Artifact Removal Method of Low Earth Orbit Image Based on Pulse Coupled Neural Network

Abstract

The present invention discloses a method for rapid artifact removal from low-orbit earth orbit (LEO) images based on a pulse-coupled neural network (PCN). The method comprises the following steps: S1: acquiring original LEO image data; S2: inputting the original LEO image data into a spaceborne hierarchical progressive PCN model; S3: forming candidate artifact regions and generating an artifact probability map based on synchronized pulse connectivity and information entropy increments; S4: generating preprocessed LEO image data and simultaneously generating a reversible repair log; S5: generating a high-quality image product with artifact removal; and S6: updating the coupling gain coefficient and link strength threshold of the spaceborne hierarchical progressive PCN model based on the processing status of the high-quality image product with artifact removal. The system of the present invention has stable closed-loop operation, significantly improved image quality, and processing latency that meets the requirements of rapid remote sensing services, demonstrating engineering feasibility.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method for quickly removing artifacts from low-orbit images based on a pulse coupled neural network. Background Art

[0002] With the rapid growth of remote sensing satellite constellations and the increasing demand for high-revisit imaging, low-Earth orbit (LEO) remote sensing satellites have become the primary platform for acquiring high-resolution surface observation images. However, due to their relatively low orbital altitudes (typically between 300 and 1000 km), they are susceptible to multiple interference sources in high-frequency imaging and high-dynamic scenes. This often results in various image artifacts in the imaging data, such as banding, high-energy particle breakdown speckling, row and column misalignment, compression errors, and image jitter caused by attitude disturbances. These artifacts not only severely interfere with the identification and analysis of ground object information but also cause systematic errors in subsequent image interpretation, classification, and inversion applications.

[0003] Currently, mainstream image artifact removal methods are mainly divided into two categories: real-time removal onboard spacecraft and post-processing on the ground. Most onboard removal methods use fixed threshold filtering, directional template filtering, or median sliding window processing. Due to limited computing resources, it is difficult to deploy complex deep networks or image segmentation models, resulting in failure in high-dynamic-contrast scenes or areas with complex textures. While ground-based deep learning artifact repair methods are effective, they are prone to over-smoothing of image details, structural distortion, or texture inconsistencies due to the fragmented structure of the space-ground model and the complex heavy compression processing chain. Furthermore, they experience high processing latency, making it difficult to meet the demand for low-latency data quality assurance in disaster monitoring and emergency response scenarios.

[0004] Furthermore, traditional methods generally lack an end-to-end collaborative mechanism for artifact removal. This means that onboard processing results cannot effectively guide fine-grained reconstruction on the ground side, and the ground side cannot provide feedback to adjust onboard processing parameters, making closed-loop optimization of the overall removal strategy difficult. This is especially true when artifacts are sparse and irregularly distributed, scale-crossing, and have complex texture overlap. Existing processing frameworks struggle to balance robustness, real-time performance, and interpretability.

[0005] Therefore, there is an urgent need for a fast image artifact removal method for low-orbit remote sensing satellites that is real-time, reversible, and has satellite-ground closed-loop adaptability to improve image quality consistency and the intelligence level of the processing link. Summary of the Invention

[0006] One purpose of the present invention is to propose a method for quickly removing artifacts from low-orbit images based on a pulse-coupled neural network. The system of the present invention has stable closed-loop operation, significantly improved image quality, and processing delay that meets the requirements of fast remote sensing services, and has engineering feasibility.

[0007] According to an embodiment of the present invention, a method for rapidly removing artifacts from low-orbit images based on a pulse coupled neural network comprises the following steps:

[0008] S1. Obtaining original low-orbit image data;

[0009] S2. Input the original LEO image data into the onboard hierarchical progressive attention adaptive pulse-coupled neural network model, and perform pulse traversal based on the synchronization signal of the original LEO image to generate a neuron firing time series;

[0010] S3. Construct a spike-fire mapping matrix based on the neuron firing time series, use the spike-fire mapping matrix to calculate the neighborhood spike synchrony, form an artifact candidate region, and generate an artifact probability map based on the neighborhood spike synchrony and information entropy increment;

[0011] S4. Construct a block-level weight mask based on the artifact probability map, use the block-level weight mask to perform pixel restoration on the artifact candidate area, generate pre-processed low-orbit image data, and simultaneously generate a reversible restoration log;

[0012] S5. The pre-processed LEO image data, the pulse fire mapping matrix, the block-level weight mask, and the reversible restoration log are transmitted to the ground enhancement unit. The ground enhancement unit reversibly expands the artifact candidate regions based on the pulse fire mapping matrix and the reversible restoration log, and performs fine-grained texture reconstruction in combination with the block-level weight mask to generate artifact-removed high-quality image products.

[0013] S6. Update the coupling gain coefficient and link strength threshold of the onboard hierarchical progressive attention adaptive pulse coupled neural network model according to the processing status of the artifact-removed high-quality image product, and return the updated parameters to the onboard hierarchical progressive attention adaptive pulse coupled neural network model.

[0014] Optionally, the S1 includes the following steps:

[0015] S11. Construct a raw low-orbit image data sequence, where each frame represents a visible light image or multispectral image acquired by a low-orbit remote sensing satellite at a specific moment. The raw low-orbit image data sequence consists of multiple image frames acquired at different time points, and each frame is labeled with the corresponding acquisition time information;

[0016] S12. Obtain the attitude quaternion and orbital position of the low-orbit remote sensing satellite corresponding to the moment of acquisition of each frame of image. The attitude quaternion and orbital position together constitute the data basis required for attitude synchronization calibration. The attitude quaternion and orbital position are used to perform attitude synchronization calibration on the original low-orbit image data sequence. The attitude synchronization calibration is used to align the image frame with the satellite attitude information;

[0017] S13. Perform preliminary radiometric calibration on the image frame after attitude synchronization calibration, which is used to correct the brightness response deviation of the original pixel grayscale value;

[0018] S14. After completing the preliminary radiometric calibration, perform strip equalization preprocessing on the image frame to remove large-scale directional brightness unevenness features to obtain the final original low-orbit image data.

