Satellite image automatic screening method based on multi-feature fusion quadtree

Abstract

The application discloses a kind of satellite image automatic screening methods based on multi-feature fusion quadtree, solve the single feature dependence of quadtree image processing, parameter static solidification problem, belong to satellite image screening field;Including: using Gaussian filter to the original satellite image denoising obtains denoising image;Denoising image is as quadtree root node, the variance of each root node, edge density and gradient direction are calculated, gradient direction is divided into a preset number of intervals, the main direction proportion is counted, and the direction consistency value is obtained;Each root node meets split condition simultaneously and triggers split;All nodes of the quadtree after splitting are traversed, and the statistical characteristics for verifying whether each node meets the split condition are collected;The comprehensive score of statistical characteristics is calculated;According to descending order of comprehensive score, sequentially select denoising image of preset proportion in sequence as screening image;The application improves satellite image target screening precision.

Description

Technical Field

[0001] This invention belongs to the field of satellite image screening technology, and relates to an automatic satellite image screening method based on a multi-feature fusion quadtree. Background Technology

[0002] A quadtree is a hierarchical spatially partitioned data structure that originated in the fields of computer graphics and geographic information systems.

[0003] Traditional quadtree methods achieve image filtering through the following steps: ① Region segmentation: Using the entire satellite image as the root node, the image is divided into four quadrants; ② Region representation: Calculate the pixel variance of each node region to quantify local complexity; ③ Split control: Decide whether to continue segmentation based on a preset pixel variance threshold; ④ Filtering decision: Filter images containing complex regions based on depth or number of nodes.

[0004] With the rapid development of aerospace technology, satellite imagery data has experienced explosive growth. However, the proportion of truly valuable images within this massive dataset is limited, with a large number of redundant or low-quality images. Given the limited onboard storage resources and severely constrained downlink bandwidth, efficiently selecting key images and prioritizing their download to ground stations is crucial for ground systems to accurately assess satellite status and optimize mission planning. Traditional quadtree methods have significant limitations in satellite image selection:

[0005] Relying solely on pixel variance as a splitting criterion cannot effectively characterize diverse features such as texture and geometry, leading to missed detections of small targets. Fixed parameter thresholds result in a lack of dynamic adaptability, making it difficult to cope with resolution differences and complex lighting interference in different satellite images. This requires repeated manual parameter tuning and has limited generalization ability. Summary of the Invention

[0006] To address the technical problems of single feature dependence and static parameter fixation in traditional quadtree image processing techniques, this invention proposes an automatic satellite image screening method based on multi-feature fusion quadtree. This method uses joint splitting decisions based on size-variance-edge density-orientation consistency features to select information-rich images from massive amounts of satellite images, providing efficient and reliable data screening support for intelligence analysis. Through image preprocessing, quadtree splitting, and multi-feature joint analysis, it achieves accurate screening of highly complex satellite images.

[0007] The objective of this invention is specifically achieved through the following technical solutions:

[0008] This invention discloses an automatic satellite image selection method based on a multi-feature fusion quadtree, the method comprising:

[0009] Step 1: Use Gaussian filtering to denoise the original satellite image to obtain a denoised image;

[0010] Step 2: Using the denoised image as the root node of the quadtree, calculate the variance, edge density, and gradient direction of each root node, divide the gradient direction into a preset number of intervals, and count the proportion of the main direction to obtain the direction consistency value.

[0011] Step 3: When the node size, variance, edge density, and orientation consistency value of each root node simultaneously meet the splitting conditions, a split is triggered; all nodes of the split quadtree are traversed and statistical features are collected; the comprehensive score of the statistical features is calculated; the images are sorted in descending order of the comprehensive score, and a preset proportion of denoised images in the sequence are selected as the screening images.

[0012] In step one, the method for obtaining a denoised image by using Gaussian filtering to denoise the original satellite image is as follows:

[0013] The GaussianBlur function in OpenCV, used in linear smoothing filters, removes Gaussian noise from the original satellite image by specifying the Gaussian kernel size and variance parameter, resulting in a denoised image.

[0014] In step one, the GaussianBlur function is:

[0015] ;

[0016] In the formula, yes The weight value at that location, These are the coordinates of the pixels in the original satellite image relative to the center of the Gaussian kernel. It is the standard deviation obtained by taking the square root of the variance parameter of the Gaussian distribution, and e is the natural constant.

