its drawback is that bicubic interpolation downsampling is used to construct the low-resolution image for each image, which does not conform to the degradation link of real remote sensing images, and the model is limited to image processing with a size of 216×216. The present invention reuses its model structure but not the weights, and trains new weights for the super-resolution reconstruction of large-width visible light in a new training set. Since the model is composed of multiple modules inside, the model can be split into multiple modules and used separately. The main modification is the way of obtaining image pairs; especially the SRADSGAN part is used, and an edge attention mechanism and a boundary loss function are added to this part.

[0045] The KernelGAN small model. The original project address is: https: / / github.com / sefibk / KernelGAN, hereinafter referred to as KernelGAN for short. The KernelGAN model is usually used to estimate the image-specific blur kernel. The present invention applies its model structure, without prior data, and does not reuse its training data, and trains in a new low-resolution image training set to generate a downsampled image with a real blur kernel.

[0046] The EESN small model. The original project address is: https: / / github.com / Jakaria08 / EESRGAN, hereinafter referred to as EESRGAN for short. The EESRGAN model is used for small target detection in remote sensing images. The present invention reuses its model weights. Since the model is composed of multiple modules inside, the model can be split into multiple modules and used separately. The main part is to specifically use its EESN part alone and use the intermediate result as an additional calculation basis. Intermediate result 1: The image obtains an edge-enhanced image after passing through EESN.

[0047] Example 1:

[0048] A method for super-resolution reconstruction of large-width remote sensing images provided by this application, and the overall implementation path includes the following steps:

[0049] S1 Use the degradation link based on KernelGAN to process the collected large-width visible light images, simulate the degradation process of real remote sensing images to obtain low-resolution images, and then use the EESN network to enhance the edge information of the low-resolution images;

[0050] S2 Use a sliding window cropping strategy based on adjacent image block overlap to crop the large-width visible light remote sensing image to obtain cropped low-resolution image blocks;

[0051] S3 uses the trained E-SRADSGAN model to perform super-resolution reconstruction on the cropped low-resolution remote sensing image patches to obtain the reconstructed image patches;

[0052] S4 uses the stitching strategy for the effective regions of the reconstructed image patches to stitch the reconstructed image patches, and finally performs edge processing to obtain a complete super-resolution reconstructed large-width visible light remote sensing image.

[0053] In the first aspect, step S1 of the above solution provides a method for constructing a low-resolution remote sensing image and image enhancement based on a real degradation link using a KernelGAN small model and an EESN small model. KernelGAN refers to a deep learning model specifically used to process low-resolution images. It can perform high-precision super-resolution reconstruction on low-resolution images without knowing the downsampling kernel. The input of KernelGAN is a low-resolution image patch (patch), and its data form is usually a two-dimensional pixel matrix, representing the pixel distribution of the image patch. The core idea of KernelGAN is that the model learns the image-specific blur kernel (SR kernel) in the internal structure of the low-resolution image without high-resolution images or any prior downsampling kernel data. The model simulates the cross-scale repetitive characteristics inside the image through generative adversarial training (GAN), and the generator is trained to generate a blurred downsampled image that matches the input image patch. The key to the KernelGAN method lies in the model's ability to learn to extract features and match image patches at different scales. Users only need to input a low-resolution image, and the model can automatically output a suitable blur kernel, thus helping existing super-resolution methods achieve higher-quality image reconstruction.

[0054] The said S1 includes the following content:

[0055] The user needs to provide a low-resolution remote sensing image sample I real , and load it as the input image into KernelGAN for blur kernel estimation. First, I realCrop it into several image patches (patches), and these image patches are input into the KernelGAN model as training data. Inside the model, the Generator will generate a new set of downsampled image patches to make it as close as possible to the distribution of the original image patches, while the Discriminator optimizes the output of the Generator by distinguishing between real and generated image patches, thus learning the true blur kernel of the image. After the Generator is trained, it can output downsampled image patches with a specific blur kernel. By performing layer-by-layer convolutional operations on the layers of the Generator, the estimated image blur kernel can be finally obtained. Assume that the size of the blur kernel is k×k, then the data structure of this blur kernel is [k,k], which can be used to generate low-resolution images with real degradation effects. Then, set the variance threshold from the visible light remote sensing image I real Crop noise patches from the flat scene of to collect noise samples, and obtain the degraded noise n; after obtaining the blur kernel k and the noise n, convolve the blur kernel with the high-resolution image I HR , perform downsampling (the downsampling scale is s), and then superimpose the noise n to generate the low-resolution image I LR . The complete mathematical expression is as follows:

[0056]

[0057] Next, apply the generated blur kernel to other low-resolution images. For example, perform a convolution operation on the generated blur kernel and another high-resolution remote sensing image Image_HR to generate a downsampled image with the same blur kernel effect, and complete the construction of the high-low resolution image pair. Then input the low-resolution image into the EESN network, and the output is the low-resolution wide-swath remote sensing image I E-LR after edge feature enhancement.

