An intelligent fine-grained target identification method, system, product, device and storage medium on a satellite fusing satellite optical remote sensing imaging characteristics

Abstract

A kind of fusion satellite optical remote sensing imaging characteristics on-board intelligent fine-grained target identification method, system, product, equipment and storage medium, belong to satellite remote sensing image on-orbit identification processing technical field, solve the prior art.Satellite completes shooting task, generates original code stream data, intercepts original code stream data block;Parse the imaging parameters corresponding to original image, correct original image using imaging parameters;Construct target identification model, identify target information in corrected image using target identification model, obtain current target type and target confidence;Compare current target type with knowledge information in knowledge base, according to the comparison result, eliminate false results in current target type, obtain the final target type;Obtain the attitude and orbit parameters of satellite, based on the attitude and orbit parameters, latitude and longitude conversion is carried out on the target, and the latitude and longitude information of the target center point is obtained.The present application is used to realize accurate and reliable on-board intelligent fine-grained target identification method.

Description

Technical Field

[0001] This invention relates to the field of on-orbit recognition and processing technology of satellite remote sensing images, specifically to on-board intelligent fine-grained target recognition methods, systems, products, equipment, and storage media that integrate satellite optical remote sensing imaging characteristics. Background Technology

[0002] To improve coverage efficiency during single overpasses, current visible light satellite constellations commonly employ side-swing imaging technology to expand ground observation range. However, large-angle side-swing significantly impacts image quality: increased angles between the sensor and observed ground features exacerbate geometric distortion, resulting in non-uniform spatial resolution and consequently causing target geometric deformation. Existing constellation systems employ an "on-board imaging-ground processing" model, relying on ground stations to perform in-depth processing of raw data using complex algorithms such as geometric correction and radiometric calibration. This systematically eliminates distortion caused by side-swing, ensuring remote sensing image quality and improving the accuracy of subsequent intelligent identification. However, its complex processing methods only support processing in ground data centers, leading to time limitations and making it difficult to meet the real-time requirements of minute-level responses in military reconnaissance and disaster emergency response.

[0003] Current satellite-based target recognition technologies primarily focus on lightweight models and hardware adaptation. Limited by onboard computing resources, related research is largely concentrated on engineering feasibility verification or single-scenario functionality implementation. For example, the "Tianzhi-1" experimental satellite launched in 2019 only carried a lightweight target detection model for space targets, used to verify the engineering feasibility of onboard processing. It did not address multi-type target recognition in complex terrain scenarios and completely ignored the raw image quality issues caused by satellite side-swing in optical imaging. While the "Jilin-1" Gaofen-03 constellation integrates an onboard intelligent recognition module supporting onboard detection of ship targets, its functionality is limited to coarse-grained detection of a single type of target. It has not been extended to fine-grained target interpretation, nor has it corrected for imaging quality issues such as geometric distortion and resolution inconsistencies caused by side-swing imaging. Although ground-based deep learning models possess powerful fine-grained recognition capabilities, the differences in radiometric and geometric representation between raw onboard data and ground-processed data can still lead to significant performance degradation when algorithms are directly deployed to the onboard edge, resulting in advanced algorithms but poor practical effects. At present, on-orbit image quality correction schemes have not been developed for satellite recognition, and there is a lack of fine-grained image recognition foundation. The algorithms are difficult to balance efficiency and accuracy, and lack the stability and applicability for practical applications.

[0004] In existing technology, Chinese patent document CN117789049A discloses an "On-orbit Image Target Detection Method Based on Spaceborne AI Payload." This method receives mission instructions from a satellite system platform, loads the corresponding mission mode according to the instructions, performs image correction on the image generated by the SAR payload, segments the corrected image into multiple sub-images, identifies targets in each sub-image according to the loaded mission mode, analyzes the target identification results using satellite status information and SAR payload imaging information to obtain target position and heading information, and transmits the target identification results, target position information, and target heading information to the onboard unit. However, this technical solution does not consider the nonlinear geometric distortion and target shape distortion problems caused by changes in imaging geometry due to large-angle side-swing observations by visible light satellites.

