A remote sensing image lightweight pre-checking and hierarchical scheduling method for on-board AI

Abstract

This invention relates to on-board remote sensing image processing and discloses a lightweight pre-detection and hierarchical scheduling method for on-board AI remote sensing images. The method includes acquiring working images using original high-resolution images, extracting image quality evaluation parameters and background interference evaluation parameters, and calculating quality scores and gating coefficients; calculating target indicators by combining morphological and intensity feature parameters; reading regional configuration files, comparing historical images, and parsing scheduling instructions to generate matching scores, change scores, and timeliness scores; generating an initial score by weighting multiple indicators, generating a base score by calculating the gating coefficients, triggering interference suppression and hard truncation mechanisms, and outputting a comprehensive score after pressure reduction; classifying value levels based on the comprehensive score and performing hierarchical physical resource scheduling. This invention utilizes basic visual operators to extract features, reducing the computational overhead of high-performance on-board models, and improves task processing efficiency under limited hardware resources by dynamically matching data value levels with node load status.

Description

Technical Field

[0001] This invention relates to on-board remote sensing image processing, specifically a lightweight pre-detection and hierarchical scheduling method for remote sensing images for on-board AI. Background Technology

[0002] With the development of satellite remote sensing technology, the demand for on-board image processing is increasing, but the power consumption, computing power, and storage space of satellite platforms are limited. In the feature extraction and target analysis stages, existing technologies usually directly transplant complex neural network models, which will generate high power consumption in the on-board environment. At the same time, a large number of images obscured by clouds or occupied by large areas of invalid background are directly entered into the subsequent processing without effective filtering, resulting in valuable computing and storage resources being occupied by low-quality data.

[0003] In terms of task scheduling, traditional mechanisms often employ static priority strategies or simple first-in-first-out logic, lacking quantitative assessment of multi-dimensional value such as image target areas, task timeliness, and changes. They also fail to establish a dynamic mapping between data value and the real-time load and storage status of heterogeneous computing nodes on the satellite. This rigid scheduling approach results in high-value data failing to receive priority access to high-performance computing loads, while low-value data is not promptly discarded to free up space, thus reducing the overall processing efficiency of hardware resources.

[0004] Furthermore, the existing on-board processing mechanism lacks sufficient correlation with the underlying hardware physical state. When the equipment is under extreme conditions such as excessive junction temperature or limited energy supply, continuous operation of heavy-load tasks can easily lead to hardware failures. There is a lack of degradation processing and field state caching mechanisms for high-load tasks. Moreover, in conventional queuing logic, medium-priority tasks are prone to long-term stagnation due to the long-term occupation of computing resources, making it difficult to balance the continuity of task execution with the physical security of the underlying system. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a lightweight pre-detection and hierarchical scheduling method for remote sensing images for onboard AI. This solves the problem that existing onboard remote sensing image processing is limited by computing and storage resources, making it difficult to process massive amounts of data in real time. Furthermore, due to the physical constraints of satellite platform power consumption, computing power limits, and downlink bandwidth, traditional indiscriminate data processing and downlink methods result in excessive resource consumption by invalid data and slow response times for critical tasks.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a lightweight pre-detection and hierarchical scheduling method for remote sensing images for onboard AI, comprising the following steps:

[0007] The working resolution image is obtained using the original high-resolution remote sensing image, and image quality evaluation parameters and background interference evaluation parameters are extracted. Based on the image quality evaluation parameters, the image quality score and quality gating coefficient are calculated.

[0008] Based on the working resolution image, morphological and intensity feature parameters are obtained, and the target index is calculated and output by combining the morphological and intensity feature parameters, background interference evaluation parameters, and image quality evaluation parameters.

[0009] Read the configuration file of the region of interest to generate a task region matching score, obtain historical reference images of the same region to perform comparison and generate a change score, and parse the urgency attribute of the satellite scheduling instructions to generate a task timeliness score;

[0010] The target index, image quality score, task region matching score, change score and task timeliness score are linearly weighted to generate an initial score. The initial score is then attenuated using a quality gating coefficient to generate a base score. The weak task region interference suppression logic is then activated for the base score, triggering a hard truncation mechanism for the upper limit of the score, and outputting a comprehensive score after pressure reduction.

[0011] Based on the comprehensive score, remote sensing images are divided into multiple value levels, and then hierarchical physical resource scheduling is performed.

[0012] Furthermore, the specific implementation method for obtaining the working resolution image from the original high-resolution remote sensing image is as follows: The working resolution image is obtained by performing a downsampling operation on the original high-resolution remote sensing image using a downsampling module; image quality evaluation parameters include cloud cover evaluation parameters, and the quality detection module generates a cloud interference flag based on the cloud cover evaluation parameters. Before calculating the image quality score and quality gating coefficient, the method further includes generating a cloud interference mask based on the working resolution image using the quality detection module, specifically including:

[0013] The working resolution image is converted to the HSV color space, and the pixel matrices of brightness, saturation, and hue channels are extracted. Pixels that simultaneously satisfy the bright and low saturation conditions are traversed and extracted to form candidate continuous regions, and the average gradient value and grayscale variance are calculated. Continuous regions that simultaneously satisfy the bright, low saturation, low gradient, and smooth texture conditions are marked as cloud interference regions, and finally a two-dimensional binarized cloud interference mask is generated. This invention identifies interference regions such as clouds, fog, or haze that affect image quality by combining bright, low saturation, low gradient, and smooth texture conditions in the color space, and uses basic pixel statistics to generate image quality evaluation parameters, reducing the computational overhead of deep learning models.

[0014] Furthermore, the cloud coverage evaluation parameters include the large-area cloud coverage ratio and the overall cloud coverage ratio. The quality detection module generates cloud coverage evaluation parameters based on the cloud interference mask, and generates cloud interference flags based on the cloud coverage evaluation parameters. Specifically, the quality detection module determines the overall cloud coverage ratio based on the ratio of the total number of pixels in the cloud interference region to the total number of pixels in the cloud interference mask; after performing connected component analysis on the cloud interference region and filtering out fragmented cloud blocks, the large-area cloud coverage ratio is determined.

[0015] The normalized sharpness parameter is calculated by combining the global gradient mean and gradient reference value, and a penalty term proportional to the overall cloud coverage ratio is subtracted to output the image quality score. When the proportion of large-area cloud coverage exceeds a preset cloud occlusion tolerance threshold, the quality gating coefficient is subjected to nonlinear attenuation and weight reduction calculation as the proportion of large-area cloud coverage increases, and a cloud interference flag in a triggered state is generated. This invention distinguishes between scattered clouds and blocky thick clouds, and introduces a nonlinear gating attenuation parameter to evaluate the impact of interference factors such as clouds, fog, or haze on image usability, preventing low-quality images from consuming subsequent computing resources.

[0016] Furthermore, the background judgment module extracts pixel sets within the hue range of vegetation or water bodies by using hard thresholding in the color space, and extracts pixel sets within the dark background range by using brightness judgment. For pixels in the hue range that satisfy the characteristics and whose grayscale variance is lower than the smooth texture threshold, a summation comparison is performed to generate the vegetation coverage ratio and the dark area coverage ratio of the water body. When the maximum value of both exceeds the background dominance threshold, the excess ratio magnitude is calculated and a positive background interference evaluation parameter is generated; when neither exceeds the threshold, it is set to zero. This invention identifies and measures the proportion of invalid backgrounds such as water bodies and vegetation, providing a negative suppression parameter for target-oriented evaluation to reduce the interference of background factors on scheduling evaluation.

[0017] Furthermore, the target estimation module calls multiple sets of basic computer vision operators to perform local feature extraction on the working resolution image and generate various local feature detection scores. Specifically, this includes: selecting local windows whose combined ranking of average gradient and gray-level variance response is within the range of the number of windows extracted, and mapping them to generate gradient target detection scores; extracting edge maps to generate candidate connected regions, and combining edge density and local intensity features to calculate and generate edge candidate detection scores.

[0018] This invention identifies regions with stable morphology, eliminates regions of abnormal size or those overlapping with cloud-interference masks, and generates stable bright and dark region detection scores. It also identifies local corner points whose grayscale variations exceed the neighborhood contrast threshold, statistically maps their density to generate local structural point detection scores, and calculates the visually salient region response based on color and brightness differences. After deducting cloud-mask-covered areas, it performs normalization processing to generate salient feature analysis scores. This invention employs a combination of basic computer vision operators to complete local feature detection, achieving response evaluation of potential target regions while reducing power consumption, thus adapting to onboard hardware constraints.

[0019] Furthermore, the target estimation module utilizes morphological and intensity feature parameters such as area, local contrast, aspect ratio of the bounding rectangle, fill rate, bounding box width and height, edge density, and shape constraint matching results to eliminate invalid noise points. A comprehensive calculation is then performed to generate candidate target analysis scores, and the number of valid candidate targets is tallied. Based on the morphological and intensity feature parameters, geometric structure matching analysis is conducted on the retained candidate connected regions to tally the number of structured targets with preset geometric structure features. This invention, through morphological filtering and geometric feature verification, obtains statistical results on the number of candidate targets and structured targets within a scene, providing fundamental data for evaluating target value.