[0017] In step two, the variance, edge density, and gradient direction are calculated as follows:

[0018] ;

[0019] ;

[0020] ;

[0021] In the formula, Let I be the variance, and I be the current node's image patch. Where is the total number of pixels, and i is the pixel index. For the first grayscale value of each pixel. The average pixel value of the region; For edge density, The gradient magnitude calculated for the Sobel operator; To compute the gradient direction using the Sobel operator, For horizontal gradient components, This represents the vertical gradient component.

[0022] In step two, the gradient direction is divided into a preset number of intervals, and the proportion of the main direction is counted to obtain the direction consistency value.

[0023] ;

[0024] In the formula, For directional consistency values, the maximum count interval in the directional histogram is: the gradient direction =0°~180° is divided into 8 direction histogram counting intervals, with each interval being 22.5°, and the maximum counting interval in the direction histogram is counted.

[0025] In step three, the splitting condition is:

[0026] ;

[0027] In the formula, the node size is the size of the current node image block I; Minimum segment size, Let Variance be the constant. for The initial variance, The edge density threshold, This is the directional consistency threshold.

[0028] In step three, the steps that trigger the split include:

[0029] The root node of the quadtree is split into four child nodes, corresponding to the four quadrants: Northwest (NW), Northeast (NE), Southwest (SW), and Southeast (SW). Each child node is further divided into four child nodes until the terminal node corresponding to the task requirement is obtained and marked as a leaf node, at which point the splitting stops.

[0030] In step three, the collected statistical characteristics are as follows:

[0031] ;

[0032] In the formula, F represents the statistical characteristic; N represents the total number of splits in the quadtree; Maximum depth; This represents the average edge density.

[0033] In step three, the calculation method for the comprehensive score of statistical features is as follows:

[0034] Score = 0.5 × +0.3×N+0.2× .

[0035] In step three, the maximum depth is calculated as follows:

[0036] ;

[0037] In the formula, Where W is the node size, and H is the height.

[0038] The beneficial effects of this invention are:

[0039] 1. The satellite image automatic screening method based on multi-feature fusion quadtree provided by this invention solves the technical problems of single feature dependence and static parameter fixation in traditional quadtree image processing technology by introducing innovative designs such as multi-feature fusion splitting conditions, orientation consistency detection and weighted scoring on the basis of traditional quadtree algorithm, and significantly improves the accuracy of satellite image target screening.

[0040] 2. This invention proposes a size-variance-edge density-orientation consistency fusion splitting condition: each quadtree node of each image needs to simultaneously calculate the region variance, edge density and orientation consistency, and splitting is triggered only when all four exceed the threshold;

[0041] 3. Directional Consistency Detection: Statistical analysis of the edge direction histogram within a node. Only when the edge direction within a node is highly concentrated and the proportion of the main direction is concentrated (e.g., the proportion of the main direction is >60%), is the node retained for further splitting, thus effectively preserving a limited number of linear structures with consistent directions (e.g., a windsurfing frame).

[0042] 4. Construct a comprehensive scoring formula to effectively filter images based on their scores. The maximum depth weight represents the richness of information; the greater the depth, the more complex the local information. The number of nodes has a weight of 0.3 to avoid overfitting due to an excessive number of nodes. The average edge density has a weight of 0.2 to aid verification and retain high edge density regions. The orientation consistency condition is already included in the splitting condition, so the retained nodes already possess orientation consistency. Recalculating subsequent scores without order would be redundant. Introducing a fourth scoring dimension might increase the risk of overfitting. Instead, using maximum depth, number of nodes, and average edge density for the final score allows for a global distinction between detail richness and structural complexity, resulting in higher computational efficiency. Attached Figure Description

[0043] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0044] Figure 1 This is a schematic diagram of the threshold given parameter provided in the embodiment of the present invention.

[0045] Figure 2This is a schematic diagram of the core code for constructing a quadtree, provided in an embodiment of the present invention, given a threshold splitting condition.

[0046] Figure 3 This is a schematic diagram of the comprehensive scoring calculation formula provided in the embodiments of the present invention.

[0047] Figure 4 This is a schematic diagram of the output of part of the running program provided in the embodiment of the present invention.