[0005] In summary, existing technologies suffer from the following problems: ground-based processing systems cannot meet the timeliness requirements of real-time application scenarios, and the low quality and limited functionality of onboard intelligent recognition input images lead to a decrease in target recognition accuracy. Summary of the Invention

[0006] This invention solves the technical problems of existing technologies, such as the inability of ground-based processing systems to meet the timeliness requirements of real-time application scenarios, and the low quality and limited functionality of on-board intelligent recognition input images, which lead to a decrease in target recognition accuracy.

[0007] The present invention discloses an on-board intelligent fine-grained target recognition method that integrates satellite optical remote sensing imaging characteristics, comprising the following steps: Step 1: After the satellite completes its imaging mission, it generates raw bitstream data and extracts blocks of the raw bitstream data. Step 2: Decode the captured raw bitstream data blocks into the raw image, parse the imaging parameters corresponding to the raw image, and use the imaging parameters to correct the raw image to obtain the corrected image. Step 3: Construct a target recognition model, use the target recognition model to identify target information in the corrected image, and obtain the current target type and target confidence level; Step 4: Compare the current target type with the knowledge information in the knowledge base. Based on the comparison results, filter out the erroneous results in the current target type to obtain the final target type. Step 5: Obtain the satellite's attitude and orbit parameters, and perform latitude and longitude conversion on the target based on the attitude and orbit parameters to obtain the latitude and longitude information of the target's center point; Step 6: Summarize the final target type, target confidence level, comparison results, and latitude and longitude information of the target center point, and output the target information.

[0008] Furthermore, in one embodiment of the present invention, the extraction of the original bitstream data block in step 1 specifically includes: A preset standard image processing unit size is used. Based on this size, raw bitstream data blocks are sequentially extracted from a specified position in the buffer. If the raw bitstream data size is smaller than the standard image processing unit size, invalid values ​​are filled into the raw bitstream data and it is marked as an invalid block.

[0009] Furthermore, in one embodiment of the present invention, the imaging parameters in step 2 include orbital altitude. Pixel size , side swing angle along the track focal length and vertical rail focal length .

[0010] Furthermore, in one embodiment of the present invention, step 2, which involves correcting the original image using imaging parameters, specifically includes: The track-side resolution and vertical resolution corresponding to each viewpoint in the original image are calculated separately. The maximum and minimum values ​​of the calculated track-side resolution and vertical resolution are averaged to obtain the output size. The original image is then resampled based on the output size to obtain the corrected image.

[0011] Furthermore, in one embodiment of the present invention, the target recognition model in step 3 adopts a single-channel type, and the target recognition model is an improvement on YOLOv11, specifically: Change the C2f module to the C3K2 module, and add the C2PSA module after the SPPF module.

[0012] Furthermore, in one embodiment of the present invention, step 5, which involves converting the latitude and longitude of the target based on the attitude and orbit parameters, specifically includes: The Coordinated Universal Time (UTC) at which the satellite completed its imaging mission is converted to Julian Day (JD). Greenwich Mean Time (GMST) is calculated based on Julian Day. The rotation matrix of ECF and ECI is calculated using GMST to obtain the satellite's current position in the ECF coordinate system. Based on the satellite's current position in the ECF coordinate system, the ECF coordinates of the target's center point are calculated. The target's ECF coordinates are then converted into the latitude and longitude of the target's center point.

[0013] The present invention discloses an on-board intelligent fine-grained target recognition system that integrates satellite optical remote sensing imaging characteristics. The system is constructed based on the aforementioned method and includes the following modules: Module 1: After the satellite completes its imaging mission, it generates raw bitstream data and extracts blocks of the raw bitstream data. Module 2 decodes the captured raw bitstream data blocks into raw images, parses the imaging parameters corresponding to the raw images, and uses the imaging parameters to correct the raw images to obtain corrected images. Module 3: Construct a target recognition model, use the target recognition model to identify target information in the corrected image, and obtain the current target type and target confidence level; Module 4 compares the current target type with the knowledge information in the knowledge base, and based on the comparison results, filters out erroneous results in the current target type to obtain the final target type; Module 5 acquires the satellite's attitude and orbit parameters, performs latitude and longitude conversion on the target based on the attitude and orbit parameters, and obtains the latitude and longitude information of the target's center point; Module 6 summarizes the final target type, target confidence level, comparison results, and latitude and longitude information of the target center point, and outputs the target information.