[0019] Optionally, the S2 includes the following steps:

[0020] S21. Input the final raw low-orbit image data into the onboard hierarchical progressive attention adaptive pulse coupled neural network model. The onboard hierarchical progressive attention adaptive pulse coupled neural network model includes a global coarse-scale PCNN layer, a mid-scale PCNN layer, and a local fine-scale attention PCNN layer. The three layers of PCNN are cascaded and based on the root mean square value of the low-orbit remote sensing satellite attitude disturbance and the imaging system instantaneous gain Initialize the global coarse-scale layer parameters:

[0021] ;

[0022] in, 、 、 is the onboard reference parameter, 、 、 is the proportional adjustment factor, is the amplitude of attitude disturbance change in the rolling observation window, is the global initial coupling gain coefficient, is the global initial adaptive time decay factor, is the global initial link strength threshold, is the reference standard value of attitude disturbance of low-orbit remote sensing satellite, is the nominal gain value of the imaging system;

[0023] S22. In the global coarse-scale PCNN layer, the adaptive time decay factor Drive the neuron internal state update and calculate the global grayscale statistical variance of the final original low-orbit image data Adaptively determine the dynamic triggering conditions of neurons and generate global coarse-scale neuron pulse outputs ;

[0024] S23. Initialize the mesoscale coupling gain coefficient in the mesoscale PCNN layer and the mesoscale initial link strength threshold , and calculate the local entropy change based on the pulse output of the global coarse-scale PCNN layer Obtaining the mesoscale dynamic coupling gain coefficient , according to the mesoscale dynamic coupling gain coefficient Dynamically adjust the trigger sensitivity of mesoscale neurons to form mesoscale pulse output ;

[0025] S24. Initialize the local scale initial link strength threshold in the local fine-scale attention PCNN layer , calculate the local attention coefficient based on the pulse output of the mid-scale PCNN layer , and use the local attention coefficient Dynamic triggering thresholds for local-scale neurons Perform real-time adaptive updates to form local fine-scale pulse outputs, enabling the positioning of complex fine-scale artifacts such as stripes, scintillation points, compression errors, and particle interference in low-orbit images;

[0026] S25. An inter-layer attention interaction mechanism is established between the global coarse-scale PCNN layer, the medium-scale PCNN layer, and the local fine-scale attention PCNN layer. The pulse output of the global coarse-scale PCNN layer guides the parameter update of the medium-scale PCNN layer, and the pulse output of the medium-scale PCNN layer guides the parameter update of the local fine-scale attention PCNN layer. The response results of the local fine-scale attention PCNN layer act reversely on the medium-scale and global coarse-scale PCNN layers, dynamically optimizing the trigger sensitivity and coupling strength of each layer, realizing a coarse-scale-medium-scale-fine-scale and fine-scale-medium-scale-coarse-scale bidirectional interactive closed loop, and finally generating a hierarchical progressive neuron excitation time series matrix.

[0027] Optionally, S3 includes the following steps:

[0028] S31. Constructing a pulse-fire mapping matrix based on the hierarchical progressive neuron excitation time series matrix ,The pulse fire mapping matrix is ​​obtained by mapping the firing time of each pixel from the hierarchical ,neuron firing time series matrix to a two-dimensional matrix consistent with the ,image coordinates;

[0029] S32. Based on the pulse fire mapping matrix, the neighborhood pulse synchronization of each pixel position within the receptive field is calculated. The neighborhood pulse synchronization is obtained by counting the relative amplitude of the difference between the first pulse triggering time of the pixel and all pixels in the neighborhood. The neighborhood pulse synchronization of all positions constitutes a synchronous pulse connectivity graph. , the synchronous spike connectivity map is used to characterize the spatial consistency distribution of the neuronal firing temporal structure of the original LEO image data:

[0030] ;

[0031] in, Pixel position Central neighborhood, is the total number of pixels in the neighborhood, Neighborhood pixels At the first pulse triggering moment in the hierarchical progressive attention adaptive pulse coupled neural network model, is the neighborhood pixel position At the first pulse triggering moment in the hierarchical progressive attention adaptive pulse coupled neural network model, is the synchronization penalty function, which reflects the degree of inhibition of connectivity caused by the difference in excitation time. Indicates pixel position The strength of the consistency with the neighboring pulse sequence;

[0032] S33. According to the synchronous pulse connection diagram The neighborhood pulse synchronization of each pixel position in the , and filter out all the neighborhood pulse synchronizations below the synchronization threshold The pixel position of The preliminary candidate region of the artifact is used to identify the potential abnormal region in which the neuronal firing timing structure is obviously separated from the surrounding structure in the original low-orbit image data. The synchronization threshold A fixed empirical threshold for demarcating normal and abnormal synchrony;

[0033] S34. Calculate the local information entropy increment for each pixel position in the preliminary candidate area of ​​the artifact ,The local information entropy increment is obtained by calculating the difference between the local Shannon entropy in the neighborhood of the pixel position and the average local Shannon entropy of its corresponding synchronously connected regions;

[0034] S35. Generate an artifact probability map by combining the neighborhood pulse synchronization and local information entropy increment of each pixel position The artifact probability is obtained by linearly weighting the neighborhood pulse synchronization and the local information entropy increment, where the neighborhood pulse synchronization is used to describe the abnormality of the timing structure and the local information entropy increment is used to describe the abnormality of the grayscale structure.

[0035] Optionally, the S4 includes the following steps:

[0036] S41. Divide the original LEO image data into multiple fixed-size image blocks according to the artifact probability map and construct a block-level weight mask ,in Represents the image block coordinates, each image block contains a set of pixels Corresponding block-level weight mask ;

[0037] S42. Weight mask at block level On this basis, the artifact restoration strength threshold is set, and all image blocks with weight values ​​greater than the artifact restoration strength threshold are defined as the restoration target block area set. ;

[0038] S43. For each set of repair target block areas Image blocks are constructed, and pixel-level one-to-one corresponding indexes are established based on the pulse fire mapping matrix. Pixel-level reversible repair operations are performed to obtain the repaired pixel values. ,The restored pixel value is obtained by weighted fusion of the original pixel and the neighborhood reference value according to the artifact probability;

[0039] S44. While performing pixel restoration, a reversible restoration log is generated, recording the position index, original pixel value, restored pixel value, and image block index of all restored pixels;

[0040] S45. Set all restored pixel values Replace the pixel values ​​at the corresponding positions in the original LEO image data to generate pre-processed LEO image data , whose structure remains consistent with the input image frame.

[0041] Optionally, the S5 includes the following steps:

[0042] S51. The pre-processed low-orbit image data, the pulse fire mapping matrix, the block-level weight mask, and the reversible repair log are transmitted to the ground enhancement unit. The ground enhancement unit receives the corresponding processing information packet for each frame of the image and constructs a mapping cache structure with a consistent structure.

