Remote sensing image compression and reconstruction method based on feature perception and latent diffusion super-resolution

Abstract

The present application belongs to the technical field of image processing, and specifically relates to a remote sensing image compression and reconstruction method based on feature perception and latent diffusion super-resolution. The method comprises an encoding and compression stage and a decoding and reconstruction stage. The encoding and compression stage performs intelligent analysis, selective compression and data packaging on the original high-resolution remote sensing image, intelligently identifies and separates the feature region and the homogeneous background region in the image through a convolutional neural network, and differentially processes the two regions, with the feature region being recorded at high precision and the homogeneous background region being aggressively down-sampled. The decoding and reconstruction stage recovers a high-resolution reconstructed image from the compressed data packet, and performs reconstruction through a latent diffusion super-resolution model. The model can utilize the semantic context of the image to generate visually natural textures, effectively avoiding overall blurring and other problems caused by traditional interpolation, and thus enabling the image to still well maintain key feature details under high compression ratio.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and in particular relates to a method for compression, storage, transmission and reconstruction of remote sensing images, specifically a remote sensing image compression and reconstruction method based on feature perception and latent diffusion super-resolution. Background Technology

[0002] With the rapid development of remote sensing technology, the data volume of high-resolution remote sensing images is growing exponentially. This poses a huge challenge to data storage space and transmission bandwidth. Traditional image compression standards (such as JPEG and JPEG2000) usually adopt coding methods based on discrete cosine transform or wavelet transform, which can lead to severe block artifacts or blurring at high compression ratios. In particular, for remote sensing images containing rich textures and sharp ground boundary (such as building edges, roads, and field ridges), the reconstruction quality often fails to meet the needs of subsequent analysis (such as ground feature classification, change detection, and target recognition).

[0003] In recent years, deep learning-based image compression and super-resolution technologies have made significant progress. However, existing deep learning compression methods typically process the entire image uniformly, failing to fully consider the uneven distribution of ground features in remote sensing images. Most areas in an image may be homogeneous regions with simple textures (such as water bodies or flat farmland), while only a small portion are high-frequency regions with complex features (such as urban building clusters or transportation hubs). Performing high-fidelity compression on homogeneous regions is a waste of computational and storage resources.

[0004] Therefore, there is an urgent need in this field for a technical solution that can intelligently distinguish different important regions in an image and apply differentiated compression strategies, while also restoring image details with high quality during reconstruction. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a remote sensing image compression and reconstruction method based on feature perception and latent diffusion super-resolution. This method combines convolutional neural network feature region extraction, selective downsampling and latent diffusion super-resolution reconstruction for efficient remote sensing image processing, enabling the image to retain key ground feature details well even at high compression ratios.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a remote sensing image compression and reconstruction method based on feature perception and latent diffusion super-resolution, including an encoding and compression stage and a decoding and reconstruction stage.

[0007] Phase 1: Encoding and Compression Phase, including: Step S11: Using a pre-trained convolutional neural network, perform pixel-by-pixel feature saliency analysis on the input original high-resolution remote sensing image to generate a feature saliency map of the same size as the input image. The value of each pixel in the feature saliency map is determined by the difference between that location and the surrounding area.

[0008] Step S12: The feature saliency map generated in step S11 is binarized according to a threshold to generate a binarized feature mask. The binarized feature mask distinguishes the image pixels into feature points to be retained and background areas to be sampled for noise reduction.

[0009] Step S13: Extract the pixel values ​​and position coordinates of all feature points in step S12 to form a feature point set, and perform interval sampling on the background area in step S12 to obtain a downsampled background image.

[0010] Step S14: Encapsulate the pixel values ​​and location coordinates of the feature point set, the downsampled background image and sampling rate, and the metadata of the original image size to form a compressed data package for storage and transportation.

[0011] Phase Two: Decoding and Reconstruction Phase, including: Step S21: Parse the compressed data packet in step S14 to obtain the pixel values ​​and location coordinates of the feature point set, the downsampled background image and sampling rate, and the metadata of the original image size.

[0012] Step S22: The downsampled background image is upsampled to restore the original image size. Based on the feature point set and its position coordinates, it is embedded into the background image restored to the original image size to generate a feature-enhanced guiding image.