[0014] The computer program product of the present invention includes a computer program or instructions, which, when executed by a processor, implement any of the above-described on-board intelligent fine-grained target recognition methods that fuse satellite optical remote sensing imaging characteristics.

[0015] The electronic device of the present invention includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus. Memory, used to store computer programs; The processor, when executing a program stored in memory, implements any of the above-described on-board intelligent fine-grained target recognition methods that fuse satellite optical remote sensing imaging characteristics.

[0016] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-described on-board intelligent fine-grained target recognition methods that fuse satellite optical remote sensing imaging characteristics.

[0017] This invention solves the technical problems of existing technologies, where ground-based processing systems cannot meet the timeliness requirements of real-time application scenarios, and where the low quality and limited functionality of on-board intelligent recognition input images lead to decreased target recognition accuracy. Specific beneficial effects include: This invention proposes an on-board intelligent fine-grained target recognition method that integrates satellite optical remote sensing imaging characteristics. Addressing the real-time processing needs of high-timeliness applications, this method deeply integrates satellite optical remote sensing imaging characteristics to intelligently process raw side-swing imaging data directly on-board. This method breaks through the traditional "ground-based centralized processing" model, eliminating the need to migrate computationally intensive image production processes to edge devices. Under the limited computing power of the satellite, it achieves adaptive compensation for side-swing distortion in the original image, significantly suppressing target shape distortion caused by side-swing. It ensures the input image quality of the intelligent recognition algorithm at the raw data level, improving its adaptability in the field of fine-grained recognition, and ultimately enhancing the accuracy and reliability of on-board real-time processing. Attached Figure Description

[0018] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the on-board intelligent fine-grained target recognition method described in Implementation Method 1; Figure 2 This is a flowchart of the on-board intelligent fine-grained target recognition method described in Implementation Method 1; Figure 3 This is a schematic diagram of the side-swing-free imaging described in Embodiment 3; Figure 4 This is a schematic diagram of the side-swing imaging described in Embodiment 3; Figure 5 This is a schematic diagram of the oil tanker case described in Example 1; Figure 6 This is a schematic diagram of the container ship example described in Example 1; Figure 7 It is the improved YOLOv11 model structure described in Implementation Method 4. Detailed Implementation

[0019] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0020] Implementation Method 1: An on-board intelligent fine-grained target recognition method integrating satellite optical remote sensing imaging characteristics, comprising the following steps: Step 1: After the satellite completes its imaging mission, it generates raw bitstream data and extracts blocks of the raw bitstream data. Step 2: Decode the captured raw bitstream data blocks into the raw image, parse the imaging parameters corresponding to the raw image, and use the imaging parameters to correct the raw image to obtain the corrected image. Step 3: Construct a target recognition model, use the target recognition model to identify target information in the corrected image, and obtain the current target type and target confidence level; Step 4: Compare the current target type with the knowledge information in the knowledge base. Based on the comparison results, filter out the erroneous results in the current target type to obtain the final target type. In this implementation, after the recognition model completes inference and the target image is processed according to the corrected image, the target's knowledge information, including pixel length, pixel width, and aspect ratio, is compared with standard knowledge within permissible limits based on the current target recognition type. The degree of matching helps confirm the target's identity, avoiding false positives and false negatives.

[0021] Step 5: Obtain the satellite's attitude and orbit parameters, and perform latitude and longitude conversion on the target based on the attitude and orbit parameters to obtain the latitude and longitude information of the target's center point; Step 6: Summarize the final target type, target confidence level, comparison results, and latitude and longitude information of the target center point, and output the target information.