[0043] S52. The ground enhancement unit performs a pixel-level reversible expansion operation on the artifact candidate area based on the reversible restoration log and the pulse fire mapping matrix. If a pixel location exists in the reversible restoration log, the corresponding original pixel value is used as the current pixel value. If the pixel location does not exist in the reversible restoration log, the restored pixel value is retained as the current pixel value, ultimately forming a reversible expanded image data frame.

[0044] S53. The ground enhancement unit performs a fine-grained texture reconstruction operation based on the reversibly expanded image data frame and the weight mask corresponding to the image block to generate a new reconstructed output value;

[0045] S54. The reconstructed output values ​​of all pixel positions are combined into a complete image frame, which constitutes an artifact-removed high-quality image product output by the ground enhancement unit.

[0046] Optionally, the S6 includes the following steps:

[0047] S61. The ground enhancement unit analyzes the processing status of the artifact-removed high-quality image product and extracts the artifact residual index in the image frame. , texture consistency index Grayscale offset indicator ;

[0048] S62. Normalize the artifact residue index, texture consistency index and grayscale offset index and fuse them into a comprehensive artifact removal performance evaluation function ;

[0049] S63. Based on the changing trend of the comprehensive artifact rejection performance evaluation function, a satellite-ground parameter update function is constructed to dynamically adjust the global coupling gain coefficient of the onboard hierarchical progressive attention adaptive pulse-coupled neural network model. Link strength threshold :

[0050] ;

[0051] in, 、 are coupling gain and threshold update rate respectively, is the target value for the desired culling performance;

[0052] S64. Update the parameters to The data is packaged into parameter packets and sent back to the onboard processing unit in real time through the satellite-to-ground link, replacing the original model parameters for the next round of image data processing, forming a closed-loop adaptive iterative update mechanism.

[0053] Optionally, the artifact residual index is:

[0054] ;

[0055] in, 、 is the height and width of the image frame, is the pixel position The probability of artifacts, represents the mean probability of residual artifacts in the entire image;

[0056] The texture consistency index:

[0057] ;

[0058] in, 、 Respectively The mean texture gradient of the region in the preprocessed and enhanced image frame, is the number of partitions;

[0059] The grayscale deviation index:

[0060] ;

[0061] in, Pixel locations in the image for artifact culling The reconstructed pixel value, is the pixel value in the original image data.

[0062] The beneficial effects of the present invention are:

[0063] (1) The present invention constructs a hierarchical progressive attention adaptive pulse coupled neural network model with satellite deployment capability. Without relying on deep convolutional networks, it achieves millisecond-level positioning of sparse irregular artifacts in low-orbit images. A hierarchical PCNN structure is used to couple the global coarse-scale, medium-scale and local fine-scale pulse propagation layers, and dynamic gating is performed through the attention mechanism, so that each layer can be synchronized and regulated according to the local excitation characteristics of the image. By dynamically adjusting the coupling gain coefficient, local threshold and timing attenuation factor, the model's sensitivity to unstructured artifacts such as particle spots and pseudo-row and column dislocation is significantly improved, ensuring that the artifacts can be accurately captured and modeled within 50ms at a 1024×1024 image resolution.

[0064] (2) The present invention proposes a block-level reversible restoration strategy based on pulse fire mapping, which achieves high coupling, high fidelity and reversible mapping between artifact positioning and pixel restoration. By constructing a pixel-level index system of pulse fire mapping matrix and artifact probability map, and performing restoration credibility control based on image block-level weight mask, the local adaptive fusion function is used to dynamically reconstruct the artifact pixel value, and a reversible restoration log is generated to ensure that the restoration status of each pixel position can be fully traced back. While achieving high-fidelity image restoration, it ensures that the ground enhancement algorithm can be expanded and reconstructed as needed. Compared with traditional compression coding restoration methods, it has stronger interpretability and inter-system compatibility.

[0065] (3) The present invention establishes a two-way collaborative update mechanism between satellite and ground, dynamically adjusts the core parameters of the pulse coupled neural network through processing performance feedback, realizes closed-loop optimization of artifact rejection capability and image quality control capability, constructs a performance evaluation function that integrates three factors: artifact residue index, texture consistency index and grayscale offset index in the ground enhancement unit, and uses the evaluation result to update the global coupling gain coefficient and link strength threshold in the satellite-borne model in real time, establishes a satellite-ground feedback closed loop, and can adapt to the changes in image quality of satellites in different orbits and different imaging conditions, achieves a dynamic balance between artifact suppression and original texture preservation, and improves the consistency of image processing under attitude disturbances. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0067] Figure 1 This is a flow chart of a method for fast removing artifacts from low-orbit images based on pulse-coupled neural networks proposed by the present invention. DETAILED DESCRIPTION

[0068] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0069] refer to Figure 1 A method for fast removing artifacts from low-orbit images based on pulse coupled neural networks includes the following steps:

[0070] S1. Acquire original low-orbit image data, which is generated by a low-orbit remote sensing satellite after attitude synchronization calibration and preliminary radiometric calibration are completed;

[0071] S2. Input the original LEO image data into the onboard hierarchical progressive attention adaptive pulse-coupled neural network model, initialize the coupling gain coefficient, link strength threshold, and time decay parameters, and perform pulse traversal based on the synchronization signal of the original LEO image to generate the neuron excitation time series;

[0072] S3. Construct a spike-fire mapping matrix based on the neuron firing time series. The spike-fire mapping matrix records the spike triggering sequence and neighborhood connectivity at each pixel. Neighborhood spike synchrony is calculated using the spike-fire mapping matrix to form artifact candidate regions. An artifact probability map is generated based on neighborhood spike synchrony and information entropy increment.

[0073] S4. Construct a block-level weight mask based on the artifact probability map. Establish a one-to-one correspondence index between the block-level weight mask and the pulse fire mapping matrix to achieve reversibility of the restoration. Perform pixel restoration on the artifact candidate area using the block-level weight mask to generate pre-processed low-orbit image data and simultaneously generate a reversible restoration log. The reversible restoration log records the pixel values ​​and corresponding indexes before and after the restoration.

[0074] S5. The pre-processed LEO image data, the pulse fire mapping matrix, the block-level weight mask, and the reversible restoration log are transmitted to the ground enhancement unit. The ground enhancement unit reversibly expands the artifact candidate regions based on the pulse fire mapping matrix and the reversible restoration log, and performs fine-grained texture reconstruction in combination with the block-level weight mask to generate artifact-removed high-quality image products.