[0013] Step S23: Using the feature enhancement guidance image from step S22 as conditional input, drive the pre-trained latent diffusion super-resolution model to perform iterative denoising and reconstruction, and output a high-resolution reconstructed image.

[0014] Preferably, the convolutional neural network in step S11 is a semantic segmentation network with an encoder-decoder structure, which is trained under supervision using remote sensing images and their corresponding binarized feature masks generated based on local variance or edge density.

[0015] Preferably, the greater the difference between each pixel position in the feature saliency map in step S11 and the surrounding area, the larger its pixel value.

[0016] Preferably, the threshold in step S12 is an adaptive threshold, which is determined by the Otsu algorithm or the quantile method based on the distribution of the feature saliency map histogram.

[0017] Preferably, the latent diffusion super-resolution model used in step S23 is obtained by training in the following way: using a large number of original high-resolution remote sensing images and their corresponding feature-enhanced guiding images pre-generated by the methods described in steps S11 to S22 as training pairs, the diffusion model is conditionally trained in the latent space; the reconstructed latent features are mapped back to the pixel space through the decoder to obtain the high-resolution reconstructed image.

[0018] Preferably, the metadata in step S14 includes at least the size of the original high-resolution remote sensing image, the sampling rate of the background area, the pixel values ​​of the feature point set, location coordinates, and number.

[0019] Preferably, it also includes a reconstruction quality assessment step, the specific assessment steps of which are as follows: Step S31: Generate binarized feature masks for the original high-resolution remote sensing image and the high-resolution reconstructed image using the convolutional neural network in step S11.

[0020] Step S32: Based on the binarized feature masks of the original high-resolution remote sensing image and the high-resolution reconstructed image generated in step S31, calculate the local structural similarity index of the original high-resolution remote sensing image and the high-resolution reconstructed image in the feature region, respectively. And the local structural similarity index in the background region .

[0021] Step S33: Calculate the gradient distortion of the original high-resolution remote sensing image and the high-resolution reconstructed image at all feature point locations. .

[0022] Step S34: Calculate the local structural similarity index in the feature region. Local structural similarity index in the background region and gradient distortion at all feature point locations The feature-preserving structural similarity index FP_SSIM is obtained by weighted fusion according to preset weights. Then, FP_SSIM is compared with the preset standard value. When FP_SSIM is greater than the preset standard value, the reconstruction quality is considered to meet the requirements.

[0023] Preferably, the calculation method of the feature-preserving structural similarity index FP_SSIM is as shown in formula (1). (1), in, , , These are the weighting coefficients, and .

[0024] Preferably, the local structural similarity index in the feature region And the local structural similarity index in the background region Using the same local window The formula is as shown in formula (2). (2), Where x represents an image patch from the original high-resolution remote sensing image, and y represents an image patch from the reconstructed high-resolution image. Let x and y be the average values ​​of all pixels in image patch x and image patch y, respectively. , These are the standard deviations of all pixels in image patch x and image patch y, respectively. It is the covariance of image patch x and image patch y. , A minimal constant set to prevent the denominator from approaching zero.

[0025] The original high-resolution remote sensing image and the high-resolution reconstructed image are processed using a sliding window method, traversing the entire image to calculate and output a full-image structural similarity map. .

[0026] Then the local structural similarity index in the feature region The arithmetic mean of all values ​​in the feature region is calculated as shown in formula (3). (3), in, With a value of 1 at the feature point and 0 at the background point, the numerator represents only accumulation. The values ​​of the pixels marked as "features" by the mask are calculated, with the denominator being the total number of pixels in the feature region.

[0027] The background region structural similarity index The arithmetic mean of all values ​​in the background region is calculated as shown in formula (4). (4), in, This is a mask for the background region, with values ​​of 1 at background points and 0 at feature points. The numerator represents the sum of all values. The values ​​of the pixels marked as "background" by the mask are calculated, with the denominator being the total number of pixels in the background area.

[0028] feature point distortion It is a natural exponential function, and its calculation method is shown in formula (5). (5), in, is a scaling factor, a constant greater than 0, used to control the "descent rate" of the exponential function, and D is the mean squared error of the gradient domain of the feature points; the larger the value, the more severe the average distortion.