[0020] Furthermore, the specific steps for calculating and outputting the targetability index are as follows: the targetability estimation module multiplies the gradient targetability detection score, edge candidate detection score, stable bright and dark region detection score, candidate target analysis score, local structure point detection score, and saliency feature analysis score by weight coefficients and sums them to calculate the basic feature score; when there is large-area cloud occlusion interference and the number of effective candidate targets is insufficient, a cloud interference suppression term is assigned.

[0021] When the background interference evaluation parameter is greater than zero and the number of valid candidate targets is insufficient, an effective background suppression deduction value is set. The basic feature score is added to the dense small target reward term, and the cloud interference suppression term and the effective background suppression deduction value are subtracted. The calculation result is then limited using a maximum and minimum value function, and the target index is output. This invention combines multiple local feature scores, cloud interference, and background suppression parameters to calculate the target index, and can adjust the image value assessment weights under cloud occlusion or background interference conditions.

[0022] Furthermore, the process of activating the weak task area interference suppression logic includes: calculating the structured target ratio based on the ratio between the number of structured targets and the number of valid candidate targets; triggering an interference risk flag when the number of valid candidate targets is higher than a preset interference risk assessment index, and the structured target ratio is lower than a preset lower limit threshold for the structured target ratio, accompanied by cloud interference and the task area matching score indicating that it is in a weak task-related area. This invention identifies and suppresses the high response phenomenon of false targets caused by environmental interference such as fragmented clouds through the structured target ratio and cloud coverage evaluation parameters, preventing non-task images from obtaining high scores.

[0023] Furthermore, the scheduling module reads the computing load rate, available video memory, storage capacity, current chip junction temperature, and real-time power consumption of each hardware computing node. When the comprehensive score meets the corresponding threshold conditions, the remote sensing images are classified into high-value, medium-value, or low-value levels. For image data entering the fidelity buffer queue, the accumulated queuing time is multiplied by a preset waiting time gain coefficient. The minimum value between the calculated result and the waiting time compensation upper limit is used to generate a waiting time compensation term. The comprehensive score is added to the waiting time compensation term to calculate the dynamic scheduling priority value, and the values ​​are arranged in descending order to generate the scheduling execution sequence. This invention combines hardware status to classify images and introduces a waiting time gain compensation mechanism to avoid long queuing delays for medium-value tasks due to computing power occupation, thus balancing scheduling priority and timeliness.

[0024] Furthermore, the process of implementing hierarchical physical resource scheduling includes: sequentially selecting high-value image data according to the execution sequence; guiding high-value image data to the high-performance processing payload on the satellite when the node load status meets the inference conditions; scheduling the task execution path of medium-value image data to low-power processing units for coarse-grained classification and screening, or placing it in the candidate cache for queuing; and directly discarding low-value image data when the storage capacity is detected to be lower than the preset capacity warning line. This invention designs task routing paths for heterogeneous computing nodes and dynamically allocates tasks of different value levels according to the node load and capacity status, thereby improving the utilization rate of on-board hardware and software resources.

[0025] Furthermore, the process of dynamically allocating on-board storage and computing resources also includes: real-time comparison of the current chip junction temperature and real-time power consumption with platform safety limits; when hardware thermal control exceeds limits or power supply is insufficient, sending hibernation or frequency reduction commands to processing nodes to interrupt currently running high-load inference tasks, and writing the task status to the fidelity cache or task recovery queue; for interrupted or unprocessed high-value image data, based on power supply and availability status, routing to the fidelity cache for queuing, and rescheduling to the on-board high-performance processing payload after the constraints are restored; if a low-power processing unit is called to participate in processing, it is only used for task status maintenance, auxiliary coarse screening, or temporary labeling, and does not replace high-performance fine inference after the constraints are restored. This invention performs task interruption and field protection when equipment thermal control exceeds limits or power is insufficient, performs fidelity caching, state preservation, and recovery scheduling for high-value tasks, and ensures hardware safety and task continuity.

[0026] Furthermore, the step of using the context evaluation module to obtain historical reference images of the same region and perform comparison to generate change scores specifically includes: image registration between the working resolution image and the historical reference images of the same region; calculating the absolute gray-level difference of the corresponding pixel position when the registration confidence reaches a threshold; marking pixels with gray-level absolute differences exceeding the difference threshold as changed pixels and counting the total number, calculating the proportion of changed pixels to generate a change rate parameter, and then converting it into a change score; when registration fails or the confidence is insufficient, gray-level difference comparison is not performed, and the change score is set to a preset neutral initial parameter. This invention quantifies the change rate parameter of the same region by comparing gray-level differences with historical reference images, increasing the scoring weight of changed region images in the overall evaluation.

[0027] Furthermore, the context evaluation module extracts the priority flag bit from the header of the task control message carried with the imaging task command or task queue, parses the encoded sequence to match the urgency attribute coefficient of the corresponding task, and queries a lookup table to convert and calculate the urgency attribute coefficient to generate a task timeliness score. This invention directly parses the urgency attribute in the satellite control message, enabling images with high timeliness requirements to obtain the corresponding scheduling priority.

[0028] Furthermore, the process of triggering the hard truncation mechanism for the scoring upper limit and outputting the comprehensive score after the pressure drop limit includes: calling multi-level task region thresholds and corresponding multiple truncation extreme values; when the task region matching score and the interference risk flag are at different trigger levels, the comprehensive score is forcibly assigned to the corresponding truncation extreme value, and the scoring upper limit is constrained within the truncation extreme value; when the interference risk flag is not triggered and the task region matching score reaches a higher-level task region threshold, the hard truncation mechanism is skipped. This invention sets multi-level region thresholds and truncation extreme values ​​to forcibly constrain the scoring upper limit of images in non-target regions or those that trigger interference risks, thus limiting the computational power consumption of low-relevance images.

[0029] Furthermore, the step of using the context evaluation module to read the region of interest configuration file and generate a task region matching score specifically includes: converting the corner points of the remote sensing image to a geographic coordinate system using real-time satellite telemetry parameters to construct the current field of view polygon; calculating the geometric intersection of the current field of view polygon and the specified region of interest polygon; determining the overlap area ratio based on the ratio of the physical area of ​​the intersection region to the total physical area of ​​the field of view; and multiplying the overlap area ratio by the task attention intensity score to calculate the task region matching score. This invention constructs a field of view polygon based on telemetry parameters, calculates the overlap area ratio through geometric intersection, and establishes a spatial matching relationship between the image evaluation and the preset region of interest.

[0030] Furthermore, the scoring gating module is used to multiply the target indicator, task region matching score, variation score, image quality score, and task timeliness score by their respective weighting coefficients and then sum them to generate the original weighted sum. A unit interval normalization operation is performed to obtain a proportional coefficient, which is then mapped and converted to a percentage range to generate the initial score. Finally, the initial score is multiplied by the quality gating coefficient and a proportional reduction limit is applied to calculate the base score. This invention normalizes multiple evaluation indicators and combines them with a quality gating coefficient for reduction calculation, uniformly converting feature parameters of different dimensions into a quantitative score for scheduling.

[0031] This invention provides a lightweight pre-detection and hierarchical scheduling method for remote sensing images for onboard AI. It offers the following advantages:

[0032] 1. This invention utilizes basic computer vision operators combined with color space thresholding to extract target features, image quality evaluation parameters, background interference evaluation parameters, and local morphological features from remote sensing image data. In the lightweight pre-detection stage, it reduces the reliance on complex neural network models that consume high computing power, and reduces the power consumption of feature extraction in the resource-constrained on-board environment. At the same time, it combines image quality gating, interference suppression, and background subtraction mechanisms to filter low-quality data, reducing the occupation of core computing and storage resources by invalid images.

[0033] 2. This invention generates a comprehensive score by weighted fusion and gating loss calculation of target indicators, image quality score, task area matching score, change score and task timeliness score, and reads the real-time load, storage capacity, thermal control status and energy supply status of the heterogeneous computing nodes on the satellite to perform hierarchical routing, allocating high-scoring images to high-performance processing payloads and discarding low-scoring images, thereby realizing dynamic matching of data of different values ​​with hardware physical status and improving the processing efficiency of on-board hardware resources.

[0034] 3. This invention monitors the chip junction temperature and power consumption in real time and compares them with the platform's safety limits. When thermal control exceeds the limit or energy is insufficient, it triggers the interruption of heavy-load tasks, downgrades the processing, or caches the field status. It also introduces queuing time gain compensation for the cache queue, ensuring hardware physical safety and task continuity under constrained operating conditions, while preventing medium-priority tasks from being stalled for a long time due to computing power congestion. Attached Figure Description

[0035] Figure 1 This is a flowchart of the lightweight pre-detection and hierarchical scheduling method for remote sensing images for on-board AI according to the present invention.