[0048] Figure 5 This is a schematic diagram of the comprehensive scoring results and selection results provided in the embodiments of the present invention. Detailed Implementation

[0049] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0050] This invention provides an automatic satellite image selection method based on a multi-feature fusion quadtree, the method comprising:

[0051] Step 1: Use Gaussian filtering to denoise the original satellite image to obtain a denoised image;

[0052] Step 2: Using the denoised image as the root node of the quadtree, calculate the variance, edge density, and gradient direction of each root node, divide the gradient direction into a preset number of intervals, and count the proportion of the main direction to obtain the direction consistency value.

[0053] Step 3: When the node size, variance, edge density, and orientation consistency value of each root node simultaneously meet the splitting conditions, a split is triggered; all nodes of the split quadtree are traversed and statistical features are collected; the comprehensive score of the statistical features is calculated; the images are sorted in descending order of the comprehensive score, and a preset proportion of denoised images in the sequence are selected as the screening images.

[0054] In step one, the method for obtaining a denoised image by using Gaussian filtering to denoise the original satellite image is as follows:

[0055] Using the GaussianBlur function in OpenCV's linear smoothing filter, Gaussian noise in the original satellite image is removed and the image is smoothed by giving the Gaussian kernel size and variance parameter, resulting in a denoised image.

[0056] In step one, the GaussianBlur function is:

[0057] ;

[0058] In the formula, yes The weight value at that location, These are the coordinates of the pixels in the original satellite image relative to the center of the Gaussian kernel. It is the standard deviation obtained by taking the square root of the variance parameter of the Gaussian distribution, and e is the natural constant.

[0059] In step two, the variance, edge density, and gradient direction are calculated as follows:

[0060] ;

[0061] ;

[0062] ;

[0063] In the formula, Let I be the variance, and I be the current node's image patch. Where is the total number of pixels, and i is the pixel index. For the first grayscale value of each pixel. The average pixel value of the region; For edge density, The gradient magnitude calculated for the Sobel operator; To compute the gradient direction using the Sobel operator, For horizontal gradient components, This represents the vertical gradient component.

[0064] In step two, the gradient direction is divided into a preset number of intervals, and the proportion of the main direction is counted to obtain the direction consistency value.

[0065] ;

[0066] In the formula, For directional consistency values, the maximum count interval in the directional histogram is: the gradient direction =0°~180° is divided into 8 direction histogram counting intervals, with each interval being 22.5°, and the maximum counting interval in the direction histogram is counted.

[0067] In step three, the splitting condition is:

[0068] ;

[0069] In the formula, the node size is the size of the current node image block I; Minimum segment size, Let Variance be the constant. for The initial variance, The edge density threshold, This is the directional consistency threshold.

[0070] In step three, the steps that trigger the split include:

[0071] The root node of the quadtree is split into four child nodes, corresponding to the four quadrants: Northwest (NW), Northeast (NE), Southwest (SW), and Southeast (SW). Each child node is further divided into four child nodes until the terminal node corresponding to the task requirement is obtained and marked as a leaf node, at which point the splitting stops.

[0072] In step three, the collected statistical characteristics are as follows:

[0073] ;

[0074] In the formula, F represents the statistical characteristic; N represents the total number of splits in the quadtree; Maximum depth; This represents the average edge density.

[0075] In step three, the calculation method for the comprehensive score of statistical features is as follows:

[0076] Score = 0.5 × +0.3×N+0.2× .

[0077] In step three, the maximum depth is calculated as follows:

[0078] ;

[0079] In the formula, Where W is the node size, and H is the height.

[0080] The verification is implemented as follows:

[0081] This invention was experimentally validated on the SPE3R: Synthetic Dataset for Satellites dataset (accessible at: https: / / www.kaggle.com / datasets / sammahoney / spe3r-synthetic-dataset-for-satellites), with a selection of satellite images used in the experiment. In this publicly available satellite image dataset, images clearly containing the spacecraft's colored solar panels were labeled as positive samples, while the rest were labeled as negative samples. The evaluation metric was accuracy, defined as the number of selected positive samples divided by the total number of selected samples.

[0082] Taking the jason1 satellite imagery from the SPE3R public dataset as an example: Figures 1 to 5 As shown:

[0083] 1) Select 400 images without an Earth background, including 98 positive samples and 302 negative samples;

[0084] 2) The image is denoised by Gaussian filtering, with a Gaussian kernel size of 3×3 and a variance of 15.

[0085] 3) such as Figure 1 As shown, the splitting operation is performed based on the given minimum segmentation size, variance threshold, edge density threshold, and orientation consistency threshold; subsequent experiments mainly adjust the threshold values.