[0029] The mean squared error D of the gradient domain of feature points is calculated as shown in formula (6). (6), Where N is the total number of feature points. This indicates the location of feature point i on the original high-resolution remote sensing image. gradient magnitude, This indicates that on a high-resolution reconstructed image, the same feature point i is at the same location. The gradient magnitude.

[0030] Preferably, when FP_SSIM > 0.9, the reconstruction quality is considered to meet the requirements.

[0031] The beneficial effects of this invention are as follows: 1. Achieving a balance between high compression ratio and high feature fidelity. This method intelligently identifies and separates feature regions from homogeneous background regions in images using a convolutional neural network, applying differentiated processing to both. Feature regions are recorded with high precision, while homogeneous background regions are aggressively downsampled. This ensures that critical information is not lost, thus achieving an extremely high compression ratio while still preserving the fine features of key features such as building edges, roads, and field ridges, overcoming the problem of severe feature degradation under high compression ratios in traditional transform coding.

[0032] 2. It achieves intelligent reconstruction rather than simple interpolation. For background areas with significantly compressed data, reconstruction is performed using a latent diffusion super-resolution model. This model, conditioned on a guide image containing fine feature information, can "understand" the semantic context of the image, thereby generating visually natural, detailed, and physically consistent textures. It can effectively avoid problems such as overall blurring and texture distortion caused by traditional interpolation.

[0033] 3. Significantly improves storage and transmission efficiency. The compression strategy of this invention can directly and drastically reduce the original data size, providing a practical solution for low-cost long-term storage of massive remote sensing data and real-time transmission under limited channels, and has great engineering application value.

[0034] 4. It provides a flexible quality-compression ratio trade-off mechanism. By adjusting the feature extraction threshold (controlling how many features are retained) and the background region downsampling rate (controlling the degree of background compression), users can make a smooth and flexible trade-off between "extreme compression" and "extreme fidelity" to adapt to different levels of application needs, making the system highly adaptable. Attached Figure Description

[0035] Figure 1 This is an overall flowchart of the method of the present invention.

[0036] Figure 2 This is a detailed process diagram of the encoding and compression stages.

[0037] Figure 3 This is a detailed process diagram of the decoding and reconstruction phases. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0039] like Figure 1 As shown, the present invention provides a remote sensing image compression and reconstruction method based on feature perception and latent diffusion super-resolution, including an encoding and compression stage and a decoding and reconstruction stage.

[0040] The first stage, encoding and compression, aims to perform intelligent analysis, selective compression, and data encapsulation of the original high-resolution remote sensing imagery. Its detailed process is as follows: Figure 2 As shown, the specific steps are as follows: S101 Input and Preprocessing: Input a raw high-resolution remote sensing image I original Its size is The image is in .RGB format and undergoes preprocessing operations such as normalization.

[0041] S102 Feature Saliency Map Generation: The preprocessed image is input into a pre-trained U-Net feature extraction network, which outputs a saliency map that is identical to the input original high-resolution remote sensing image. original A single-channel feature saliency map S of the same size, where each pixel value is S(i,j). [0,1] represents the importance (saliency score) of the corresponding position (i,j) in the image. The higher the value, the greater the likelihood that the pixel belongs to a key feature region (such as a greenhouse in a farmland area, the edge of a building, etc.).

[0042] S103 Binarization Feature Mask Generation: An adaptive thresholding algorithm is applied to the feature saliency map S to generate a binary feature mask M, the calculation method of which is shown in formula (7). (7), in, This is an adaptive threshold.

[0043] This invention provides two threshold determination methods: the Otsu algorithm and the quantile method. The Otsu algorithm automatically calculates the threshold that maximizes the inter-class variance between the feature region and the background region. The quantile method, on the other hand, sets a threshold to retain a certain percentage of the most significant pixels, for example, retaining the top 15% of the most significant pixels. The value of all pixels in the saliency map S is located in the upper 15th percentile after being sorted in descending order.

[0044] Pixels with a value of 1 in the mask M are defined as feature points, while regions with a value of 0 are defined as background regions.

[0045] S104 Selective Downsampling and Data Extraction: Differentiated processing of the original high-resolution remote sensing image based on mask M: 1. Feature point set extraction: Extract all features that satisfy... For pixels with a value of 1, record their coordinates (x, y). i y i ) and original RGB pixel values ​​(R i G i B i ), constituting the feature point set P= , where N is the total number of feature points.