[0036] Figure 2 This is a flowchart of the multi-dimensional lightweight target estimation process of the present invention;

[0037] Figure 3 This is a flowchart of the hierarchical scheduling and on-board heterogeneous resource allocation of the present invention;

[0038] Figure 4 This is a bar chart showing the multidimensional sub-index feature response under different scenarios of the present invention;

[0039] Figure 5 This is a bar chart showing the comprehensive scoring and hierarchical scheduling results of the present invention.

[0040] Figure 6 This is an example image of a high-value scene in the core port area of ​​this invention, obtained through remote sensing.

[0041] Figure 7 This is an example image of a remote sensing image of a valuable scene in an urban building area according to the present invention;

[0042] Figure 8 This is an example image of a low-value scene in an open sea area, as presented in this invention. Detailed Implementation

[0043] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] See attached document Figure 1This invention provides a lightweight pre-detection and hierarchical scheduling method for remote sensing images for on-board AI. This method relies on a lightweight pre-detection and hierarchical scheduling system for remote sensing images for on-board AI. The system includes a downsampling module, a quality detection module, a background judgment module, a target estimation module, a context evaluation module, a scoring gating module, and a scheduling module. The scheduling control logic from the downsampling module to the scoring gating module and the scheduling module can run in low-power hardware processing units such as ARM, RISC-V, or FPGA. The high-performance target detection inference task is guided by the scheduling module to a graphics processor or a dedicated artificial intelligence acceleration chip for execution based on the resource status.

[0045] A lightweight pre-detection and hierarchical scheduling method for remote sensing images, geared towards on-board AI, performs feature parameter calculation and logical evaluation using basic image feature extraction operators. The method includes the following steps:

[0046] S10. Data Access and Quality Assessment. A downsampling module is used to perform downsampling on the original high-resolution remote sensing image to obtain a working resolution image. A quality detection module is used to extract image quality evaluation parameters. In one specific implementation, these parameters include cloud cover evaluation parameters, and the quality detection module generates a cloud interference flag based on these parameters. Specifically, the working resolution image is converted to the HSV color space, and pixel matrices of the luminance, saturation, and hue channels are extracted to generate a cloud interference mask. Based on the mask, cloud cover evaluation parameters including the proportion of large-area cloud coverage and the overall cloud coverage are output, and a cloud interference flag is generated. Then, image quality scores and quality gating coefficients are calculated based on the proportion of large-area cloud coverage and the overall cloud coverage. A background judgment module receives the working resolution image and statistically analyzes its grayscale variance and HSV color distribution. Vegetation coverage and water body dark area coverage are calculated, and background interference evaluation parameters are extracted. These parameters include scene background suppression factors generated based on the vegetation coverage and water body dark area coverage.

[0047] S20. Calculation of Multi-Source Target Indicators. The target estimation module receives the working resolution image, calls multiple sets of basic computer vision operators to perform local feature extraction on the working resolution image to statistically analyze the average gradient and gray-level variance response within a multi-scale window, calculates the edge density using the edge extraction operator in the basic computer vision operators, and filters out stable extreme value regions and local corner points. At the same time, it calculates the visually salient region response based on the color and brightness differences of the working resolution image.

[0048] The target estimation module extracts the connected components of the working resolution image, calculates the morphological and intensity feature parameters of each connected component, including area, local contrast, aspect ratio of the bounding rectangle and fill rate. Based on the morphological and intensity feature parameters, the number of effective candidate targets and the number of structured targets are counted, and the number of effective candidate targets and the number of structured targets are output as auxiliary gating parameters.

[0049] The target estimation module integrates the average gradient and gray-level variance response, edge density, stable extreme regions, local corners, visually salient region responses, and connected component features to calculate the basic feature score. When the number of candidate targets reaches the target quantity judgment index and the number of structured targets reaches the morphology judgment index, it accumulates and densifies the small target reward term. When the number of candidate targets does not reach the preset lower limit threshold, it combines the large area cloud coverage ratio in the image quality evaluation parameters and the scene background suppression factor in the background interference evaluation parameters to subtract the suppression term and output the target index.

[0050] S30. Context Feature Extraction. The context evaluation module extracts spatial and temporal contextual semantic information, reads the system's preset region of interest configuration file, calculates the overlap ratio between the current image and the specified region of interest, and generates a task region matching score based on the overlap ratio. The context evaluation module obtains historical reference images of the same region, performs grayscale difference comparison to generate a change score, and sets the change score to a preset neutral initial parameter when historical reference images are missing. The context evaluation module parses the urgency attribute of real-time satellite scheduling instructions and generates a task timeliness score based on the urgency attribute.

[0051] S40. Multidimensional Comprehensive Score Calculation. The scoring gating module receives the target index output by the target estimation module, the image quality score output by the quality detection module, and the task region matching score, change score, and task timeliness score output by the context evaluation module. The target index, image quality score, task region matching score, change score, and task timeliness score are linearly weighted to generate an initial score. The initial score is then attenuated using the quality gating coefficient output by the quality detection module to generate a base score. The scoring gating module activates the weak task area interference suppression logic for the base score. When the task region matching score indicates a weak task area, the number of candidate targets is higher than the preset interference risk assessment index, but the proportion of structured targets does not reach the preset proportion threshold, and the cloud interference flag is in the triggered state, a hard truncation mechanism for the upper limit of the score is triggered, and the comprehensive score after pressure reduction is output.

[0052] S50. Physical resource hierarchical scheduling. The scheduling module receives the comprehensive score, classifies the remote sensing image into high-value, medium-value, and low-value levels based on the comprehensive score and preset hierarchical thresholds, and performs hierarchical physical resource scheduling based on the classification results;

[0053] For high-value images, the scheduling module generates an early warning marker signal to guide the image data to the onboard high-performance processing payload. When the computing power load, hardware thermal control exceed the limit, or the energy supply status does not meet the conditions for high-performance fine-grained inference, the original resolution data of the high-value image data, or lossless or low-loss image data that meets the input requirements for high-performance fine-grained inference, is written to the fidelity buffer according to the restricted scheduling rules in the system configuration file. The corresponding task identifier, comprehensive score, value level identifier, and processing type to be executed are saved. After the above computing power, thermal control, or energy constraints are restored, the high-value image data is resent to the onboard high-performance processing payload according to the dynamic scheduling priority. If a low-power processing unit needs to be called to participate in processing during the restricted period, the processing is only used for task status maintenance, auxiliary coarse screening, or temporary marking, and does not change the fidelity storage status of the high-value image data in the fidelity buffer, nor does it replace the high-performance fine-grained inference after the constraints are restored.

[0054] For images of medium value, the scheduling module directs the image data to a lightweight classification model or candidate buffer. For images of low value, the scheduling module performs high-rate data compression, delayed computation, or discarding / removing actions, thus dynamically allocating on-board storage and computing resources. The fidelity buffer stores the original image data or low-loss compressed image data, and simultaneously stores task identifiers, comprehensive scores, value level identifiers, scheduling status information, completed processing stage markers, or recoverable intermediate processing states.

[0055] Step S10 in the lightweight pre-detection and hierarchical scheduling method for remote sensing images for on-board AI specifically includes the following sub-steps:

[0056] S11. Image Downsampling Processing. The downsampling module performs a downsampling operation on the original high-resolution remote sensing image to obtain a working resolution image. The downsampling module receives high-resolution image pixel data and uses a preset interpolation algorithm to scale the spatial dimensions of the image to a fixed working resolution size. The remote sensing image data in this invention includes a complete frame of remote sensing image, image blocks obtained by segmenting a complete frame of remote sensing image, or local image regions obtained using a sliding window method.

[0057] In one specific embodiment of the invention, the resolution of the original remote sensing image or image patch is scaled down from 512×512 pixels to the working resolution size (256×256 pixels); in the operating mode with a low-computing-power central processing unit on board, the working resolution size can be further scaled down to 128×128 pixels. In another specific embodiment, after the onboard optical payload captures a frame of original high-resolution remote sensing image of 4096×4096 pixels, it can be divided into multiple image patches, such as 512×512 pixel image patches.

[0058] The downsampling module generates working resolution images of 256×256 pixels or 128×128 pixels for each 512×512 pixel image block. Then, it calculates the targetability index, image quality score, task region matching score, variation score, task timeliness score, and comprehensive score for each image block, and determines the value level and performs hierarchical physical resource scheduling accordingly. The working resolution image reduces the total number of pixels in the input data, preserving the core image structure while reducing the computational load for subsequent feature extraction. The interpolation algorithm used for image scaling is a well-known technique in the field and will not be described further here.

[0059] S12. Color space conversion and mask generation. It should be noted that the image quality evaluation parameters are not limited to the cloud cover evaluation parameters in this embodiment. In other alternative embodiments, the image quality evaluation parameters may also include evaluation indicators used to characterize haze, blur, overexposure, shadows, noise, low-texture background, or other image quality degradation factors. For clarity, this embodiment uses the extraction of cloud cover evaluation parameters and the generation of cloud interference markers as examples for detailed explanation.