[0075] S6. Update the coupling gain coefficient and link strength threshold of the onboard hierarchical progressive attention adaptive pulse coupled neural network model based on the processing status of the artifact rejection high-quality image product, and return the updated parameters to the onboard hierarchical progressive attention adaptive pulse coupled neural network model to form a closed-loop adaptive artifact rejection process between the satellite and the ground.

[0076] In this embodiment, S1 includes the following steps:

[0077] S11. Construct a raw low-orbit image data sequence, where each frame represents a visible light image or multispectral image acquired by a low-orbit remote sensing satellite at a specific moment. The raw low-orbit image data sequence consists of multiple image frames acquired at different time points, and each frame is labeled with the corresponding acquisition time information;

[0078] S12. Obtain the attitude quaternion and orbital position of the low-orbit remote sensing satellite corresponding to the moment of image acquisition for each frame. The attitude quaternion represents the attitude information of the low-orbit remote sensing satellite in three-dimensional space, and the orbital position represents the position coordinate of the low-orbit remote sensing satellite in inertial space. The attitude quaternion and orbital position together constitute the data basis required for attitude synchronization calibration. The attitude quaternion and orbital position are used to perform attitude synchronization calibration on the original low-orbit image data sequence. The attitude synchronization calibration is used to align the image frame with the satellite attitude information, so that each pixel in the image establishes a spatially consistent relationship with the physical ground position.

[0079] S13. Perform preliminary radiometric calibration on the image frames after attitude synchronization calibration. This preliminary radiometric calibration is used to correct the brightness response deviation of the original pixel grayscale value. The brightness response value is obtained by multiplying the brightness response value by the radiation gain coefficient and adding the radiation offset term. The radiation gain coefficient is used to represent the amplification ratio of the imaging system to the input signal. The radiation offset term is used to correct the zero-point drift of the sensor. The brightness response value serves as the basis for the pixel intensity of the image in the physical radiation sense.

[0080] S14. After completing the preliminary radiometric calibration, perform strip equalization preprocessing on the image frame. Strip equalization preprocessing is performed along the image column direction. All pixel values ​​in each column of the image frame are averaged to obtain a column average value. The column average value represents the overall brightness level of the column in the vertical direction. The column average value is used to determine the offset trend of directional brightness unevenness. The column average value is used for pixel intensity normalization processing to eliminate large-scale directional brightness unevenness features and obtain the final original low-orbit image data.

[0081] In this embodiment, S2 includes the following steps:

[0082] S21. Input the final raw low-orbit image data into the onboard hierarchical progressive attention adaptive pulse coupled neural network model. The onboard hierarchical progressive attention adaptive pulse coupled neural network model includes a global coarse-scale PCNN layer, a mid-scale PCNN layer, and a local fine-scale attention PCNN layer. The three layers of PCNN are cascaded and based on the root mean square value of the low-orbit remote sensing satellite attitude disturbance and the imaging system instantaneous gain Initialize the global coarse-scale layer parameters:

[0083] ;

[0084] in, 、 、 is the onboard reference parameter, 、 、 is the proportional adjustment factor, is the amplitude of attitude disturbance change in the rolling observation window, is the global initial coupling gain coefficient, is the global initial adaptive time decay factor, is the global initial link strength threshold, is the reference standard value of attitude disturbance of low-orbit remote sensing satellite, is the nominal gain value of the imaging system;

[0085] S22. In the global coarse-scale PCNN layer, the adaptive time decay factor Drive the neuron internal state update and calculate the global grayscale statistical variance of the final original low-orbit image data Adaptively determine the dynamic triggering conditions of neurons and generate global coarse-scale neuron pulse outputs to preliminarily locate large-scale bands and obvious artifacts of high-energy particle breakdown in the original low-orbit image data:

[0086] ;

[0087] in, is the pixel position At the moment The internal state voltage, is the final original low-orbit image data pixel grayscale value, for Receptive field neighborhood, is the long-range coupling weight, is the pulse output at the previous moment, is the global variance amplification factor;

[0088] S23. Initialize the mesoscale coupling gain coefficient in the mesoscale PCNN layer and the mesoscale initial link strength threshold , and calculate the local entropy change based on the pulse output of the global coarse-scale PCNN layer Obtaining the mesoscale dynamic coupling gain coefficient , according to the mesoscale dynamic coupling gain coefficient Dynamically adjust the trigger sensitivity of mesoscale neurons to form mesoscale pulse outputs to capture the edge and contour features of low-orbit image artifacts:

[0089] ;

[0090] The dynamic coupling gain coefficient is calculated as:

[0091] ;

[0092] in, is the entropy sensitive factor, is the upper limit of entropy, is the pixel position Neighborhood grayscale The instantaneous probability of is a grayscale set, Correction weights for coarse layer pulses;

[0093] S24. Initialize the local scale initial link strength threshold in the local fine-scale attention PCNN layer , calculate the local attention coefficient based on the pulse output of the mid-scale PCNN layer , and use the local attention coefficient Dynamic triggering thresholds for local-scale neurons Perform real-time adaptive updates to form local fine-scale pulse outputs, enabling the positioning of complex fine-scale artifacts such as stripes, scintillation points, compression errors, and particle interference in low-orbit images:

[0094] ;

[0095] The dynamic trigger threshold is calculated as:

[0096] ;

[0097] in, for neighborhood, is the number of neighborhood pixels, is the mesoscale pulse output, is the local reference threshold, is the attention sensitivity factor, is the gradient gain coefficient, Pixels Grayscale gradient amplitude, is the coupling feedback weight;

[0098] S25. An inter-layer attention interaction mechanism is established between the global coarse-scale PCNN layer, the medium-scale PCNN layer, and the local fine-scale attention PCNN layer. The pulse output of the global coarse-scale PCNN layer guides the parameter update of the medium-scale PCNN layer, and the pulse output of the medium-scale PCNN layer guides the parameter update of the local fine-scale attention PCNN layer. The response results of the local fine-scale attention PCNN layer act reversely on the medium-scale and global coarse-scale PCNN layers, dynamically optimizing the trigger sensitivity and coupling strength of each layer, realizing a coarse-scale-medium-scale-fine-scale and fine-scale-medium-scale-coarse-scale bidirectional interactive closed loop, and finally generating a hierarchical progressive neuron excitation time series matrix.