[0046] 2. Background Area Downsampling: Background areas marked as 0 in mask M are sampled at intervals. The downsampling rate is set to r, meaning one pixel is retained in both the row and column directions every r pixels. The resolution of the original high-resolution remote sensing image in this area then increases from H... W decreases to (H / r) (W / r) yields the downsampled background image I. bg_down .

[0047] S105 Metadata Recording and Data Encapsulation: Generate and record the metadata for this compression, including at least the original size (H). W), downsampling rate (r), number of feature points (N), version of the feature extraction network U-Net, adaptive threshold Information such as file storage path, and metadata information are stored via XML files. Finally, the feature point set P and the downsampled background image I are combined... bg_down Encapsulate the data along with metadata (e.g., using a zip format) to generate a final compressed data packet for storage or transmission.

[0048] Compared to the original image, the compressed data size is significantly reduced. For example, the original image is 2048×2048×3≈12.6MB. Assuming the feature point ratio is 5%, the number of feature points N is approximately 2048×2048×0.05≈209715. Storing their coordinates and RGB values ​​requires approximately 209715×(2+2+3) bytes≈1.47 MB. The downsampled background image is 1024×1024×3≈3.15MB. The total data size is approximately 4.62 MB, and the compression ratio is approximately 2.7:1.

[0049] In this stage, the training of the feature extraction network U-Net uses remote sensing datasets (such as LoveDA and InriaAerial Image Labeling). The Sobel edge intensity map of the image is calculated, and a threshold is set to generate a binary mask as a training label to supervise U-Net's learning to recognize feature regions.

[0050] The second stage, decoding and reconstruction, aims to recover a high-resolution reconstructed image from the compressed data packet. The detailed process is as follows: Figure 3 As shown, the specific steps are as follows: S201 Data Unpacking and Parsing: Receives and decompresses compressed data packets, parses metadata, and reads the feature point set P and downsampled background image I. bg_down .

[0051] S202 generates feature-enhanced guided images: 1. Background upsampling: Based on the original size H in the metadata. W and downsampling rate r, for I bg_down Perform bilinear interpolation upsampling to restore its size to H. W, obtain the initial full-size background image I bg_up .

[0052] 2. Feature point embedding: Traverse the feature point set P, and extract the original RGB values ​​(R values) of each feature point. i G i B i Write back I precisely bg_up The corresponding coordinates (x) i y i If the interpolated pixels at a given location are overwritten, a feature-enhanced guide image I is generated. guide The image contains precise original information at the feature point locations, while the background area is a coarse interpolation result.

[0053] S203 Super-resolution reconstruction based on latent diffusion model: Feature-enhanced guiding image I guideAs a conditional input, it is fed into a pre-trained latent diffusion super-resolution model for high-quality reconstruction. The specific steps are as follows: 1. Network Input: The latent diffusion super-resolution model is a general model pre-trained using a large amount of data. During the training phase of this model, guide images generated from a large number of images through the encoding phase of this invention and their corresponding high-resolution images are used as training pairs. In this model, I... guide As a condition, we start with a randomly sampled noise tensor in the latent space.

[0054] 2. Iterative Denoising: The model performs a multi-step iterative denoising process. In each step, the denoising U-Net combines the noise latent variables from the previous step with those from I... guide The extracted conditional features are used as input to predict and remove some noise.

[0055] 3. Output Results: After all denoising steps, a clean latent feature representation is obtained. This representation is then converted back to pixel space by a decoder, and the final high-resolution reconstructed image I is output. reconstructed This process intelligently diffuses clear feature point information to the surrounding area, generating natural and realistic background textures. For example, around clear road feature points, the super-resolution model reconstructs a continuous, smooth road texture; around building outlines, it reconstructs clear edge and roof textures. Ultimately, it outputs a high-quality, high-resolution reconstructed image. reconstructed It is visually very close to the original image, and key features have been perfectly preserved.

[0056] Using the original high-resolution image as the target, corresponding guiding images are generated through the encoding stage process of this invention (using a pre-trained feature extraction network U-Net) to form training pairs. A standard LDM training process is employed to train the diffusion model in the latent space.