[0060] The quality inspection module converts the working resolution image to the HSV color space, extracts the pixel matrices of the luminance, saturation, and hue channels, and generates a cloud interference mask based on the pixel matrices. The quality inspection module acquires the RGB color channel information of the working resolution image and performs a non-linear transformation, mapping the RGB color channel information into independent hue, saturation, and luminance pixel matrices. Since clouds, fog, and haze in remote sensing scenes typically possess characteristics of high brightness, low saturation, continuous regions, and relatively smooth textures, and their edge gradient responses are usually weaker than those of real targets such as vehicles, ships, and aircraft, the cloud interference mask generation conditions in this embodiment include high brightness, low saturation, smoothness, and low gradient conditions.

[0061] The quality detection module traverses the brightness pixel matrix to filter pixels with brightness values ​​higher than a preset brightness threshold, and simultaneously filters the saturation pixel matrix to filter pixels with saturation values ​​lower than a preset saturation threshold. It calculates the average gradient value within the candidate continuous regions. When the average gradient value is lower than a preset gradient threshold, the corresponding candidate continuous region is determined to meet the low gradient condition. The module also verifies the texture smoothness within the spatial continuous region using grayscale variance. Continuous regions that simultaneously meet the conditions of high brightness, low saturation, low gradient, and smoothness are marked as cloud interference regions, ultimately generating a two-dimensional binarized cloud interference mask. The brightness threshold, saturation threshold, preset gradient threshold, and preset grayscale variance determination criteria are all calibrated based on statistical prior data from historical remote sensing images.

[0062] S13. Cloud Coverage Evaluation Parameter Statistics and Quality Assessment. As a specific implementation method for image quality evaluation parameters, the quality detection module outputs the large-area cloud coverage ratio and the overall cloud coverage ratio, and uses the large-area cloud coverage ratio and the overall cloud coverage ratio as cloud coverage evaluation parameters. Based on the large-area cloud coverage ratio and the overall cloud coverage ratio, the image quality score and quality gating coefficient are calculated.

[0063] Quality inspection module statistical cloud interference mask The total number of pixels marked as cloud interference is used to calculate the overall cloud coverage ratio. The overall cloud coverage ratio is calculated as follows: the total number of pixels in the cloud interference mask that are cloud interference markers is divided by the total number of pixels in the working resolution image.

[0064] Meanwhile, cloud interference masking Connectivity analysis is performed on the internal cloud interference region. After filtering out fragmented cloud patches with areas smaller than the area threshold, the remaining area is calculated to determine the proportion of large-area cloud coverage. The area threshold is calculated as a pixel area threshold at the working resolution image based on the pre-calibrated lower limit of the physical area of ​​the cloud cluster and the ground sampling distance or load calibration imaging ratio corresponding to the current image. The calculation logic for the large-area cloud coverage ratio is as follows: divide the total pixel area of ​​the remaining connected components after filtering out fragmented cloud blocks by the total number of pixels in the working resolution image.

[0065] The quality inspection module combines the global gradient distribution information of the working resolution image with the overall cloud coverage ratio. The image sharpness and effective field of view integrity are assessed by subtracting the percentage of cloud cover from the normalized sharpness parameter. A proportional penalty term is used to output image quality scores within a fixed range. The more severe the cloud cover, the worse the image quality. The lower the value, the richer the visually effective information within the image. In one embodiment, the quality detection module calculates the global gradient mean of the image at the current working resolution. And read gradient reference values ​​from the system configuration file. ,in It is obtained by statistical calibration of the global gradient mean from clear historical remote sensing images of the same payload, and Configured to a value greater than 0; when the system configuration file is missing or When the value is not greater than 0, the quality detection module uses the default gradient reference value calibrated before launch as... Image quality score is calculated using the following formula:

[0066] ;

[0067] ;

[0068] in, This is a normalized sharpness parameter; Both the overall cloud coverage ratio and the preset reference value can be obtained directly from the current image. This function represents the minimum value among the elements within the parentheses. This function represents the maximum value among the elements within the parentheses.

[0069] The computational logic aims to preserve the image quality baseline with an effective field of view; it only applies when global sharpness is low or cloud occlusion is severe. The quality score decreases as the normalized sharpness parameter decreases or the overall cloud coverage increases, and is truncated to 0 when the calculated result is below 0. For example, when the overall cloud coverage reaches a certain level but is within an acceptable range, the quality score can still maintain a high level (e.g., above 200), in which case the subsequent quality gating factor is mainly relied upon. The interference suppression logic is used to reduce its weight.

[0070] Furthermore, the quality inspection module incorporates a large-area cloud coverage ratio. Generate quality gating coefficients with values ​​ranging from 0 to 1.0. And generate cloud interference flags. .when At that time, the quality detection module will flag the cloud interference. Set to 1; when At that time, cloud interference flags Set to 0. Cloud interference flag. This is used by the scoring gating module to determine whether there is large-area cloud interference in the current image.

[0071] Quality gate coefficient The base value is 1.0, when the proportion of large-area cloud coverage is... Greater than the preset cloud occlusion tolerance threshold At that time, the quality gate coefficient With the proportion of large-area cloud coverage The nonlinear attenuation and weighting calculation is performed due to the increase in cloud occlusion tolerance threshold. The settings are determined by analyzing historical task statistics and written into the system configuration file. satisfy The quality gate factor is calculated according to the following rules:

[0072] when hour, ;

[0073] when hour, .

[0074] Using the above calculation method, the quality gate coefficient The value of is limited to between 0 and 1, and depends only on the proportion of large-area cloud coverage in the current image and the system's preset cloud occlusion tolerance threshold. The cloud occlusion tolerance threshold represents the upper limit of the proportion of large-area cloud occlusion that the spaceborne visual preprocessing task can tolerate. Quality Gating Coefficient This is used to suppress the final overall score to prevent low-quality images from receiving abnormally high scores.

[0075] S14. Scene background judgment and suppression factor generation. The background judgment module receives the working resolution image and statistically analyzes the grayscale variance and HSV color distribution of the working resolution image. It calculates the vegetation coverage ratio and the water body dark area coverage ratio, and extracts background interference evaluation parameters. In one specific embodiment, the background interference evaluation parameters include a scene background suppression factor generated based on the vegetation coverage ratio and the water body dark area coverage ratio.

[0076] In remote sensing scenes, vast farmlands, vegetated areas, low-texture water bodies, or dark background areas typically lack significant structural textures and exhibit continuous and uniform color or brightness distributions. The background judgment module extracts the hue and brightness pixel matrices from the working resolution image. It then uses hard thresholding in the color space to extract pixel sets within preset vegetation or water body hue intervals, and uses brightness judgment to extract pixel sets within preset dark background intervals. The vegetation, water, and dark background hue intervals are statically configured based on standard color feature distribution models of typical natural features in the color space. These intervals are statistically derived from the HSV color distribution of vegetation, water, and dark background areas in historical remote sensing samples and written into the system configuration file. Hard thresholding of color intervals is a well-known technique in this field and will not be elaborated upon here.

[0077] The background assessment module evaluates the texture smoothness of connected regions by combining the gray-level variance of local image patches. The preset smooth texture threshold is set based on the prior value of the gray-level fluctuation variance of uniform ground without target objects. Subsequently, pixels in regions whose tonal range meets vegetation characteristics and whose gray-level variance is lower than the smooth texture threshold are summed and compared to calculate the vegetation coverage ratio. Furthermore, for pixels whose color tone range meets the preset water body color tone range or whose brightness range meets the dark background characteristics, and whose grayscale variance is lower than the smooth texture threshold, the summation and comparison are performed to uniformly calculate and generate the water body dark area coverage ratio. .

[0078] When vegetation coverage ratio Coverage ratio of dark areas in water bodies The maximum value exceeds the preset background dominance threshold. At that time, the difference between the maximum value and the background dominance threshold is calculated as the excess ratio magnitude value, and a positive scene background suppression factor is generated. When the background dominance threshold is not exceeded At that time, the scene background suppression factor will be used. Set to 0. The background dominance threshold is... The average environmental proportion of valid targets in a conventional remote sensing field of view is set and written into the system configuration file. satisfy .

[0079] In one embodiment, when hour, ;when hour, Therefore, the scene background suppression factor Mapped to a scale of 0 to 255, consistent with other scores in the targetability index calculation. The background dominance threshold is used to define whether the current field of view is dominated by large areas of monotonous terrain features, set based on the statistical mean of the proportion of effective targets in a typical remote sensing view. Scene background suppression factor. Used to reduce the target index in subsequent processing to suppress false alarms caused by random noise in the smooth background region.

[0080] See attached document Figure 2 Step S20 in the lightweight pre-detection and hierarchical scheduling method for remote sensing images for on-board AI specifically includes the following sub-steps:

[0081] S21. Statistics on Local Average Gradient and Gray-Level Variance Response Based on Multi-Scale Windows. The target estimation module receives the working resolution image and calls multiple sets of basic computer vision operators to perform local feature extraction on the working resolution image to statistically analyze the average gradient and gray-level variance response within a multi-scale window. Specifically, based on the design concept of the BING algorithm, the target estimation module slides local windows of different sizes on the working resolution image, statistically analyzing the average gradient and gray-level variance response of pixels within each window. The set of local windows of different sizes is calculated based on the true physical size of typical targets in conventional remote sensing images and the ground sampling distance or payload calibration imaging ratio corresponding to the current remote sensing image.