[0022] In this embodiment, after completing the latitude and longitude conversion of the target, all information of the target is summarized, including target type, target center point longitude, target center point latitude, target confidence level, target length and target width, etc. The information is stored in the form of binary code stream, and when the satellite enters the mission orbit, it is transmitted back to the ground station via data transmission.

[0023] Visible light satellite constellations expand coverage through side-swing imaging, but large-angle side-swing can cause geometric distortion and resolution inconsistencies. Existing systems rely on complex corrections from ground stations, making it difficult to meet minute-level real-time response requirements. On-board recognition technology, limited by computing power, only achieves coarse detection of single targets, failing to address the issue of image quality degradation. This leads to a significant decrease in the performance of advanced ground-based algorithms deployed on satellites, and there is a lack of an integrated on-orbit correction-recognition solution that balances efficiency and accuracy.

[0024] To address the aforementioned problems, this embodiment proposes an on-board intelligent fine-grained target recognition method that integrates satellite optical remote sensing imaging characteristics, such as... Figure 1 and Figure 2 As shown, this method can reduce the impact of side-swing imaging on the original image during satellite on-orbit recognition processing, alleviate target shape distortion, effectively improve the accuracy of target recognition algorithms under large side-swing conditions, and expand the application prospects of on-board recognition processing. It solves the technical problems of existing technologies where ground-based processing systems cannot meet the timeliness requirements of real-time application scenarios, and where the low quality and limited functionality of on-board intelligent recognition input images lead to decreased target recognition accuracy.

[0025] Implementation Method Two: This implementation method further defines the on-board intelligent fine-grained target recognition method that integrates satellite optical remote sensing imaging characteristics as described in Implementation Method One. Specifically, step 1 involves extracting the original bitstream data block as follows: A preset standard image processing unit size is used. Based on this size, raw bitstream data blocks are sequentially extracted from a specified position in the buffer. If the raw bitstream data size is smaller than the standard image processing unit size, invalid values ​​are filled into the raw bitstream data and it is marked as an invalid block.

[0026] In this embodiment, after the satellite completes the shooting tasks specified in the predetermined mission plan and instructions, the raw image data generated by its camera is first stored in the onboard solid-state storage unit in the form of a binary stream file. This raw data contains complete sensor output information (such as pixel values, auxiliary parameter headers, etc.), but has not yet undergone any filtering or segmentation processing, and needs to be extracted in a targeted manner according to the specific needs of the subsequent intelligent recognition module.

[0027] When the onboard intelligent processing unit receives the identification command from the ground, the system will accurately locate the target data file from the solid-state storage unit, read the file into the buffer through the high-speed serial transmission interface, and the identification module will extract continuous binary code stream segments from the buffer in sequence according to the pre-set standard image processing unit size.

[0028] Specifically, the module first parses the bitstream header information, confirms that it matches the preset processing requirements, and then reads data from the beginning of the buffer or a specified offset position, using fixed-size "data blocks" as units, such as the binary bitstream length corresponding to a 1024×1024 pixel block. If the original image size is smaller than a standard unit, a padding mechanism is triggered (marking invalid blocks).

[0029] Implementation Method 3: This implementation method further defines the on-board intelligent fine-grained target recognition method that integrates satellite optical remote sensing imaging characteristics described in Implementation Method 1. In step 2, the original image is corrected using imaging parameters, specifically as follows: The track-side resolution and vertical resolution corresponding to each viewpoint in the original image are calculated separately. The maximum and minimum values ​​of the calculated track-side resolution and vertical resolution are averaged to obtain the output size. The original image is then resampled based on the output size to obtain the corrected image.

[0030] In this embodiment, the identification module parses remote sensing image imaging parameters from the auxiliary data of the central unit. The parsed parameters include orbital height. Pixel size , side swing angle along the track focal length and vertical rail focal length .

[0031] Image correction aims to address distortion issues caused by resolution changes resulting from lateral tilting during satellite imaging. It involves calculating the true pixel resolution for each pixel using the analyzed parameters and then re-outputting the corrected image.