[0057] S204 Reconstruction Quality Assessment: After image reconstruction is completed, the feature-preserving structural similarity index FP_SSIM proposed in this invention is used to quantitatively assess the reconstruction quality. The specific steps are as follows: Using the same U-Net and thresholding methods as in the encoding stage, the original high-resolution remote sensing image I... original and high-resolution reconstructed images I reconstructed The process is performed to obtain a feature mask for evaluation.

[0058] Calculate the local structural similarity index of the feature regions respectively. Local structural similarity index in the background region and gradient distortion at all feature point locations .

[0059] Then, the FP_SSIM value is synthesized according to formula (1). (1), in, , , These are the weighting coefficients, and In this embodiment, according to taking =0.6、 =0.2、 =0.2, when A value greater than 0.9 indicates that the reconstruction quality meets the requirements of most remote sensing analysis applications. This index allows for an objective and quantitative assessment of the performance advantages of the method described in this invention in preserving key ground features, providing guidance for the selection of compression parameters and algorithm optimization.

[0060] The local structural similarity index of the feature region Local structural similarity index in the background region and gradient distortion at all feature point locations The calculation method is as follows: Local structural similarity index in the feature region And the local structural similarity index in the background region Using the same local window The formula is as shown in formula (2). (2), Where x represents an image patch from the original high-resolution remote sensing image, and y represents an image patch from the reconstructed high-resolution image. Let x and y be the average values ​​of all pixels in image patch x and image patch y, respectively. , These are the standard deviations of all pixels in image patch x and image patch y, respectively. It is the covariance of image patch x and image patch y. , A minimal constant set to prevent the denominator from approaching zero.

[0061] The original high-resolution remote sensing image and the high-resolution reconstructed image are processed using a sliding window method, traversing the entire image to calculate and output a full-image structural similarity map. .

[0062] Then the local structural similarity index in the feature region The arithmetic mean of all values ​​in the feature region is calculated as shown in formula (3). (3), in, With a value of 1 at the feature point and 0 at the background point, the numerator represents only accumulation. The values ​​of the pixels marked as "features" by the mask are calculated, with the denominator being the total number of pixels in the feature region.

[0063] The background region structural similarity index The arithmetic mean of all values ​​in the background region is calculated as shown in formula (4). (4), in, This is a mask for the background region, with values ​​of 1 at background points and 0 at feature points. The numerator represents the sum of all values. The values ​​of the pixels marked as "background" by the mask are calculated, with the denominator being the total number of pixels in the background area.

[0064] feature point distortion It is a natural exponential function, and its calculation method is shown in formula (5). (5), in, is a scaling factor, a constant greater than 0, used to control the "descent rate" of the exponential function, and D is the mean squared error of the gradient domain of the feature points; the larger the value, the more severe the average distortion.

[0065] The mean squared error D of the gradient domain of feature points is calculated as shown in formula (6). (6), Where N is the total number of feature points. This indicates the location of feature point i on the original high-resolution remote sensing image. gradient magnitude, This indicates that on a high-resolution reconstructed image, the same feature point i is at the same location. The gradient magnitude.

[0066] It should be noted that, in this invention, all "features" refer to regions in remote sensing images that, in order to meet specific downstream analysis tasks (such as land cover classification, target detection, and change detection), must prioritize maintaining their structural integrity and detail clarity. These typically include, but are not limited to, building outlines, road networks, field ridges, and water body boundaries. Convolutional neural networks predict feature regions by learning common patterns from a large number of such samples, rather than simply fitting edge maps. The training labels generated in the above manner essentially mark regions in the image that are structurally complex and information-rich. Practice shows that these regions have a high degree of overlap with the key feature regions of interest in downstream semantic analysis tasks. Therefore, networks trained under the supervision of these labels can effectively predict feature regions valuable for subsequent applications. A latent diffusion super-resolution model refers to a generative model that reconstructs a high-resolution image from a low-resolution conditional image through an iterative denoising process in a low-dimensional latent space. All of the above are terms well-known to those skilled in the art.