[0086] 4) such as Figure 2 As shown, if all nodes meet the splitting conditions, continue splitting and proceed to step 3); otherwise, stop splitting and proceed to step 5.

[0087] 5) Statistical analysis of node count (total number of splits in the quadtree), maximum depth, and average edge density features;

[0088] 6) For example Figure 3 As shown, the comprehensive score is given according to the comprehensive scoring formula;

[0089] 7) For example Figure 4 As shown, the process iterates through all images, calculating the maximum depth, number of nodes, average edge density, and peak orientation consistency for each image. The splitting operation continues only if the node size, variance, average edge density, and orientation consistency values ​​all exceed a given threshold; otherwise, the splitting process terminates. Then, combining the maximum depth, number of nodes, and average edge density values, the top 25% of images are selected according to a comprehensive scoring formula.

[0090] 8) For example Figure 5 As shown, the accuracy evaluation index was calculated. Based on the output image results, the proportion of positive sample images to the selected results was calculated, and the statistical results are shown in Table 1. It can be seen that compared with the traditional quadtree method, the present invention has a significant improvement in effectively selecting positive sample images. Table 1 is a comparison chart of results with different threshold parameters.

[0091] Table 1

[0092] parameter Traditional quadtree method Method of the present invention promote Minimum segmentation size = 8; Variance threshold = 100; Edge density threshold = 0.05; Orientation consistency threshold = 0.6 54% 78% 24%↑ Minimum segmentation size = 4; Variance threshold = 100; Edge density threshold = 0.05; Orientation consistency threshold = 0.6 62% 78% 16%↑ Minimum segmentation size = 8; Variance threshold = 50; Edge density threshold = 0.05; Orientation consistency threshold = 0.6 54% 78% 24%↑

[0093] In summary, this invention significantly improves the accuracy of target selection in satellite images by introducing multi-feature fusion splitting conditions, directional consistency detection, and a weighted scoring mechanism into the traditional quadtree algorithm. Traditional methods rely solely on pixel variance as the splitting criterion, easily misclassifying high-noise areas (such as cloud reflections) as complex structures, and failing to identify key targets on metal surfaces such as solar panel hinges (where uniform grayscale results in low variance but significant geometric structures). To address this deficiency, this invention proposes a joint decision-making mechanism based on size, variance, edge density, and directional consistency features, achieving collaborative perception of multiple attributes of space targets: variance features capture subtle grayscale changes in solar panels, edge density features identify the continuous linear structure of the solar panel frame, and directional consistency features accurately extract the directional regularity of satellite components through an 8-interval histogram.

[0094] Given the strong directional regularity of components such as satellite solar panels and antennas, the directional consistency detection mechanism designed in this invention exhibits significant advantages. By dividing the edge direction into eight intervals (each interval being 22.5°) and selecting areas with a high proportion of the dominant direction, the parallel edges of the solar panel frame can be accurately distinguished from the chaotic textures of the space scene.

[0095] In the screening and decision-making stage, the comprehensive scoring formula designed in this invention deeply integrates the physical characteristics of space targets: the maximum depth has a high weight (50%) because intricate structures such as sail hinges and sensors inevitably trigger deep splitting; the number of nodes has a second high weight (30%) corresponding to the surge in child nodes generated by complex targets such as sail deployment; and the average edge density weight (20%) avoids false targets with large local depth but global simplicity (such as misjudging a single antenna). In the SPE3R dataset test, the final accuracy is improved compared to the traditional quadtree method.

[0096] The beneficial effects of the embodiments of the present invention are:

[0097] 1. The satellite image automatic screening method based on multi-feature fusion quadtree provided by this invention solves the technical problems of single feature dependence and static parameter fixation in traditional quadtree image processing technology by introducing innovative designs such as multi-feature fusion splitting conditions, orientation consistency detection and weighted scoring on the basis of traditional quadtree algorithm, and significantly improves the accuracy of satellite image target screening.

[0098] 2. This invention proposes a size-variance-edge density-orientation consistency fusion splitting condition: each quadtree node of each image needs to simultaneously calculate the region variance, edge density and orientation consistency, and splitting is triggered only when all four exceed the threshold;

[0099] 3. Directional Consistency Detection: Statistical analysis of the edge direction histogram within a node. Only when the edge direction within a node is highly concentrated and the proportion of the main direction is concentrated (e.g., the proportion of the main direction is >60%), is the node retained for further splitting, thus effectively preserving a limited number of linear structures with consistent directions (e.g., a windsurfing frame).