[0032] During remote sensing satellite side-tilt imaging, the camera's line of sight is laterally deflected relative to the direction perpendicular to the satellite's flight path. It no longer points perpendicularly at the ground target but tilts at a certain angle. This increases the actual distance between the satellite and the target, making it appear farther than when observed perpendicularly. Consequently, a target of the same size on the ground occupies fewer pixels on the imaging sensor, leading to target shape distortion. The imaging illustrations of a satellite with and without side-tilt are shown below. Figure 3 and Figure 4 As shown.

[0033] In the case of lateral tilt, the method for calculating the true resolution of each pixel is as follows, where the field point position is A( : ; ; in, For connecting ground targets and satellites, Let L be the projected length of the line connecting ground point A and the satellite on the optical axis. This is the satellite's nominal altitude. The mean radius of curvature of the Earth. These are the parameters for correcting the yaw angle.

[0034] ; ; ; ; in, and These represent the along-rail resolution and the perpendicular-rail resolution corresponding to a certain point in the image plane, respectively. By calculating the true resolution of each point, the size of the new image can be calculated. and These are the optical focal lengths of the camera in the x and y directions, respectively. and These are the field of view correction values ​​in the x and y directions, respectively.

[0035] Because the resolution of each viewpoint is uneven, irregular two-dimensional matrices cannot be directly formed into image matrices for subsequent deep learning models to recognize and process. Processing these matrices using complex methods is computationally impractical given the limitations of satellite edge detection. Considering the generalization capabilities of deep learning models, a certain degree of error in the image itself is acceptable. Therefore, the mean value of the maximum and minimum values ​​of pixels in different directions in the image is used to correct the output size of the image. The corresponding steps are as follows: Calculate the true cell length of different rows and columns, the first... The actual cell size of the row: ; No. The actual cell size of the column: ; The corrected image size is calculated as follows: ; ; in, For the first GSD mean of all cells in the row, For the first The GSD mean of all cells. and These represent the width and height of the corrected image, respectively.

[0036] Based on the output size of the corrected image, the original image is resampled using OpenCV functions. Once interpolation is completed, the corrected image with reduced target shape distortion can be output.

[0037] Implementation Method Four: This implementation method further defines the on-board intelligent fine-grained target recognition method that integrates satellite optical remote sensing imaging characteristics as described in Implementation Method One. The target recognition model in step 3 adopts a single-channel type, and the target recognition model is an improvement on YOLOv11, specifically: Change the C2f module to the C3K2 module, and add the C2PSA module after the SPPF module.

[0038] In this embodiment, the deep learning model recognition process is completed on the edge device of the satellite. This process differs from traditional ground-based recognition and requires optimization strategies for various aspects such as satellite computing power, satellite data, and satellite deployment.

[0039] To address the issue of limited computing power on satellites, a single-stage framework recognition model with faster inference speed was adopted; to address the single-channel nature of satellite data, the target recognition model was modified from three channels to a single channel to reduce the amount of input data computation; to address the deployment issues on satellites, the model was converted into an Offline Model structure for efficient inference on Ascend hardware.

[0040] In terms of the inference model, the single-stage model used is an improvement based on YOLOv11, such as... Figure 7 As shown, the C2f module was modified to a C3K2 module, and a C2PSA module was added after the SPPF module. The C3K2 module has a simple structure, enhances feature diversity through dual convolutional kernels, and removes the gradient flow in the C2f module, making it suitable for edge platform processing. The C2PSA module combines the CSP (Cross Stage Partial) structure and the PSA (Pyramid Squeeze Attention) attention mechanism, aiming to improve the model's multi-scale feature extraction capability, and is used to adapt to targets with certain shape changes in the original satellite images.

[0041] General target recognition models are designed for processing three-channel RGB natural image data. However, considering that satellite data is panchromatic and single-channel, the model is modified from a three-channel to a single-channel type. Theoretically, replicating panchromatic images as pseudo-three channels is acceptable for the model, but it is redundant and increases the load on edge computing power without information gain. The single-channel model adapts to real data sources, requires no data conversion, reduces forward computation by about two-thirds, and improves inference speed and computational efficiency.