[0058] In the second aspect, step S2 of the above solution uses a sliding window cropping strategy based on adjacent image patch overlap to crop the wide-swath visible light remote sensing image to obtain the cropped low-resolution image patches.

[0059] S2 includes the following content:

[0060] First, set the initial size of the wide-swath remote sensing image I E-LR as H×W, where H is the image height and W is the image width. When cropping the wide-swath remote sensing image, use a sliding window with a sliding window size of SIZE h ×SIZE w (by default, 320×320 in the present invention), and the padding width is P (by default, 32 pixels in the present invention).

[0061] To ensure that no edge pixels of the image block are lost during subsequent splicing, during the cropping process, it is necessary to expand the length and width of the original image I E-LR The mathematical expression for the size of the expanded region O is as follows:

[0062]

[0063] where H′ and W′ are the height and width of region O respectively; [·] represents the ceiling operation.

[0064] The cropping process uses a sliding window for operation. Let I E-LR ′ be the new image after I E-LR is expanded within region O. The width and height of I E-LR ′ are W′ and H′ respectively. To ensure that the sliding window can fully traverse the expanded image I E-LR ′ during the cropping process and ensure that the cropped image blocks have an expanded edge with a width of P in both the length and width directions, it is necessary to set the cropping step size to S h and S w , where S h represents the cropping step size in the horizontal direction, and S w represents the cropping step size in the vertical direction. The mathematical expression for the cropping step size is as follows:

[0065] S h =SIZE h -2P, S W =SIZE w -2P.

[0066] Subsequently, perform the zero-value initialization operation on region O. The mathematical expression for the above operation process is as follows:

[0067] I E-LR ′=zeros(H′,W′)

[0068] where zeros(·) represents the zero-value initialization operation, that is, creating a matrix composed of zero values. Obtained according to the above formula, and place the original image I E-LR in the central region of region O as the initial image to be cropped. The mathematical expression for the above operation process is as follows:

[0069]

[0070] where i and j represent pixel indices. When cropping on the expanded image, the sliding window starts from the (0,0) coordinate and slides with S h and S w as the step sizes. The upper-left coordinate of the cropping sliding window is (i,j), and its mathematical expression is as follows:

[0071] (i,j) = (m·S h , n·S w );

[0072] Wherein, m and n respectively represent the indices of the cropped sliding window in the y and x directions.

[0073] The above cropping method can ensure that there is an overlapping area O of 2P pixels between adjacent image patches obtained after cropping the sliding window, overlap thus avoiding the loss of image edge information.

[0074] In the third aspect, step S3 of the above solution is to perform super-resolution reconstruction on the cropped low-resolution remote sensing image patches using the trained E-SRADSGAN model, with the aim of obtaining the image patches after super-resolution reconstruction. First, use the generator SRDSRAN of the improved E-SRADSGAN to extract features from the low-resolution image patches. An edge attention mechanism is added to the generator SRDSRAN to improve the edge quality. The generator combines the dense sampling mechanism with the residual network and the edge attention module to extract the shallow and deep features of the image layer by layer, fuse and enhance them, and perform weighted processing on the edge features through the attention module to make the edge information more prominent and improve the detail clarity of the super-resolution reconstruction. Subsequently, during the model training process, the loss function combines pixel loss (L1 loss), perceptual loss (extracting high-dimensional features using a pre-trained VGG network), adversarial loss (feedback through the discriminator), and edge information loss function (strengthening the constraint on the high-frequency details of the image through the Sobel edge detection operator), and optimizes the image reconstruction quality from multiple perspectives. These loss functions are controlled by weight coefficients to ensure that the generated high-resolution image has a natural visual effect while restoring details.

[0075] At this time, the world coordinate range can be customized to reconstruct the remote sensing image slices again. The spliced image is generated into a GeoTIFF file according to the geographic coordinate information, including the reconstructed RGB image and the corresponding geographic reference information. This enables the image to be directly used for further geographic information analysis and classification tasks, such as scene classification, object detection, etc.

[0076] In the fourth aspect, step S4 of the above solution is to splice the reconstructed image patches using the effective area splicing strategy of the reconstructed image patches, and finally perform edge processing to obtain the complete large-width visible light remote sensing image after super-resolution reconstruction.