[0099] In this embodiment, S3 includes the following steps:

[0100] S31. Constructing a pulse-fire mapping matrix based on the hierarchical progressive neuron excitation time series matrix The pulse-fire mapping matrix records the first pulse triggering moment of the neuron corresponding to each pixel position in the original low-orbit image data. The first pulse triggering moment is used to measure the response priority of the neuron at that position and its positioning order in the temporal structure. The pulse-fire mapping matrix is ​​obtained by mapping the excitation time of each pixel point from the hierarchical progressive neuron excitation time series matrix to a two-dimensional matrix consistent with the image coordinates;

[0101] S32. Based on the pulse fire mapping matrix, the neighborhood pulse synchronization of each pixel position within the receptive field is calculated. The neighborhood pulse synchronization is used to characterize the consistency of the pulse triggering timing of the pixel position and its neighboring pixels. The neighborhood pulse synchronization is obtained by counting the relative amplitude of the difference between the first pulse triggering time of the pixel and all pixels in the neighborhood. The neighborhood pulse synchronization of all positions constitutes a synchronous pulse connectivity graph. , the synchronous spike connectivity map is used to characterize the spatial consistency distribution of the neuronal firing temporal structure of the original LEO image data:

[0102] ;

[0103] in, Pixel position Central neighborhood, is the total number of pixels in the neighborhood, Neighborhood pixels At the first pulse triggering moment in the hierarchical progressive attention adaptive pulse coupled neural network model, is the neighborhood pixel position At the first pulse triggering moment in the hierarchical progressive attention adaptive pulse coupled neural network model, is the synchronization penalty function, which reflects the degree of inhibition of connectivity caused by the difference in excitation time. Indicates pixel position The strength of the consistency with the neighboring pulse sequence;

[0104] S33. According to the synchronous pulse connection diagram The neighborhood pulse synchronization of each pixel position in the , and filter out all the neighborhood pulse synchronizations below the synchronization threshold The pixel position of The preliminary candidate region of the artifact is used to identify the potential abnormal region in which the neuronal firing timing structure is obviously separated from the surrounding structure in the original low-orbit image data. The synchronization threshold A fixed empirical threshold for demarcating normal and abnormal synchrony;

[0105] S34. Calculate the local information entropy increment for each pixel position in the preliminary candidate area of ​​the artifact The local information entropy increment is used to measure the degree of deviation of the grayscale complexity of the local area where the pixel is located relative to the average grayscale complexity of the synchronously connected area to which it belongs. The local information entropy increment is obtained by calculating the difference between the local Shannon entropy in the neighborhood of the pixel position and the average local Shannon entropy of its corresponding synchronously connected area. The larger the local information entropy increment, the more inconsistent the grayscale structure of the pixel position is, and the more likely it is that there is image artifact interference;

[0106] S35. Generate an artifact probability map by combining the neighborhood pulse synchronization and local information entropy increment of each pixel position The artifact probability map is used to indicate the probability of each pixel being an artifact. The artifact probability is obtained by linearly weighting the neighborhood pulse synchronization and the local information entropy increment. The neighborhood pulse synchronization is used to describe the abnormality of the temporal structure, and the local information entropy increment is used to describe the abnormality of the grayscale structure:

[0107] ;

[0108] in, 、 is the linear weighting factor, is the normalized upper bound of the local entropy set.

[0109] In this embodiment, S4 includes the following steps:

[0110] S41. Divide the original LEO image data into multiple fixed-size image blocks according to the artifact probability map and construct a block-level weight mask ,in Represents the image block coordinates, each image block contains a set of pixels The corresponding block-level weight mask is defined as:

[0111] ;

[0112] in, Represents the total number of pixels in the image block, is the pixel position Artifact probability value, block-level weight mask Indicates the overall artifact intensity of the image block;

[0113] S42. Weight mask at block level On this basis, the artifact restoration strength threshold is set, and all image blocks with weight values ​​greater than the artifact restoration strength threshold are defined as the restoration target block area set. ;

[0114] S43. For each set of repair target block areas Image blocks are constructed, and pixel-level one-to-one corresponding indexes are established based on the pulse fire mapping matrix. Pixel-level reversible repair operations are performed to obtain the repaired pixel values. , the restored pixel value is obtained by weighted fusion of the original pixel and the neighborhood reference value according to the artifact probability:

[0115] ;

[0116] in, represents the current pixel artifact weight, is the original pixel value, Indicates pixel position The mean of the neighborhood pixels;

[0117] S44. While performing pixel restoration, a reversible restoration log is generated, recording the position index, original pixel value, restored pixel value, and image block index of all restored pixels;

[0118] S45. Set all restored pixel values Replace the pixel values ​​at the corresponding positions in the original LEO image data to generate pre-processed LEO image data , whose structure remains consistent with the input image frame.

[0119] In this embodiment, S5 includes the following steps:

[0120] S51. The pre-processed low-orbit image data, the pulse fire mapping matrix, the block-level weight mask, and the reversible repair log are transmitted to the ground enhancement unit. The ground enhancement unit receives the corresponding processing information packet for each frame of the image and constructs a mapping cache structure with a consistent structure.

[0121] S52. The ground enhancement unit performs a pixel-level reversible expansion operation on the artifact candidate area based on the reversible restoration log and the pulse fire mapping matrix. The pixel-level reversible expansion operation is used to restore the value of each restored pixel. The restoration is based on the original pixel value before restoration recorded in the reversible restoration log. If a pixel position exists in the reversible restoration log, the corresponding original pixel value is used as the current pixel value. If the pixel position does not exist in the reversible restoration log, the restored pixel value is retained as the current pixel value, ultimately forming a reversible expanded image data frame.

[0122] S53. The ground enhancement unit performs a fine-grained texture reconstruction operation based on the reversibly expanded image data frame and the weight mask of the corresponding image block. The fine-grained texture reconstruction operation is used to fuse the texture credibility weight of the image block where the pixel is located with the weighted average of the cross-frame observed pixels to generate a new reconstructed output value. The texture credibility weight is obtained by subtracting the block-level weight mask value of the image block to which the pixel belongs from 1, and represents the degree to which the pixel area is a credible texture. The weighted average of the cross-frame observed pixels is used to introduce multi-frame redundant information to enhance the image texture quality. The generated reconstructed output value is a weighted fusion result of the current pixel value and the cross-frame average.

[0123] S54. The reconstructed output values ​​of all pixel positions are combined into a complete image frame, which constitutes an artifact-removed high-quality image product output by the ground enhancement unit.