[0042] When deploying models to edge devices, standard pt and pth format model weight files cannot be used. Furthermore, to enable the model to run on the domestically produced Huawei Ascend AI processor, it needs to be converted to the Ascend offline model om format. The om format is a high-performance inference format optimized specifically for Ascend chips. Its high performance, low latency, and independence from native frameworks provide a standard for AI deployment and inference on satellite edge devices.

[0043] Implementation Method 5: This implementation method further defines the on-board intelligent fine-grained target recognition method that integrates satellite optical remote sensing imaging characteristics as described in Implementation Method 1. Specifically, step 5 involves converting the latitude and longitude of the target based on attitude and orbit parameters: The Coordinated Universal Time (UTC) at which the satellite completed its imaging mission is converted to Julian Day (JD). Greenwich Mean Time (GMST) is calculated based on Julian Day. The rotation matrix of ECF and ECI is calculated using GMST to obtain the satellite's current position in the ECF coordinate system. Based on the satellite's current position in the ECF coordinate system, the ECF coordinates of the target's center point are calculated. The target's ECF coordinates are then converted into the latitude and longitude of the target's center point.

[0044] In this embodiment, the conversion of latitude and longitude on-orbit requires data such as the satellite's time and position information to convert the satellite's attitude and orbit information at a certain moment into the geographic information of the ground target. This process involves the transformation of multiple coordinate systems, the application of orbital dynamics models, and geometric calculations.

[0045] The complete process for coordinate system transformation using attitude parameters is as follows: 1. Time processing: First, the mission time is converted. The satellite uses Coordinated Universal Time (UTC), a globally unified time standard based on atomic seconds. Time processing involves converting UTC to Julian Day (JD), a timekeeping method that calculates time in consecutive days within the Julian cycle, facilitating astronomical calculations.

[0046] ; ; ; Then, Greenwich Sidereal Time (GMST) needs to be calculated based on the Julian day. The R2P5 function takes the Julian day as input, and if the decimal part is greater than or equal to 0.5, takes its integer part and increments it by 0.5; if the decimal part is less than 0.5, takes its integer part and subtracts 0.5, which is then used as the result. The value of .

[0047] ; ; ; ; ; ; 0.01745329251994329576924; Finally, GMST is used to calculate satellite orbital parameters in astrometry, and throughout the process, it is used to calculate the rotation from the Earth-Fixed Frame (ECF) to the Inertial Frame (ECI).

[0048] 2. Earth-fixed system ECF to ECI inertial frame: The rotation matrix for the transformation between the two can be calculated using GMST, and this matrix can be used to transform the position and velocity vectors in the Earth-Fixed frame to the inertial frame. The ECF to ECI transformation only requires a rotation about the Z-axis.

[0049] The position conversion method is as follows: ; ; The speed conversion method is as follows: ; in, Velocity in ECI coordinate system The angular velocity of Earth's rotation. Let Z be the rotation matrix about the Z-axis. For rotation angle, , and These are the X, Y, and Z components of the position vector in the ECI coordinate system. , and These are the X, Y, and Z components of the position vector in the ECF coordinate system, respectively.

[0050] 3. Track extrapolation: Using the orbital dynamics model, the satellite's position and velocity at the GPS time are extrapolated to the mission time to calculate the satellite's new position and velocity under ECF at the current time.

[0051] ; ; ; in, The velocity is in the ECF coordinate system.

[0052] 4. Comprehensive transformation: After calculating the ECF position of the satellite at the mission time, based on information such as the sensor mounting matrix, attitude quaternions, and side-swing angle, a unit direction vector ρ pointing from the satellite to the ground is constructed. The ray distance from the satellite to the target point is then calculated. The ECF coordinates of the target point are calculated using geometric solutions. The formula is as follows: ; ; in, The satellite's position in the ECF coordinate system. The ECF coordinates of the target center point.

[0053] 5. Coordinate transformation: After obtaining the ECF coordinates of the target point, This represents the distance from the target point to the Earth's center. The latitude and longitude conversion formula is as follows: ; ; in, and These are the latitude and longitude of the target point, respectively. , , These are the three coordinate components of the target point in the ECF coordinate system.