[0082] The target estimation module calls a preset system configuration file to obtain the number of windows to extract. The number of windows to extract is determined by the pixel scale range of typical targets in the working resolution image, the target coverage requirement, and the computing power budget of the low-power processing unit. It then selects local windows whose combined ranking of average gradient and gray-level variance response falls within the range of this number of windows to extract, calculates the average response, and maps it to generate a gradient target detection score. Step S21 aims to quickly determine whether there is a region with a clear target outline in the image.

[0083] S22. Edge density calculation and separation of stable extreme value regions and local corner features. The target estimation module obtains the area lower limit threshold, area upper limit threshold, aspect ratio anomaly threshold, and neighborhood contrast threshold pre-calibrated based on historical remote sensing image datasets by calling the preset system configuration file, and performs the following three parts of lightweight feature calculation and separation:

[0084] The first part is edge density calculation: The target estimation module combines edge extraction operators from basic computer vision operators to extract edges from the working resolution image, and generates candidate connected regions based on the edge map or preliminary threshold segmentation results. Within each candidate connected region, the number of edge pixels is counted, and then the ratio of edge pixels to the total number of pixels in the connected region is calculated to determine the edge density. Combining the edge density and the local intensity features of the candidate connected regions, the target estimation module comprehensively calculates and generates an edge candidate detection score. This helps distinguish man-made targets with sharp edge structures from smooth cloud formations.

[0085] The second part is the feature separation of stable MSER regions: The target estimation module calls the MSER region extraction operator to find regions that maintain morphological stability under multiple grayscale thresholds, and filters out stable MSER regions. The target estimation module removes noise points with an area below the lower area threshold or background blocks with an area above the upper area threshold, as well as regions with an aspect ratio exceeding the aspect ratio abnormal threshold. It also obtains the cloud interference mask output by the quality detection module, calculates the overlap pixel ratio between each stable MSER region and the cloud interference mask, and excludes the corresponding stable MSER region when the overlap pixel ratio is higher than the preset cloud mask overlap threshold. The preset cloud mask overlap threshold is determined by the statistical results of the mask overlap ratio between historical cloud and fog region samples and real target region samples, and is written into the system configuration file. The target estimation module integrates the remaining effective and stable extreme value regions to generate stable bright and dark region detection scores. .

[0086] The third part is local corner feature separation: The target estimation module calls the corner detection algorithm to perform feature extraction in the working resolution image, filtering and separating local corners. The target estimation module identifies local corners in the image whose gray-level changes exceed the neighborhood contrast threshold, and calculates the density of local corners within the effective region. Since local corners are highly sensitive to texture, the target estimation module only maps the density of local corners to generate local structure point detection scores within the local area containing candidate connected regions. As supplementary evidence, it is used to reflect the texture richness of target clusters or complex man-made facilities.

[0087] S23. Calculation of Visually Satisfactory Region Response Based on Color and Brightness Differences. The target estimation module calculates the visually salient region response based on the color and brightness differences of the working resolution image. Specifically, the target estimation module converts the working resolution image to the LAB color space and calculates the color feature distance between the color feature vector of each pixel and the global average color feature vector of the entire image. The larger the color feature distance, the stronger the salience of the corresponding pixel compared to the background, and the higher the corresponding visually salient region response. The target estimation module subtracts the areas covered by cloud interference masks and normalizes the visually salient region responses of the remaining areas to generate a saliency feature analysis score. .

[0088] For the specific calculation processes of the BING algorithm, edge extraction operator, MSER extreme region extraction operator, FAST corner detection algorithm, and color space conversion, those skilled in the art can use well-known operators or equivalent lightweight operators such as gradient norm sorting, Sobel or Canny edge extraction, MSER region extraction, FAST corner extraction, and RGB to HSV or LAB color space conversion to implement them. The specific underlying calculation principles are well-known technologies in this field and will not be elaborated here. This invention focuses on the combined scoring of the aforementioned lightweight features, environmental gating, and on-board hierarchical scheduling logic.

[0089] S24. Extraction of Morphological and Intensity Feature Parameters Based on Connected Components. The target estimation module first calculates the local background mean of the image, identifies target regions that are brighter or darker than the local background, and determines whether the current region stands out more than the surrounding background. Subsequently, the target estimation module extracts the connected components of the working resolution image and calculates the morphological and intensity feature parameters of each connected component. The morphological and intensity feature parameters include area, local contrast, aspect ratio of the bounding rectangle, and fill rate. In order to comprehensively obtain the physical properties of the connected regions, the morphological and intensity feature parameters also include the bounding box width and height, edge density, and shape constraint matching results. The shape constraint matching results are determined based on the range of connected component area, the range of bounding rectangle aspect ratio, the range of fill rate, and the range of edge density. The above-mentioned range of connected component area, the range of bounding rectangle aspect ratio, the range of fill rate, and the range of edge density are all obtained by statistically calibrating the geometric and texture feature distribution of high-resolution remote sensing image datasets of historical typical targets (such as vehicles, ships, and aircraft) collected in advance and are pre-loaded into the system configuration file.

[0090] S25. Statistical Determination of the Number of Valid Candidate Targets and Structured Targets. The target estimation module calls a preset system configuration file to obtain the elongated shape threshold value. This threshold value is determined by statistically analyzing the aspect ratio, fill rate, edge density, and shape rationality distribution of the bounding rectangles of ships, vehicles, aircraft, and other man-made targets in historical remote sensing images, and is written into the system configuration file. The target estimation module removes invalid noise points based on morphological and intensity feature parameters and counts the number of valid candidate targets. Simultaneously, the target estimation module further determines the number of structured targets with preset geometric features (such as the slender outline of a ship, the rectangular structure of a vehicle, and the specific aspect ratio frame of an aircraft) by shape analysis within connected regions where the aspect ratio of the circumscribed rectangle exceeds the critical value for slender shapes. The target estimation module calculates and generates candidate target analysis scores based on the feature parameters of the preserved connected components. and the number of candidate targets With the number of structured targets As an auxiliary gating parameter, it is output to the scoring gating module.

[0091] S26. Conditional accumulation mechanism for reward items for dense small targets. The target estimation module obtains preset target quantity determination index and morphology determination index by calling the preset system configuration file. The target quantity determination index is determined by the statistical results of the number of valid candidate targets in historical dense small target image samples, and the morphology determination index is determined by the statistical results of the number of structured targets in historical dense small target image samples, and is written into the system configuration file.

[0092] During the feature score fusion process in the target estimation module, a conditional accumulation mechanism is set up: when the number of candidate targets reaches the target quantity judgment index and the number of structured targets reaches the morphology judgment index, the conditional accumulation mechanism of the dense small target reward item is activated to generate and accumulate the dense small target reward item. If the conditions are not met, then the intensive small goal reward items will be applied. The value is 0. Among them, the reward items for intensive small goals are... The preset reward score in the system configuration file ranges from 0 to 255 and is obtained by calibrating the basic feature score of manually confirmed target images from historical dense small target samples to meet the improvement requirements.

[0093] S27. The suppression term deduction and target index output are integrated with the background judgment and quality detection parameters. The target estimation module integrates the average gradient and gray-level variance response, edge density, stable extreme value regions, local corner points, visually salient region responses, and connected component features to calculate the basic feature score. This is then combined with a conditional accumulation mechanism for the dense small target reward term, and further reduced by the suppression term deduction using the cloud interference mask generated during image quality assessment and the scene background suppression factor in the background interference evaluation parameters. Finally, an anti-environmental interference target index is output. The target estimation module calls a pre-configured system configuration file to obtain the pre-calibrated weight coefficients of each feature operator according to different task scenarios. Based on these weight coefficients and the set suppression logic, the calculation process is performed using the following two formulas:

[0094] First, calculate the basic feature score based on the weight coefficients of each feature operator. The calculation formula is as follows:

[0095] ;

[0096] Subsequently, in basic feature segmentation Based on this, by combining dense small-objective reward and inhibition terms, and applying range constraints, the final target indicator is output. The calculation formula is as follows:

[0097] ;

[0098] in, Basic feature score; These are the average gradient and gray-level variance response weighting coefficients; The gradient target detection score; The edge density is used to derive the score weighting coefficient; The edge candidate detection score; To stabilize the scoring weight coefficients for bright and dark areas; To stabilize the detection score in both bright and dark areas; These are the weighting coefficients for the connected component feature scores; Analyze scores for candidate targets; The weighting coefficient for local corner point scores; The score is the local structural point detection score; The response weight coefficients for visually salient regions; The score is calculated based on the significance of the features.