[0077] The S4 includes the following content:

[0078] First, set the actual splicing area of each image patch after super-resolution reconstruction as O effective , O effective The area range is shown in the following formula:

[0079] O effective = SIZE h′ × SIZE w′ ;

[0080] Wherein, SIZE h′ and SIZE w′ respectively represent the height and width of the actual stitching area before super-resolution reconstruction (by default in the present invention, SIZE h′ = SIZE w′ = 320 - 2P). Assume that the magnification of the super-resolution reconstruction task is scale, and the actual stitching area is O SR . After super-resolution reconstruction, the area range of O SR is shown in the following formula:

[0081] O SR = scale · O effective = (scale · SIZE h′ ) × (scale · SIZE w′ ).

[0082] The large-width remote sensing image after super-resolution reconstruction is stitched and restored by each reconstructed image block I SR (m, n). The width W″ and height H″ of the area O stitched of the large-width remote sensing image after reconstruction can be calculated based on H, W, P, SIZE h , SIZE w and scale, and its mathematical expression is shown in the following formula:

[0083]

[0084] where [·] represents the ceiling operation. It is necessary to place I SR (m,n) at the corresponding position of O stitched . By analogy with the cropping process, assume that m and n respectively represent the indices of I SR in the height and width directions. Then, the mathematical expression of the upper-left corner coordinates of each I SR during the stitching process is shown in the following formula:

[0085] Position = (m · scale · (SIZE h′ - P), n · scale · (SIZE w′ - P)).

[0086] After super-resolution processing, the size of each image block I SR (m,n) becomes SIZE SR × SIZE SR (by default in this article, SIZESR = scale · SIZE h′ = scale · SIZE w′ ), but due to SIZE SR there is a filled size P · scale in it. We only need to take the effective image area in the middle part for stitching. Let the stitched large-width remote sensing image be I stitched The mathematical expression of this stitching process is shown as follows:

[0087]

[0088] where the upper left coordinate of I SR (m, n) is (i, j), and i and j represent pixel indices. The above operation removes the edge part from each super-resolution image block and only retains the effective area in the middle for stitching. Since the length and width are rounded up when constructing the region O, there is a zero-value matrix filled during the rounding-up operation at the right and lower edges of the final stitched image. It is necessary to intercept the first H″ rows and W″ columns. The final obtained stitched image I output The mathematical expression is shown as follows:

[0089] I output = I stitched , 0 < i < H″; 0 < j < W″

[0090] The above stitching process places each I SR (m, n) in the corresponding position and adjusts the stitching step according to the magnification factor, finally effectively reducing the loss of edge pixels of the image block caused by super-resolution reconstruction and realizing the super-resolution reconstruction of the large-width remote sensing image.

[0091] Compared with the method of performing super-resolution reconstruction on small-size (such as 216×216) remote sensing image blocks, this method can directly process the entire large-width remote sensing image, effectively avoiding the problem of incomplete feature extraction caused by the lack of neighboring pixels in the edge area of the image block, thus effectively reducing the loss of edge information and having obvious advantages in various downstream visual task applications.

[0092] Based on the above entire scheme, the method example of actual operation is as follows:

[0093] Application step 1: Determine the low-resolution image data for super-resolution reconstruction. Data can be selected, paid for, and downloaded from existing remote sensing satellite data sharing service platforms. The specific satellite platform can be customized, depending on the resolution of the satellite platform. At this time, there is no requirement for the repetition rate of the image, and only large-width visible light remote sensing images are needed. Through the above channels, the low-resolution large-width remote sensing image I LR can be obtained.

[0094] Application Step 2: Construction of high-low resolution dataset and image preprocessing. The obtained I LR is loaded as input into KernelGAN for blur kernel estimation. This model outputs the blur kernel k of the image. Subsequently, a variance threshold is set to crop noise blocks from the flat scene of I LR to collect noise samples n. By convolving the blur kernel k with the high-resolution image I HR , downsampling, and then adding the noise n. This step can obtain the low-resolution image I LR , and complete the construction of the high-low resolution image pair. Then, the low-resolution image is input into the EESN network, and the output is the low-resolution image with enhanced edge features.

[0095] Application Step 3: Sliding window cropping based on adjacent image block overlap. The low-resolution wide-swath remote sensing image with enhanced edge features is used as input and fed into the interface for cropping. The purpose of this step is to obtain low-resolution image blocks suitable for processing by the super-resolution reconstruction network.

[0096] Application Step 4: Super-resolution reconstruction. The cropped image blocks are used as the E-SRADSGAN super-resolution reconstruction network. The purpose of this step is to obtain the super-resolution reconstruction results of the low-resolution image blocks.

[0097] Application Step 5: Stitching of effective regions of reconstructed image blocks. Select the effective regions after super-resolution reconstruction of the image blocks for stitching to obtain the complete reconstruction result of the wide-swath visible light remote sensing image. The overall flowchart is shown in Figure 4 as follows.