[0054] This implementation aims to integrate the characteristics of satellite optical remote sensing imaging, reduce the target shape distortion problem in the case of side-swing when the original image is processed on orbit, and improve the ability of on-board intelligent fine-grained target recognition.

[0055] Example 1: This embodiment uses remote sensing image data as the satellite data source, and the satellites used all have sub-meter resolution. The above method uses a sample of naval vessels as an example to illustrate the effectiveness of the correction method fused with satellite optical remote sensing imaging features. Figure 5 and Figure 6 As shown, the left side represents the target in its normal state, the middle side represents the target with shape distortion, and the right side represents the corrected state.

[0056] Its specific experimental performance is as follows: Table 1 Accuracy Indicators

[0057] Table 2 Speed ​​Indicators

[0058] In terms of algorithm accuracy, a total of 179 remote sensing panoramic images with 220 targets were tested in the satellite maritime target sample. The recall rate improved from 73.3% to 89.1%, and the precision improved from 72.1% to 83.5%. In terms of algorithm speed, the performance improved from 5 FPS to 5.6 FPS.

[0059] In summary, the on-board intelligent fine-grained target recognition method that integrates satellite optical remote sensing imaging characteristics proposed in this embodiment can effectively improve the accuracy and speed of on-orbit intelligent target recognition processing.

[0060] Implementation Method Six: This implementation method describes an on-board intelligent fine-grained target recognition system that integrates satellite optical remote sensing imaging characteristics. The system is constructed based on the method described in Implementation Method One and includes the following modules: Module 1: After the satellite completes its imaging mission, it generates raw bitstream data and extracts blocks of the raw bitstream data. Module 2 decodes the captured raw bitstream data blocks into raw images, parses the imaging parameters corresponding to the raw images, and uses the imaging parameters to correct the raw images to obtain corrected images. Module 3: Construct a target recognition model, use the target recognition model to identify target information in the corrected image, and obtain the current target type and target confidence level; Module 4 compares the current target type with the knowledge information in the knowledge base, and based on the comparison results, filters out erroneous results in the current target type to obtain the final target type; Module 5 acquires the satellite's attitude and orbit parameters, performs latitude and longitude conversion on the target based on the attitude and orbit parameters, and obtains the latitude and longitude information of the target's center point; Module 6 summarizes the final target type, target confidence level, comparison results, and latitude and longitude information of the target center point, and outputs the target information.

[0061] Implementation Method Seven: A computer program product, comprising a computer program or instructions, which, when executed by a processor, implements any of the above-described on-board intelligent fine-grained target recognition methods that integrate satellite optical remote sensing imaging characteristics.

[0062] Implementation Method 8: An electronic device, comprising a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus; Memory, used to store computer programs; The processor, when executing a program stored in memory, implements any of the above-described on-board intelligent fine-grained target recognition methods that fuse satellite optical remote sensing imaging characteristics.

[0063] Implementation Method Nine: A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements any of the above-described on-board intelligent fine-grained target recognition methods that integrate satellite optical remote sensing imaging characteristics.

[0064] The above provides a detailed description of the on-board intelligent fine-grained target identification method, system, product, device, and storage medium that integrates satellite optical remote sensing imaging characteristics proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A satellite-based intelligent fine-grained target recognition method integrating satellite optical remote sensing imaging characteristics, characterized in that, Includes the following steps: Step 1: After the satellite completes its imaging mission, it generates raw bitstream data and extracts blocks of the raw bitstream data. Step 2: Decode the captured raw bitstream data blocks into the raw image, parse the imaging parameters corresponding to the raw image, and use the imaging parameters to correct the raw image to obtain the corrected image. Step 3: Construct a target recognition model, use the target recognition model to identify target information in the corrected image, and obtain the current target type and target confidence level; Step 4: Compare the current target type with the knowledge information in the knowledge base. Based on the comparison results, filter out the erroneous results in the current target type to obtain the final target type. Step 5: Obtain the satellite's attitude and orbit parameters, and perform latitude and longitude conversion on the target based on the attitude and orbit parameters to obtain the latitude and longitude information of the target's center point; Step 6: Summarize the final target type, target confidence level, comparison results, and latitude and longitude information of the target center point, and output the target information.