[0067] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A remote sensing image compression and reconstruction method based on feature perception and latent diffusion super-resolution, characterized in that: Includes the following steps: The encoding and compression stages include: Step S11: Perform pixel-by-pixel feature saliency analysis on the input original image to generate a feature saliency map of the same size as the input image. The value of each pixel in the feature saliency map is determined by the difference between the pixel position and the surrounding area. The greater the difference between the pixel position and the surrounding area in the feature saliency map, the larger its pixel value. Step S12: Generate a binarized feature mask from the feature saliency map. The binarized feature mask distinguishes the image pixels into feature points to be retained and background areas to be sampled for noise reduction. Step S13: Extract the pixel values ​​and position coordinates of all feature points to form a feature point set, perform interval sampling on the background area, and obtain a downsampled background image; Step S14: Encapsulate the pixel values ​​and location coordinates of the feature point set, the downsampled background image and sampling rate, and the metadata of the original image size to form a compressed data package; The decoding and reconstruction phase includes: Step S21: Parse the compressed data packet in step S14 to obtain the corresponding feature point set data, downsampled background data, and metadata of the original image size; Step S22: Background upsampling: Based on the original image size in the metadata and downsampling rate, for downsampled background images Perform bilinear interpolation upsampling to restore its size to To obtain the initial full-size background image ; Feature point embedding: Traverse the feature point set and extract the original RGB values ​​of each feature point. （R i ，G i ，B i ） Accurately write back the original full-size background image The corresponding coordinates （x i ，y i ） If the interpolated pixels at a given location are covered, a feature-enhanced guide image is generated. This image has precise original information at the feature point locations, while the background area is a coarse interpolation result. Step S23: Input the feature-enhanced guided image into the pre-trained latent diffusion super-resolution model for iterative denoising and reconstruction, and output a high-resolution reconstructed image.

2. The remote sensing image compression and reconstruction method based on feature perception and latent diffusion super-resolution according to claim 1, characterized in that: The convolutional neural network in step S11 is a semantic segmentation network with an encoder-decoder structure, which is trained under supervision using remote sensing images and their corresponding binary feature masks generated based on local variance or edge density.

3. The remote sensing image compression and reconstruction method based on feature perception and latent diffusion super-resolution according to claim 1, characterized in that: The threshold in step S12 is an adaptive threshold, which is determined by the Otsu algorithm or the quantile method based on the distribution of the feature saliency map histogram.

4. The remote sensing image compression and reconstruction method based on feature perception and latent diffusion super-resolution according to claim 1, characterized in that: The latent diffusion super-resolution model used in step S23 is obtained by training a large number of original high-resolution remote sensing images and their corresponding feature-enhanced guiding images pre-generated by the methods described in steps S11 to S22 as training pairs, and conditionally training the diffusion model in the latent space. The decoder maps the reconstructed latent features back to pixel space to obtain a high-resolution reconstructed image.

5. The remote sensing image compression and reconstruction method based on feature perception and latent diffusion super-resolution according to claim 1, characterized in that: The metadata in step S14 includes at least the size of the original high-resolution remote sensing image, the sampling rate of the background area, the pixel values ​​of the feature point set, the location coordinates, and the number of points.

6. The remote sensing image compression and reconstruction method based on feature perception and latent diffusion super-resolution according to claim 1, characterized in that: It also includes a reconstruction quality assessment step, the specific assessment steps are as follows: Step S31: Generate binarized feature masks for the original high-resolution remote sensing image and the high-resolution reconstructed image using the convolutional neural network in step S11. Step S32: Based on the binarized feature masks of the original high-resolution remote sensing image and the high-resolution reconstructed image generated in step S31, calculate the local structural similarity index of the original high-resolution remote sensing image and the high-resolution reconstructed image in the feature region, respectively. And the local structural similarity index in the background region ; Step S33: Calculate the gradient distortion of the original high-resolution remote sensing image and the high-resolution reconstructed image at all feature point locations. ; Step S34: Calculate the local structural similarity index in the feature region. Local structural similarity index in the background region and gradient distortion at all feature point locations The feature-preserving structural similarity index is obtained by weighting and fusing according to preset weights. Then by comparison Compared with the preset standard value, when If the value exceeds the preset standard value, the reconstruction quality is considered to meet the requirements.

7. The remote sensing image compression and reconstruction method based on feature perception and latent diffusion super-resolution according to claim 6, characterized in that: The feature-preserving structural similarity index The calculation method is shown in formula (1). （1）， in, 、 、 These are the weighting coefficients, and .

8. The remote sensing image compression and reconstruction method based on feature perception and latent diffusion super-resolution as described in claim 6, characterized in that: when When the value is greater than 0.9, the reconstruction quality is considered to meet the requirements.

Images