[0098] This method uses three modeling methods based on the E-SRADSGAN large model and the KernelGAN and EESN small models to achieve the super-resolution reconstruction task. First, the KernelGAN and EESN small models are used to construct the high-low resolution image pair dataset and complete the image preprocessing with enhanced edge information. Then, the wide-swath visible light remote sensing image with enhanced edge information is cropped by the sliding window cropping strategy based on adjacent image block overlap. The cropped image blocks are input into the E-SRADSGAN large model for super-resolution reconstruction, and combined with the stitching strategy of the effective regions of the reconstructed image blocks to stitch the reconstructed image blocks, and finally obtain the complete super-resolution reconstruction result of the wide-swath visible light remote sensing image. It has improved in terms of processing speed and processing accuracy compared with traditional super-resolution reconstruction methods. On the basis of existing super-resolution reconstruction methods, the super-resolution reconstruction of small-size images is extended to wide-swath remote sensing images, providing a new method for the super-resolution reconstruction of wide-swath remote sensing images.

[0099] The specific application ways of the present invention are numerous. The above description is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art of this technology, without departing from the principle of the present invention, several improvements can still be made, and these improvements should also be regarded as the protection scope of the present invention.

Claims

1. A method for super-resolution reconstruction of large-width optical remote sensing images, characterized in that, The method includes the following steps: S1: Collect a wide-field visible light remote sensing image, downsample the wide-field visible light remote sensing image using a degradation link based on blur kernel and noise estimation, simulate the degradation process of a real remote sensing image to obtain a low-resolution image, and then use an edge information enhancement auxiliary network to enhance the edge information of the low-resolution image to obtain a low-resolution image with enhanced edge features; S2: Use a sliding window cropping strategy based on adjacent image block overlap to crop the low-resolution image with enhanced edge features to obtain cropped low-resolution image blocks; S3: Use the trained E-SRADSGAN model to perform super-resolution reconstruction on the cropped low-resolution remote sensing image blocks to obtain reconstructed image blocks; S4: Use a reconstructed image block effective area stitching strategy to stitch the reconstructed image blocks, and finally perform edge processing to obtain a complete super-resolution reconstructed wide-field visible light remote sensing image.

2. The method for super-resolution reconstruction of a large-width optical remote sensing image according to claim 1, characterized in that, Specifically, S2 is as follows: S11, Obtain the real low-resolution wide-field visible light remote sensing image I real ; S12, downsample the wide-field visible light remote sensing image I real to obtain a high-resolution image I HR ; S13. Input the wide - width visible - light remote - sensing image I real into a preset fuzzy kernel extraction model to obtain fuzzy kernel information k; S14, set the variance threshold to crop noise blocks from the flat scene of the wide-field visible light remote sensing image I real to collect noise samples and obtain the degraded noise n; S15, convolve the blurred kernel with the high-resolution image I HR and perform downsampling, then add noise n to generate the low-resolution image I LR ; S16, input the low-resolution image I LR into a preset edge information-assisted enhancement network to obtain a low-resolution image I with enhanced edge features E-LR .

3. A method for super-resolution reconstruction of a large-width optical remote sensing image according to claim 1, characterized in that, Specifically, S2 is as follows: S21, obtain the low-resolution image I after enhancing the edge features E-LR and the height and width information O = H' × W' of where H′ and W′ are the height and width of region O respectively; [·] represents the ceiling operation; S22 then initializes a new area by filling P pixels in four directions of I E-LR according to the height and width information; S23: Calculate the cropping step size of the sliding window according to the height and width information and the filled pixel P; S24, place I E-LR at the center of the expanded new area to complete image preprocessing and obtain the new image I E-LR '; where i and j represent pixel indices; S25, use a sliding window to crop the new image I to be cropped according to the cropping step size E-LR ′, the upper left corner coordinates of the sliding window for cropping are (i, j), and the image patches obtained by cropping are saved in a unified format, and its mathematical expression is shown as follows: (i, j) = (m·S h , n·S w ); where S h and S w are step sizes, and m and n respectively represent the indices of the cropping sliding window in the y and x directions.

4. A method for super-resolution reconstruction of a large-width optical remote sensing image according to claim 1, characterized in that Specifically, S4 is as follows: S41: Read the graph blocks after super-resolution reconstruction; S42: Stitch the image blocks after super-resolution reconstruction in the cropping order, discard the filled areas around the image blocks, and only take the middle effective areas for stitching; S43: Remove the extended outer frame area from the stitched wide-field visible light remote sensing image above to obtain the super-resolution reconstruction result of the complete wide-field visible light remote sensing image.

Images