[0099] , , , , and All feature scores are mapped to the range of 0 to 255. Each feature score is obtained by linearly normalizing and truncating the corresponding original response value according to the preset response lower limit and preset response upper limit. The preset response lower limit and preset response upper limit are determined by the statistical distribution of the corresponding feature responses in historical remote sensing samples and written into the system configuration file. to All are non-negative weight coefficients, and satisfy the following conditions: The weighting coefficients are calibrated by comparing the pre-detection results of historical remote sensing samples with the manually labeled results and then written into the system configuration file.

[0100] For target indicators; Rewards for intensive, small-scale goals; This is a cloud interference suppression term generated based on image quality evaluation parameters; The scene background suppression factor is a parameter in the background interference evaluation parameters output by the background judgment module. The effective background suppression deduction value is substituted into the target indicator calculation formula. Depend on Number of candidate targets The number of effective candidates is determined together with the lower limit threshold.

[0101] The feature value is used to implement dynamic reward and punishment adjustment based on basic feature scores, including cloud interference suppression, effective background suppression, and dense entity weighting. To constrain the dynamic reward and punishment adjustment feature value to the truncation extreme value limit within the effective value range of 0 to 255, the output result is limited.

[0102] When the target estimation module performs suppression term deduction to construct the input parameters of the formula, it first calls the preset system configuration file to obtain the lower limit threshold of the number of effective candidates and the preset penalty score, and then obtains the cloud interference flag output by the quality detection module. and the scene background suppression factor generated and output by the background judgment module. Among them, cloud interference indicator The quality inspection module determines the cloud coverage ratio over a large area. Cloud occlusion tolerance threshold The comparison results are generated, and the cloud occlusion tolerance threshold is determined. The effective candidate number lower limit threshold is determined by the statistical results of the large-area cloud coverage ratio of historical cloud-occluded image samples and written into the system configuration file; the effective candidate number lower limit threshold is determined by the statistical results of the number of candidate targets in historical real target image samples; the preset penalty score is calibrated by the target score distribution of false alarm images in historical large-area cloud-occluded samples and written into the system configuration file.

[0103] Specifically, when cloud interference flag When the target estimation module is in a triggered state and there is large-area cloud obstruction interference, and the number of candidate targets is insufficient to reach the lower limit threshold of the effective candidate number, the target estimation module will use the preset penalty score as the cloud interference suppression term. Assign a value; otherwise, remove the cloud interference suppression term. The value is assigned to 0;

[0104] At the same time, when scene background suppression factor When the value is greater than 0 (i.e., the field of view is dominated by a large area of ​​monotonous ground features), and the number of candidate targets is also insufficient to reach the lower limit threshold of the effective candidate number, the target estimation module will subtract the effective background suppression value. Set as This is used for subsequent deduction calculations; otherwise (i.e., when the number of candidate targets reaches the lower limit threshold of the number of effective candidates), the target estimation module will deduct the effective background suppression value. Set to 0, and maintain the scene background suppression factor output by the background judgment module. The cloud interference suppression term remains unchanged to avoid falsely penalizing areas with truly dense targets. and effective background suppression deduction value After the assignment, the target estimation module will , , and Substituting the values ​​into the target calculation formula, and after being constrained by the maximum and minimum value functions, the final target index is output. .

[0105] Step S30 in the lightweight pre-detection and hierarchical scheduling method for remote sensing images for on-board AI includes: extracting spatial and temporal contextual semantic information using a context evaluation module, specifically including the following sub-steps:

[0106] S31. Spatial Dimension Task Region Matching Evaluation. The context evaluation module reads the system's pre-set Region of Interest (ROI) configuration file to extract spatial contextual semantic information. The ROI configuration file can configure task-focused regions of interest for the global or specified image. Each ROI carries a preset task focus intensity score, which is pre-assigned by the ground control station based on the timeliness and strategic importance of the observation task, representing the relevance of the ROI to the current observation task. The ROI configuration file contains the geographic latitude and longitude coordinate boundary information of the specific observation task area annotated by the ground control station.

[0107] The context evaluation module obtains the real-time satellite telemetry parameters corresponding to the current remote sensing image, and uses these parameters to transform the four corner points of the current remote sensing image to the geographic coordinate system to construct the current field of view polygon. The transformation mapping process from remote sensing image pixel coordinates to geographic coordinates is a well-known technique in this field and will not be elaborated here.

[0108] Subsequently, the context evaluation module calculates the geometric intersection of the current field-of-view polygon and the region of interest polygon specified in the region of interest configuration file to statistically analyze the physical area of ​​the intersection region. It then divides the physical area of ​​the intersection region by the total physical area of ​​the current field-of-view polygon to calculate the overlap ratio between the current image and the specified region of interest. Task Attention Intensity Score The task region matching score is calculated and generated according to the following formula. :

[0109] ;

[0110] in, This is a function that takes the minimum value and is used to match and divide the task region. The numerical range is constrained to be between 0 and 255. The overlap area ratio reaches its maximum when the current image is entirely located within a region of high task interest, such as a port or airport (i.e.,...). (equal to 1), the highest value of the task attention intensity score. Make the task area matching division The maximum value can reach 255; when the current image is located in a low-task area such as ocean or farmland, a lower configuration is corresponding to each. The score is used to calculate a lower value, such as 100 for ocean or farmland. Task area matching score. This indicates the degree to which the current image meets the requirements for ground observation in the spatial dimension. When there are multiple regions of interest in the region of interest profile, the context evaluation module calculates the matching score for each region of interest and takes the maximum value as the task region matching score for the current image. .

[0111] S32. Extraction of Historical Image Change Features in the Time Dimension. The context evaluation module queries the onboard storage medium to obtain historical reference images of the same region. Upon successful retrieval of a historical reference image, image registration is performed between the current working resolution image and the historical reference image to ensure spatial pixel-level alignment. If the historical reference image is successfully retrieved but image registration fails or the registration confidence level is lower than a preset registration confidence threshold (wherein the preset registration confidence threshold is set by statistical analysis of the correct registration probability from historical remote sensing image feature matching experiments and written into the system configuration file), the context evaluation module does not perform grayscale difference comparison and will not perform grayscale difference comparison. Set to preset neutral initial parameters. Historical reference images for the same area are remote sensing images of the same geographical location taken during previous transits and orthorectified.

[0112] For feature matching and spatial registration between two remote sensing images, those skilled in the art can extract image feature points and calculate homography matrix for resampling. The image registration process between two remote sensing images is a well-known technique in the field and will not be described in detail here.

[0113] After pixel-level registration is completed, the context evaluation module performs gray-level difference comparison, calculates the absolute difference in gray levels at corresponding pixel locations between the two images, marks pixels with absolute gray-level differences exceeding a preset difference threshold as changed pixels, and then counts the total number of changed pixels to calculate the ratio of the total number of changed pixels to the total number of pixels in the entire image, generating a change rate parameter. .

[0114] The preset difference threshold is calculated by statistically analyzing historical, unchanged images of the same region to extract the system noise standard deviation, and a predetermined multiple of this standard deviation is set as the difference threshold, with the multiple set between 2 and 3. Combining this with the magnitude of the rate of change parameter, the context evaluation module converts the rate of change parameter into a change score using the following formula. : .

[0115] The greater the difference between the image before and after, the larger the value of the rate of change parameter, and the greater the corresponding change score generated by the context assessment module. The higher. Change points It is applicable to tasks such as disaster assessment, illegal construction monitoring, military deployment, port changes, and road changes. Examples include assessing pre- and post-disaster building damage, identifying newly appearing vessels in ports, monitoring road blockages, and analyzing changes in newly constructed areas. When historical reference images are missing due to onboard storage limitations, initial image registration failure, or registration confidence levels below a preset threshold, the context assessment module will analyze the changes. The initial parameters are set to preset neutral parameters (e.g., a fixed constant of 128), where the fixed constant of 128 is the midpoint of the score scale from 0 to 255. The neutral initial parameters can be written into the system configuration file according to the task's conservative requirements to ensure the continuous and stable operation of the overall evaluation logic.

[0116] S33. Task Timeliness Classification and Task Timeliness Score Generation. The urgency attribute of real-time satellite scheduling instructions is parsed using the context evaluation module. Priority flags are extracted from the task control message headers annotated with imaging task instructions or synchronously carried by the onboard task queue. The encoded sequences of different levels, such as routine inspection, key monitoring, and emergency response, contained in the priority flags are then parsed to match the corresponding task's urgency attribute coefficient. The urgency attribute coefficient is a weighted parameter between 0 and 1, used to characterize the timeliness requirements of the observation task.

[0117] In one embodiment of the present invention, the context evaluation module has a built-in fixed lookup table, and queries the lookup table according to the parsed encoding sequence to obtain the corresponding urgency attribute coefficient, and then calculates and generates the task timeliness score according to the following formula. : .in, The output range is constrained to be between 0 and 255, which is the task timeliness score. This refers to the urgency attribute coefficients retrieved from a lookup table. The lookup table defines the correspondence between different priority flags and urgency attribute coefficients.