[0124] In this embodiment, S6 includes the following steps:

[0125] S61. The ground enhancement unit analyzes the processing status of the artifact-removed high-quality image product and extracts the artifact residual index in the image frame. , texture consistency index Grayscale offset indicator ;

[0126] S62. Normalize the artifact residue index, texture consistency index and grayscale offset index and fuse them into a comprehensive artifact removal performance evaluation function :

[0127] ;

[0128] in, 、 、 are the normalized values ​​of the indicators, , , is the weighting coefficient;

[0129] S63. Based on the changing trend of the comprehensive artifact rejection performance evaluation function, a satellite-ground parameter update function is constructed to dynamically adjust the global coupling gain coefficient of the onboard hierarchical progressive attention adaptive pulse-coupled neural network model. Link strength threshold :

[0130] ;

[0131] in, 、 are coupling gain and threshold update rate respectively, is the target value for the desired culling performance;

[0132] S64. Update the parameters to The data is encapsulated into parameter packets and sent back to the onboard processing unit in real time through the satellite-to-ground link, replacing the original model parameters for the next round of image data processing, forming a closed-loop adaptive iterative update mechanism to achieve dynamic optimization of the artifact removal performance of the satellite-to-ground integration.

[0133] In this embodiment, the artifact residual index is:

[0134] ;

[0135] in, 、 is the height and width of the image frame, is the pixel position The probability of artifacts, represents the mean probability of residual artifacts in the entire image;

[0136] Texture consistency index:

[0137] ;

[0138] in, 、 Respectively The mean texture gradient of the region in the preprocessed and enhanced image frame, is the number of partitions;

[0139] Grayscale deviation index:

[0140] ;

[0141] in, Position in the image for artifact removal The reconstructed pixel value, is the pixel value in the original image data. Example

[0142] In the continuous downlink image data stream, when a low-orbit remote sensing satellite processed five frames of image data numbered 010438-010442, the system detected significant striping artifacts in the column direction of these images. The artifact area was mainly concentrated in the lower left quadrant of the image, with a strip width of approximately 6-12 pixels, and an average brightness fluctuation amplitude of more than 1.8 times that of the surrounding background. The initial detection of the artifact was carried out through the on-board synchronization signal trigger mechanism. After the pulse traversal of each frame of the image, multiple continuous synchronous "pulse-dense" band blocks appeared in the neuron excitation time series, and the triggering time was highly close. The system automatically marked these areas as high-risk trigger areas.

[0143] Then, the system constructs a pulse fire mapping matrix based on the excitation time series and Neighborhood pulse synchronization is calculated. For frame 010440, the system identified a low-synchronization region with connectivity less than 0.42 within the pixel coordinate range (314:320, 240:410), covering an area exceeding 2.9% of the total image pixels. Furthermore, the mean local entropy increment in this region reached 0.161, 3.4 times that of other image regions. Based on this information, the system constructed an artifact probability map, with the maximum probability point occurring at position (317, 252) with an artifact probability of 0.91.

[0144] According to the artifact probability map, the image is divided into A total of 256 image blocks were obtained, of which 19 had block-level weight mask values ​​exceeding 0.75 and were identified by the system as heavily contaminated artifact areas. These image blocks were then submitted to the pixel restoration engine for processing. Each restored pixel generated a log record, including the pre-restoration value, the post-restoration value, and the corresponding image frame, image block, and coordinate information. For example, in the image block numbered B-118, the grayscale value of the pixel position (321,260) was restored from the original 136 to 122, with a weight coefficient of 0.88, which was recorded in the reversible restoration log entry R-0047591.

[0145] All restoration logs and fire mapping matrices are transmitted along with the preprocessed image to the ground processing module. After receiving image 010440, the ground module, by reversing the restoration logs and weight masks, discovered that image blocks B-118, B-120, and B-135 had high restoration coefficients and, after texture reconstruction, still exhibited artifact edges with inconsistent frequencies. The system initiated a fine-grained texture enhancement process, first reversibly expanding these image blocks to restore them to their original pixels. Multiple frames of images from the corresponding regions over the past three trajectory segments were then retrieved and fused using a time-windowed sliding weighted averaging method to obtain a temporally smoothed value for each pixel. For example, the weighted average value for pixel position (321, 260) was 127.4, resulting in a final fused output value of 125.

[0146] After the ground enhancement output is completed, the system performs artifact residual and texture consistency evaluation on the current frame image. The artifact residual index is calculated to be 0.036, a decrease of 58% compared with the previous version; the texture consistency index is improved from 0.86 to 0.93; the average repair grayscale offset is reduced from 11.2 to 4.3; and the image relative mean error is less than 2.4%. At the same time, the comprehensive artifact removal performance function of this image frame is The updated value is 0.128, which is lower than the system-set threshold of 0.15.

[0147] According to the performance score results, the ground system automatically adjusts the coupling gain coefficient and synchronization threshold in the onboard model. The new parameter pairs are and The system encodes the parameter update package and sends it back to the onboard processing end with the number UPD-01883. It takes effect when the image frame number 010443 is processed.

[0148] The image frame after this parameter update is shown in the downlink data. The high-value area in the overall artifact probability map is greatly reduced, and the weight mask value of image block B-118 is reduced from the original 0.81 to 0.44. At the same time, image block B-120 is no longer marked as a candidate area after processing.

[0149] After the entire process was completed, the system's output evaluation report showed that in this case, a total of 107,392 pixels of suspected striping artifacts were eliminated, accounting for approximately 10.2% of the total image area. The final average probability of artifact residue dropped to 0.039, a decrease of 65.8% compared to the control group image (0.114) without the introduction of this method; the texture structure similarity index SSIM increased by 12.7%, and the manual review accuracy increased by 15.3%.

[0150] This Example 1 verifies the ability of the method of the present invention to identify stripes, bright spots, and dynamic compression artifacts and the efficiency of automatic repair under high interference conditions. At the same time, through satellite-ground collaborative feedback, adaptive control of the core parameters of the model is achieved. The system closed-loop operation is stable, the image quality is significantly improved, the processing delay meets the requirements of fast remote sensing services, and it has engineering feasibility.

[0151] The present invention constructs a hierarchical progressive attention adaptive pulse coupled neural network model with satellite deployment capability. Without relying on deep convolutional networks, it achieves millisecond-level positioning of sparse irregular artifacts in low-orbit images. A hierarchical PCNN structure is used to couple the global coarse-scale, medium-scale and local fine-scale pulse propagation layers, and dynamic gating is performed through the attention mechanism, so that each layer can be synchronously regulated according to the local excitation characteristics of the image. By dynamically adjusting the coupling gain coefficient, local threshold and timing attenuation factor, the model's sensitivity to unstructured artifacts such as particle spots and pseudo-row and column dislocations is significantly improved, ensuring that artifacts can be accurately captured and modeled within 50ms at a 1024×1024 image resolution.