2. The on-board intelligent fine-grained target recognition method based on the fusion of satellite optical remote sensing imaging characteristics as described in claim 1, characterized in that, The step 1 of extracting the original bitstream data block specifically involves: A preset standard image processing unit size is used. Based on this size, raw bitstream data blocks are sequentially extracted from a specified position in the buffer. If the raw bitstream data size is smaller than the standard image processing unit size, invalid values ​​are filled into the raw bitstream data and it is marked as an invalid block.

3. The on-board intelligent fine-grained target recognition method based on the fusion of satellite optical remote sensing imaging characteristics as described in claim 1, characterized in that, The imaging parameters in step 2 include orbital height. Pixel size , side swing angle along the track focal length and vertical rail focal length .

4. The on-board intelligent fine-grained target recognition method based on the fusion of satellite optical remote sensing imaging characteristics according to claim 1 or 3, characterized in that, In step 2, the original image is corrected using imaging parameters, specifically as follows: The track-side resolution and vertical resolution corresponding to each viewpoint in the original image are calculated separately. The maximum and minimum values ​​of the calculated track-side resolution and vertical resolution are averaged to obtain the output size. The original image is then resampled based on the output size to obtain the corrected image.

5. The on-board intelligent fine-grained target recognition method based on the fusion of satellite optical remote sensing imaging characteristics according to claim 1, characterized in that, The target recognition model in step 3 is a single-channel type, and the target recognition model is an improvement on YOLOv11, specifically: Change the C2f module to the C3K2 module, and add the C2PSA module after the SPPF module.

6. The on-board intelligent fine-grained target recognition method based on the fusion of satellite optical remote sensing imaging characteristics according to claim 1, characterized in that, Step 5 involves converting the latitude and longitude of the target based on its attitude and orbit parameters, specifically as follows: The Coordinated Universal Time (UTC) at which the satellite completed its imaging mission is converted to Julian Day (JD). Greenwich Mean Time (GMST) is calculated based on Julian Day. The rotation matrix of ECF and ECI is calculated using GMST to obtain the satellite's current position in the ECF coordinate system. Based on the satellite's current position in the ECF coordinate system, the ECF coordinates of the target's center point are calculated. The target's ECF coordinates are then converted into the latitude and longitude of the target's center point.

7. A satellite-based intelligent fine-grained target recognition system integrating satellite optical remote sensing imaging characteristics, said system being constructed based on the method described in claim 1, characterized in that... Includes the following modules: Module 1: After the satellite completes its imaging mission, it generates raw bitstream data and extracts blocks of the raw bitstream data. Module 2 decodes the captured raw bitstream data blocks into raw images, parses the imaging parameters corresponding to the raw images, and uses the imaging parameters to correct the raw images to obtain corrected images. Module 3: Construct a target recognition model, use the target recognition model to identify target information in the corrected image, and obtain the current target type and target confidence level; Module 4 compares the current target type with the knowledge information in the knowledge base, and based on the comparison results, filters out erroneous results in the current target type to obtain the final target type; Module 5 acquires the satellite's attitude and orbit parameters, performs latitude and longitude conversion on the target based on the attitude and orbit parameters, and obtains the latitude and longitude information of the target's center point; Module 6 summarizes the final target type, target confidence level, comparison results, and latitude and longitude information of the target center point, and outputs the target information.

8. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by the processor, they implement the on-board intelligent fine-grained target recognition method that integrates satellite optical remote sensing imaging characteristics as described in any one of claims 1-6.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; The processor, when executing a program stored in memory, implements the on-board intelligent fine-grained target recognition method according to any one of claims 1-6, which integrates satellite optical remote sensing imaging characteristics.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the on-board intelligent fine-grained target recognition method according to any one of claims 1-6, which integrates satellite optical remote sensing imaging characteristics.

Images