[0118] Task time limit This is used to indicate the urgency of a ground observation mission. For the same image, a lower urgency attribute coefficient corresponds to a mission timeliness score generated by the context evaluation module under a routine patrol mission. The low priority level allows the image data to be included in the candidate cache; the urgency attribute coefficient, which approaches 1 in high-priority states such as disaster emergencies, corresponds to the task timeliness score generated by the context evaluation module. The higher resolution makes the image data more likely to be prioritized for processing, making it suitable for on-board dynamic resource scheduling scenarios.

[0119] Step S40 in the lightweight pre-detection and hierarchical scheduling method for remote sensing images for on-board AI specifically includes the following sub-steps:

[0120] S41. Multidimensional feature linear weighting and initial score generation. The scoring gating module receives the target index output by the target estimation module and the image quality score output by the quality detection module, and receives the task region matching score, change score and task timeliness score output by the context evaluation module.

[0121] The scoring gating module obtains the set weight coefficients, performs a linear weighted calculation on the target indicator, task region matching score, variation score, image quality score, and task timeliness score, and maps the weighted sum from a numerical range of 0 to 255 to a percentage range to generate an initial score. The formula for calculating the initial score is as follows:

[0122] ;

[0123] in, The initial score is out of 100. For target indicators; Match scores to the task area; Divide into categories of change; Image quality score; Assign time points to tasks; The weighting coefficient corresponding to the target indicator is set to 0.55. The weight coefficient corresponding to the task region is set to 0.15; The weighting coefficient corresponding to the change score is set to 0.10; The weighting coefficient corresponding to the image quality score is set to 0.10. The weighting coefficient corresponding to the task timeliness score is set to 0.10;

[0124] The original weighted sum result is used to integrate the target, spatial region, temporal variation, image quality and timeliness requirements; This is the scaling factor after normalizing the original weighted sum to a unit interval.

[0125] The weighting coefficients of each indicator are calibrated before launch based on the specific payload type of the satellite and the routine observation needs of the ground tracking and control stations, and are dynamically updated via commands during on-orbit operation. The dynamically updated... to All are non-negative numbers and satisfy the following conditions: Through linear weighted calculation, the scoring gating module integrates the low-level visual features of the image with the high-level spatial-temporal semantics of the task within the same quantization dimension.

[0126] S42. Score Attenuation Adjustment Based on Quality Gating Coefficient. The scoring gating module receives the quality gating coefficient output from the quality detection module and uses it to perform attenuation adjustment on the initial score to generate a base score. The calculation formula is as follows: ;in, Basic score; This is the initial score; This is the quality gate coefficient.

[0127] By introducing a quality gating coefficient that is directly multiplied by the initial score, the scoring gating module can reduce the weight of contaminated image data at the global level when the image quality is poor or there is significant cloud occlusion. This prevents residual local strong edges or significant features in the image from being faked as high targetability and thus obtaining an unreasonablely high initial score.

[0128] S43. Weak Task Area Interference Flag Determination. The scoring gating module activates the weak task area interference suppression logic based on the basic score. The scoring gating module obtains the number of candidate targets and the number of structured targets counted by the target estimation module, calculates the structured target ratio by dividing the number of structured targets by the number of candidate targets, and reads the cloud interference flag generated by the quality detection module based on image quality evaluation parameters. and overall cloud coverage ratio When the number of candidate targets When the value is 0, the scoring gating module sets the structured target ratio to 0; when the number of candidate targets is 0... When the value is greater than 0, the proportion of structured targets is determined according to... calculate.

[0129] In specific physical scenarios such as ocean clouds, wave reflections, or complex farmland edges, basic computer vision operators are prone to misidentifying discrete interference patches as a large number of candidate targets. Real vehicles and ships typically possess relatively stable geometric features. The scoring gating module detects a higher number of candidate targets than a preset interference risk assessment index, and a structured target ratio lower than a preset lower limit threshold, while cloud interference indicators are also present. In a triggered state or overall cloud coverage ratio When the preset fragmented cloud interference threshold (used to characterize the presence of discontinuous highlights or fragmented cloud interference) is reached, and the task region matching score indicates that the current image is in a weakly task-relevant region, the image is determined to have natural environmental pseudo-feature interference, triggering an interference risk flag. The interference risk assessment index, the lower limit threshold of the structured ratio, and the fragmented cloud interference threshold are statistically analyzed and calibrated by the ground control center based on historical scene image sample sets, and are preset as system configuration parameters and imported into the scoring gating module.

[0130] S44. Score Upper Limit Hard Truncation and Comprehensive Score Output. The score gating module triggers a truncation mechanism based on the score upper limit for image data that triggers interference risk flags or is located in non-core areas of interest, outputting a comprehensive score after voltage drop limitation. .

[0131] The scoring gating module reads the task region matching score and calls the system's preset multi-level task region thresholds. The multi-level task region thresholds (including the lowest-level task region threshold and higher-level task region thresholds) and the corresponding cutoff extreme values ​​at each level (including the first cutoff extreme value, the second cutoff extreme value, and the third cutoff extreme value) are calculated or calibrated based on the theoretical upper limit of the credibility score under different task region constraints, the comprehensive score distribution of historical false alarm samples, the task region level configuration, and the results of manual review, and are pre-fixed in the system.

[0132] After determining the conditions for truncation of weak task regions, edge task regions, and interference risk, when two or more truncation conditions are met simultaneously, the scoring gating module compares the extreme values ​​of each triggered condition and selects the smallest extreme value as the current scoring cap; if the base score... If the score exceeds the current scoring limit, the overall score will be adjusted. The value will be assigned to the current maximum score; otherwise, the overall score will be calculated. Assignment as a basis for scoring .

[0133] When the task region matching score indicates that the current region is a weak task region, a hard truncation operation is performed. The specific truncation logic is as follows:

[0134] When the task region matching score is less than the lowest-level task region threshold (e.g., the lowest-level task region threshold is set to 110), the scoring gating module determines that the current image is in a background region with weak task relevance. At this point, if the base score is higher than the first cutoff extreme value (e.g., the first cutoff extreme value is set to 54), the scoring gating module will apply a comprehensive score. The base score is assigned the first cutoff extreme value; if the base score is lower than or equal to the first cutoff extreme value, the composite score will be adjusted. Assign a value to the original base score.

[0135] When the task region matching score is greater than or equal to the lowest-level task region threshold and less than the higher-level task region threshold (e.g., the higher-level task region threshold is set to 180), the scoring gating module determines that the current image is located at the edge of the observed target. At this point, if the base score is higher than the second cutoff extreme value (e.g., the second cutoff extreme value is set to 66), the scoring gating module will perform a comprehensive scoring. The base score is assigned the second cutoff extreme value; if the base score is lower than or equal to the second cutoff extreme value, the composite score will be adjusted. Assign the original value as the base score. The value of the second cutoff extreme value is greater than that of the first cutoff extreme value.

[0136] The value of the third cutoff extreme value (e.g., 44) is preset to be less than the first cutoff extreme value (e.g., 54) and the second cutoff extreme value (e.g., 66). When the interference risk flag is triggered, it has absolute deweighting priority, and the scoring gating module calls the third cutoff extreme value to force the comprehensive score to be reduced. The upper limit constraint is within the third cutoff extreme value, thereby directly suppressing the score of image data with high-density texture interference to the low-value range.

[0137] For image data that does not trigger the interference risk flag and whose task region matching score is greater than or equal to the threshold of a higher-level task region, thus indicating that it is in the core task region, the scoring gating module skips the hard truncation mechanism and directly uses the basic score as the comprehensive score. Output. Overall score. Finally, the output is in the form of normalized values ​​to ensure that image data in different scenes and at different task levels have a unified comparison benchmark.

[0138] S45. Encapsulated output of the comprehensive score and auxiliary evaluation indicators. The scoring gating module will truncate and limit the comprehensive score. The data, along with quality gating coefficients, interference risk indicators, and mission area matching scores, is encapsulated in a data structure and output to the scheduling module. The scoring gating module only provides a quantitative comprehensive evaluation benchmark and risk indicator. The specific physical resource routing, cache migration, compression or elimination actions, and their corresponding value level classifications are ultimately executed by the scheduling module in step S50 based on the comprehensive score and the real-time onboard resource status. For routine data compression processing and high-ratio compression of image data, those skilled in the art can use JPEG2000 or CCSDS image data compression standards for encoding. The specific onboard image compression mechanisms are well-known technologies in the field and will not be elaborated here.

[0139] See attached document Figure 3 Step S50 in the lightweight pre-detection and hierarchical scheduling method for remote sensing images for on-board AI specifically includes the following sub-steps:

[0140] S51. Real-time Status Monitoring of On-board Heterogeneous Resources. The scheduling module acquires the current hardware operating parameters of the on-board heterogeneous processing platform in real time, while simultaneously recording the processing time during image block processing. The processing time is the actual measured time in a single run, reflecting the processing time of the image block under current hardware and parameter conditions, and is not used as a fixed threshold or fixed algorithm parameter. The on-board heterogeneous processing platform includes various low-power processing units such as ARM architecture processors, RISC-V architecture processors, or field-programmable gate arrays, as well as graphics processors or dedicated AI acceleration chips for performing high-load tasks. The scheduling module polls each hardware computing node via the internal system bus, reading the computing load rate, available video memory, on-chip or external storage margin, current chip junction temperature, and real-time power consumption of each hardware computing node. The communication and status reading process of the scheduling module using integrated circuit bus protocols or serial peripheral interface protocols to obtain underlying register values ​​is a well-known technology and will not be elaborated further.