[0100] 4. Construct a comprehensive scoring formula to effectively filter images based on their scores. The maximum depth weight represents the richness of information; the greater the depth, the more complex the local information. The number of nodes has a weight of 0.3 to avoid overfitting due to an excessive number of nodes. The average edge density has a weight of 0.2 to aid verification and retain high edge density regions. The orientation consistency condition is already included in the splitting condition, so the retained nodes already possess orientation consistency. Recalculating subsequent scores without order would be redundant. Introducing a fourth scoring dimension might increase the risk of overfitting. Instead, using maximum depth, number of nodes, and average edge density for the final score allows for a global distinction between detail richness and structural complexity, resulting in higher computational efficiency.

[0101] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for automatic satellite image selection based on multi-feature fusion quadtree, characterized in that, The method includes: Step 1: Use Gaussian filtering to denoise the original satellite image to obtain a denoised image; Step 2: Using the denoised image as the root node of the quadtree, calculate the variance, edge density, and gradient direction of each root node, divide the gradient direction into a preset number of intervals, and count the proportion of the main direction to obtain the direction consistency value. Step 3: When the node size, variance, edge density, and orientation consistency value of each root node simultaneously meet the splitting conditions, a split is triggered; all nodes of the split quadtree are traversed and statistical features are collected; the comprehensive score of the statistical features is calculated; the images are sorted in descending order of the comprehensive score, and a preset proportion of denoised images in the sequence are selected as the screening images.

2. The method as described in claim 1, characterized in that, In step one, the method for obtaining a denoised image by using Gaussian filtering to denoise the original satellite image is as follows: The GaussianBlur function in OpenCV, used in linear smoothing filters, removes Gaussian noise from the original satellite image by specifying the Gaussian kernel size and variance parameter, resulting in a denoised image.

3. The method as described in claim 2, characterized in that, In step one, the GaussianBlur function is: ； In the formula, yes The weight value at that location, These are the coordinates of the pixels in the original satellite image relative to the center of the Gaussian kernel. It is the standard deviation obtained by taking the square root of the variance parameter of the Gaussian distribution, and e is the natural constant.

4. The method as described in claim 3, characterized in that, In step two, the variance, edge density, and gradient direction are calculated as follows: ； ； ； In the formula, Let I be the variance, and I be the current node's image patch. Where is the total number of pixels, and i is the pixel index. For the first grayscale value of each pixel. The average pixel value of the region; For edge density, The gradient magnitude calculated for the Sobel operator; To compute the gradient direction using the Sobel operator, For horizontal gradient components, This represents the vertical gradient component.

5. The method as described in claim 4, characterized in that, In step two, the gradient direction is divided into a preset number of intervals, and the proportion of the main direction is counted to obtain the direction consistency value. ； In the formula, For directional consistency values, the maximum count interval in the directional histogram is: the gradient direction =0°~180° is divided into 8 direction histogram counting intervals, with each interval being 22.5°, and the maximum counting interval in the direction histogram is counted.

6. The method as described in claim 5, characterized in that, In step three, the splitting condition is: ； In the formula, the node size is the size of the current node image block I; Minimum segment size, Let Variance be the constant. for The initial variance, The edge density threshold, This is the directional consistency threshold.

7. The method as described in claim 6, characterized in that, In step three, the steps that trigger the split include: The root node of the quadtree is split into four child nodes, corresponding to the four quadrants: Northwest (NW), Northeast (NE), Southwest (SW), and Southeast (SW). Each child node is further divided into four child nodes until the terminal node corresponding to the task requirement is obtained and marked as a leaf node, at which point the splitting stops.

8. The method as described in claim 7, characterized in that, In step three, the collected statistical characteristics are as follows: ； In the formula, F represents the statistical characteristic; N represents the total number of splits in the quadtree; Maximum depth; This represents the average edge density.

9. The method as described in claim 8, characterized in that, In step three, the calculation method for the comprehensive score of statistical features is as follows: Score=0.5× +0.3×N+0.2× 。 10. The method as described in claim 9, characterized in that, In step three, the maximum depth is calculated as follows: ； In the formula, Where W is the node size, and H is the height.

Images