[0152] The present invention proposes a block-level reversible repair strategy based on pulse fire mapping, which realizes high coupling, high fidelity and reversible mapping between artifact positioning and pixel repair. By constructing a pixel-level index system of pulse fire mapping matrix and artifact probability map, and performing repair credibility control according to image block-level weight mask, the artifact pixel value is dynamically reconstructed by local adaptive fusion function, and a reversible repair log is generated to ensure that the repair status of each pixel position can be fully traced back. While achieving high-fidelity image restoration, it ensures that the ground enhancement algorithm can be expanded and reconstructed the original texture structure as needed. Compared with traditional compression coding repair methods, it has stronger interpretability and inter-system compatibility.

[0153] The present invention establishes a two-way collaborative update mechanism between satellite and ground, dynamically adjusts the core parameters of the pulse-coupled neural network through processing performance feedback, realizes closed-loop optimization of artifact rejection capability and image quality control capability, constructs a performance evaluation function that integrates three factors: artifact residue index, texture consistency index and grayscale offset index in the ground enhancement unit, and uses the evaluation result to update the global coupling gain coefficient and link strength threshold in the satellite-borne model in real time, establishes a satellite-ground feedback closed loop, can adapt to the changes in image quality of satellites in different orbits and different imaging conditions, achieves a dynamic balance between artifact suppression and original texture preservation, and improves the consistency of image processing under attitude disturbances.

[0154] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for removing artifacts from low-orbit images based on pulse coupled neural networks, characterized in that: The steps include: S1. Obtaining original low-orbit image data; S2. Input the original LEO image data into the onboard hierarchical progressive attention adaptive pulse-coupled neural network model, and perform pulse traversal based on the synchronization signal of the original LEO image to generate a neuron firing time series; S3. Construct a spike-fire mapping matrix based on the neuron firing time series, use the spike-fire mapping matrix to calculate the neighborhood spike synchrony, form an artifact candidate region, and generate an artifact probability map based on the neighborhood spike synchrony and information entropy increment; The S3 includes the following steps: S31. Constructing a pulse-fire mapping matrix based on the hierarchical progressive neuron excitation time series matrix ,The pulse fire mapping matrix is ​​obtained by mapping the firing time of each pixel from the hierarchical ,neuron firing time series matrix to a two-dimensional matrix consistent with the ,image coordinates; S32. Based on the pulse fire mapping matrix, the neighborhood pulse synchronization of each pixel position within the receptive field is calculated. The neighborhood pulse synchronization is obtained by counting the relative amplitude of the difference between the first pulse triggering time of the pixel and all pixels in the neighborhood. The neighborhood pulse synchronization of all positions constitutes a synchronous pulse connectivity graph. , the synchronous spike connectivity map is used to characterize the spatial consistency distribution of the neuronal firing temporal structure of the original LEO image data: ； in, Pixel position Central neighborhood, is the total number of pixels in the neighborhood, Neighborhood pixels At the first pulse triggering moment in the hierarchical progressive attention adaptive pulse coupled neural network model, is the neighborhood pixel position At the first pulse triggering moment in the hierarchical progressive attention adaptive pulse coupled neural network model, is the synchronization penalty function, which reflects the degree of inhibition of connectivity caused by the difference in excitation time. Indicates pixel position The strength of the consistency with the neighboring pulse sequence; S33. According to the synchronous pulse connection diagram The neighborhood pulse synchronization of each pixel position in the , and filter out all the neighborhood pulse synchronizations below the synchronization threshold The pixel position of The preliminary candidate region of the artifact is used to identify the potential abnormal region in which the neuronal firing timing structure is obviously separated from the surrounding structure in the original low-orbit image data. The synchronization threshold A fixed empirical threshold for demarcating normal and abnormal synchrony; S34. Calculate the local information entropy increment for each pixel position in the preliminary candidate area of ​​the artifact ,The local information entropy increment is obtained by calculating the difference between the local Shannon entropy in the neighborhood of the pixel position and the average local Shannon entropy of its corresponding synchronously connected regions; S35. Generate an artifact probability map by combining the neighborhood pulse synchronization and local information entropy increment of each pixel position , the artifact probability is obtained by linearly weighting the neighborhood pulse synchronization and the local information entropy increment, where the neighborhood pulse synchronization is used to describe the abnormality of the temporal structure, and the local information entropy increment is used to describe the abnormality of the grayscale structure; S4. Construct a block-level weight mask based on the artifact probability map, use the block-level weight mask to perform pixel restoration on the artifact candidate area, generate pre-processed low-orbit image data, and simultaneously generate a reversible restoration log; S5. The pre-processed LEO image data, the pulse fire mapping matrix, the block-level weight mask, and the reversible restoration log are transmitted to the ground enhancement unit. The ground enhancement unit reversibly expands the artifact candidate regions based on the pulse fire mapping matrix and the reversible restoration log, and performs fine-grained texture reconstruction in combination with the block-level weight mask to generate artifact-removed high-quality image products. S6. Update the coupling gain coefficient and link strength threshold of the onboard hierarchical progressive attention adaptive pulse coupled neural network model according to the processing status of the artifact-removed high-quality image product, and return the updated parameters to the onboard hierarchical progressive attention adaptive pulse coupled neural network model.

2. The method for removing artifacts from low-orbit images based on pulse coupled neural networks according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Construct a raw low-orbit image data sequence, where each frame represents a visible light image or multispectral image acquired by a low-orbit remote sensing satellite at a specific moment. The raw low-orbit image data sequence consists of multiple image frames acquired at different time points, and each frame is labeled with the corresponding acquisition time information; S12. Obtain the attitude quaternion and orbital position of the low-orbit remote sensing satellite corresponding to the moment of acquisition of each frame of image. The attitude quaternion and orbital position together constitute the data basis required for attitude synchronization calibration. The attitude quaternion and orbital position are used to perform attitude synchronization calibration on the original low-orbit image data sequence. The attitude synchronization calibration is used to align the image frame with the satellite attitude information; S13. Perform preliminary radiometric calibration on the image frame after attitude synchronization calibration, which is used to correct the brightness response deviation of the original pixel grayscale value; S14. After completing the preliminary radiometric calibration, perform strip equalization preprocessing on the image frame to remove large-scale directional brightness unevenness features to obtain the final original low-orbit image data.

3. The method for removing artifacts from low-orbit images based on pulse coupled neural networks according to claim 1, characterized in that: The S2 comprises the following steps: S21. Input the final raw low-orbit image data into the onboard hierarchical progressive attention adaptive pulse coupled neural network model. The onboard hierarchical progressive attention adaptive pulse coupled neural network model includes a global coarse-scale PCNN layer, a mid-scale PCNN layer, and a local fine-scale attention PCNN layer. The three layers of PCNN are cascaded and based on the root mean square value of the low-orbit remote sensing satellite attitude disturbance and the imaging system instantaneous gain Initialize the global coarse-scale layer parameters: ； in, 、 、 is the onboard reference parameter, 、 、 is the proportional adjustment factor, is the amplitude of attitude disturbance change in the rolling observation window, is the global initial coupling gain coefficient, is the global initial adaptive time decay factor, is the global initial link strength threshold, is the reference standard value of attitude disturbance of low-orbit remote sensing satellite, is the nominal gain value of the imaging system; S22. In the global coarse-scale PCNN layer, the adaptive time decay factor Drive the neuron internal state update and calculate the global grayscale statistical variance of the final original low-orbit image data Adaptively determine the dynamic triggering conditions of neurons and generate global coarse-scale neuron pulse outputs ; S23. Initialize the mesoscale coupling gain coefficient in the mesoscale PCNN layer and the mesoscale initial link strength threshold , and calculate the local entropy change based on the pulse output of the global coarse-scale PCNN layer Obtaining the mesoscale dynamic coupling gain coefficient , according to the mesoscale dynamic coupling gain coefficient Dynamically adjust the trigger sensitivity of mesoscale neurons to form mesoscale pulse output ; S24. Initialize the local scale initial link strength threshold in the local fine-scale attention PCNN layer , calculate the local attention coefficient based on the pulse output of the mid-scale PCNN layer , and use the local attention coefficient Dynamic triggering thresholds for local-scale neurons Perform real-time adaptive updates to form local fine-scale pulse outputs, enabling the positioning of complex fine-scale artifacts such as stripes, scintillation points, compression errors, and particle interference in low-orbit images; S25. An inter-layer attention interaction mechanism is established between the global coarse-scale PCNN layer, the medium-scale PCNN layer, and the local fine-scale attention PCNN layer. The pulse output of the global coarse-scale PCNN layer guides the parameter update of the medium-scale PCNN layer, and the pulse output of the medium-scale PCNN layer guides the parameter update of the local fine-scale attention PCNN layer. The response results of the local fine-scale attention PCNN layer act reversely on the medium-scale and global coarse-scale PCNN layers, dynamically optimizing the trigger sensitivity and coupling strength of each layer, realizing a coarse-scale-medium-scale-fine-scale and fine-scale-medium-scale-coarse-scale bidirectional interactive closed loop, and finally generating a hierarchical progressive neuron excitation time series matrix.

4. The method for removing artifacts from low-orbit images based on pulse coupled neural networks according to claim 1, characterized in that: The S4 comprises the following steps: S41. Divide the original LEO image data into multiple fixed-size image blocks according to the artifact probability map and construct a block-level weight mask ,in Represents the image block coordinates, each image block contains a set of pixels Corresponding block-level weight mask ; S42. Weight mask at block level On this basis, the artifact restoration strength threshold is set, and all image blocks with weight values ​​greater than the artifact restoration strength threshold are defined as the restoration target block area set. ; S43. For each set of repair target block areas Image blocks are constructed, and pixel-level one-to-one corresponding indexes are established based on the pulse fire mapping matrix. Pixel-level reversible repair operations are performed to obtain the repaired pixel values. ,The restored pixel value is obtained by weighted fusion of the original pixel and the neighborhood reference value according to the artifact probability; S44. While performing pixel restoration, a reversible restoration log is generated, recording the position index, original pixel value, restored pixel value, and image block index of all restored pixels; S45. Set all restored pixel values Replace the pixel values ​​at the corresponding positions in the original LEO image data to generate pre-processed LEO image data , whose structure remains consistent with the input image frame.

5. The method for removing artifacts from low-orbit images based on pulse coupled neural networks according to claim 4, characterized in that: The S5 comprises the following steps: S51. The pre-processed low-orbit image data, the pulse fire mapping matrix, the block-level weight mask, and the reversible repair log are transmitted to the ground enhancement unit. The ground enhancement unit receives the corresponding processing information packet for each frame of the image and constructs a mapping cache structure with a consistent structure. S52. The ground enhancement unit performs a pixel-level reversible expansion operation on the artifact candidate area based on the reversible restoration log and the pulse fire mapping matrix. If a pixel location exists in the reversible restoration log, the corresponding original pixel value is used as the current pixel value. If the pixel location does not exist in the reversible restoration log, the restored pixel value is retained as the current pixel value, ultimately forming a reversible expanded image data frame. S53. The ground enhancement unit performs a fine-grained texture reconstruction operation based on the reversibly expanded image data frame and the weight mask corresponding to the image block to generate a new reconstructed output value; S54. The reconstructed output values ​​of all pixel positions are combined into a complete image frame, which constitutes an artifact-removed high-quality image product output by the ground enhancement unit.

6. The method for removing artifacts from low-orbit images based on pulse coupled neural networks according to claim 5, characterized in that: The S6 comprises the following steps: S61. The ground enhancement unit analyzes the processing status of the artifact-removed high-quality image product and extracts the artifact residual index in the image frame. , texture consistency index Grayscale offset indicator ; S62. Normalize the artifact residue index, texture consistency index and grayscale offset index and fuse them into a comprehensive artifact removal performance evaluation function ; S63. Based on the changing trend of the comprehensive artifact rejection performance evaluation function, a satellite-ground parameter update function is constructed to dynamically adjust the global coupling gain coefficient of the onboard hierarchical progressive attention adaptive pulse-coupled neural network model. Link strength threshold : ； in, 、 are coupling gain and threshold update rate respectively, is the target value for the desired culling performance; S64. Update the parameters to The data is packaged into parameter packets and sent back to the onboard processing unit in real time through the satellite-to-ground link, replacing the original model parameters for the next round of image data processing, forming a closed-loop adaptive iterative update mechanism.

7. The method for removing artifacts from low-orbit images based on pulse coupled neural networks according to claim 6, characterized in that: The artifact residual index is: ； in, 、 is the height and width of the image frame, is the pixel position The probability of artifacts, represents the mean probability of residual artifacts in the entire image; The texture consistency index: ； in, 、 Respectively The mean texture gradient of the region in the preprocessed and enhanced image frame, is the number of partitions; The grayscale deviation index: ； in, Pixel locations in the image for artifact culling The reconstructed pixel value, is the pixel value in the original image data.

Images