[0141] S52. Dynamic priority construction based on comprehensive score and queuing time. The scheduling module receives the final comprehensive score output by the scoring gating module and presets high-value threshold and medium-value threshold.

[0142] The high-value and medium-value thresholds are pre-calculated and calibrated by engineers based on the critical conditions of the gating constraint logic. Since the upper limit of the comprehensive score when the task region matching score is below 180 is constrained to 66, the high-value threshold is set to a value greater than 66 to ensure that only strong task region images can trigger high value, for example, the high-value threshold is set to 80; the medium-value threshold is set to 60 to distinguish general candidate targets from low-value interference.

[0143] The scheduling module makes a logical judgment based on the final comprehensive score: when the judgment condition is met... When the conditions are met, remote sensing images are classified into high-value categories; when the criteria are met... At that time, remote sensing images are classified into medium value levels; when the judgment criteria are met... At that time, remote sensing images were classified into low-value categories. Among them, This indicates the final overall score. Indicates a high-value threshold. This represents the value threshold.

[0144] After completing the multi-dimensional value level classification, the scheduling module calculates the dynamic scheduling priority for image data entering the cache queue and waiting to be processed. High-value image data corresponds to the fidelity cache area, while medium-value image data corresponds to the candidate cache area or the lightweight processing queue. This prevents data backlog at the end of the cache queue due to prolonged inactivity. The dynamic scheduling priority calculation formula is constructed as follows:

[0145] ;

[0146] in, This indicates the final overall score; This indicates the cumulative queuing time, and its unit of measurement is set by the system configuration file to one of seconds, minutes, or system scheduling cycles. This represents the waiting time compensation item subject to upper limit constraints. This indicates the upper limit of waiting time compensation. Its value is determined by the task's tolerable latency and the cache scheduling strategy and written into the system configuration file. This represents the dynamic scheduling priority value; This represents the preset waiting time gain coefficient, used to indicate the score compensation amount corresponding to a unit waiting time. Its value is calculated and calibrated by the onboard data management system based on the system clock cycle and the mission tolerance delay duration. and They are all on the same score scale. At the beginning of each hardware computing cycle, the scheduling module traverses the backlog of image data in the fidelity buffer, candidate buffer, or lightweight processing queue, and sorts them in descending order according to the calculated dynamic scheduling priority to generate a real-time scheduling execution sequence.

[0147] S53. High-performance payload guidance and early warning marking for high-value images. The scheduling module selects high-value image data sequentially according to the scheduling execution sequence, and evaluates the current computing load rate, available video memory, on-chip storage margin, or external storage margin of the graphics processor and dedicated AI acceleration chip. If the computing load rate is below the full load threshold and the available video memory, on-chip storage margin, or external storage margin meets inference requirements, the scheduling module generates an early warning marking signal, guiding the image data to the on-board high-performance processing payload, and simultaneously sending a storage physical address pointer. The on-board high-performance processing payload is specifically represented by a graphics processor or a dedicated AI acceleration chip. The full load threshold is calibrated by engineers before satellite launch based on stress tests conducted on the heat dissipation characteristics and concurrent processing capabilities of heterogeneous computing nodes; for example, it may be set to 85% of the maximum computing load.

[0148] A graphics processing unit (GPU) or dedicated AI acceleration chip receives the image data stream and invokes an internally deployed high-performance AI model to perform a deep object detection task. In one embodiment of the invention, the high-performance AI model employs a convolutional neural network based on a single-stage detection architecture. Its internal hierarchical structure sequentially includes a backbone network for extracting image features, a feature pyramid neck network for multi-scale feature fusion, and a detection head for outputting prediction results. In the context of onboard remote sensing scenarios, the input data needs to be preprocessed, cropped, or scaled into a fixed-dimensional image tensor, such as a 256×256 pixel RGB three-channel image. The model's output represents specific business meaning and physical state, specifically including: target category, bounding box coordinates, and confidence score.

[0149] The construction and training steps of the high-performance artificial intelligence model are executed in a ground-based data center. Remote sensing images containing various ground targets are collected as a training sample set and manually labeled to generate true bounding boxes and category labels. During the training phase, the training samples are input into the model, and the intersection-union ratio (IU) is calculated using the predicted bounding boxes and the true bounding boxes output by the model. An overall loss function is constructed by combining the classification cross-entropy loss and the bounding box regression loss. Subsequently, the convolutional kernel weight parameters in the network are updated using the backpropagation algorithm until the model loss converges to a stable state. After training, the model parameters are fixed and uploaded for deployment on satellite. The use of the automatic differentiation mechanism and optimization algorithm of deep learning frameworks to implement gradient calculation and backpropagation updates of the neural network model is a well-known technique and will not be elaborated further.

[0150] S54. Lightweight processing and data removal / compression of low-to-medium value images. The scheduling module allocates the task execution path for medium-value image data to low-power processing units. This allows the medium-value image data to either enter a lightweight classification model for coarse-grained classification and screening, or to queue it in a candidate buffer to await transmission to the ground data center for analysis. The lightweight classification model has fewer network layers and parameters than high-performance AI models, enabling rapid identification of potential target outlines.

[0151] For low-value image data—that is, data deemed invalid by the system to be heavily obscured by clouds, pure water bodies, or severely interfered with—the scheduling module executes degradation processing logic. This involves directly calling image processing operators to perform high-rate data compression, adopting a delayed calculation strategy, or executing a discard / removal action. Specifically, the discard / removal action is triggered when the available video memory, on-chip storage capacity, or external storage capacity falls below a preset capacity warning threshold (i.e., the removal condition is met). This directly discards the low-value image data, thereby freeing up storage space. The capacity warning threshold is pre-defined by the on-board data management system based on the total physical capacity of on-chip or external memory and data security management policies; for example, it may be set to 10% of the total capacity.

[0152] S55. A dynamic routing mechanism for a fidelity buffer based on hardware thermal control constraints. The scheduling module compares the monitored current chip junction temperature and real-time power consumption with the system safety thermal limits and power consumption limits set by the satellite platform in real time. The system safety thermal limits and power consumption limits are fixed safety parameters calibrated at the factory based on the physical limits of the hardware by the satellite thermal control system and power management system. When the hardware thermal control exceeds the limit or the system is in the shadow zone, resulting in insufficient power supply, the scheduling module triggers a hardware protection strategy. The scheduling module sends a hibernation or frequency reduction command to the high-power graphics processor or dedicated AI acceleration chip, while interrupting the currently running high-load inference task, and writes the image data, task identifier, scheduling status information, completed processing stage markers, or recoverable intermediate processing states corresponding to the interrupted task into the fidelity buffer or task recovery queue.

[0153] For interrupted high-value image data and unprocessed high-value image data, the scheduling module reads the restricted scheduling rules in the system configuration file and makes a judgment based on the current chip junction temperature, real-time power consumption, energy supply status, and the availability status of low-power processing units. The high-value image data is directly routed to the fidelity buffer to queue and wait, and the task identifier, comprehensive score, value level identifier, completed processing stage mark, or recoverable intermediate processing state are recorded. After the chip junction temperature, real-time power consumption, or energy supply status recovers to meet the high-performance inference conditions, the data is rescheduled to the high-performance processing load according to the dynamic scheduling priority.

[0154] If a low-power processing unit is invoked during the fidelity caching period, only auxiliary processing, state maintenance, or temporary coarse screening that does not destroy the fidelity data will be performed. This will not be considered the final, refined inference result for high-value image data, nor will it replace the processing path subsequently rescheduled to a high-performance processing payload. This invention performs task interruption and field protection when the equipment's thermal control exceeds limits or its energy is insufficient. It performs fidelity caching, state preservation, and recovery scheduling for high-value tasks, ensuring hardware safety and task continuity until the satellite platform's thermal control and energy status return to normal operating conditions.

[0155] To intuitively demonstrate the system's decision-making basis for hierarchical scheduling of image data, the system underwent simulation evaluations for three typical remote sensing imaging scenarios, combined with... Figures 4 to 8 This paper provides a joint explanation of feature quantification and final judgment results under different scenarios. Based on the above application scenarios, a specific processing embodiment for urban building area (medium value) image data is provided to elaborate on the operational logic of the lightweight pre-detection and hierarchical scheduling method for remote sensing images for on-board AI.

[0156] Data Access and Quality Assessment. The onboard complementary metal-oxide-semiconductor (CMOS) optical payload captures a raw high-resolution remote sensing image (4096×4096 pixels in this example, an urban building area scene). The image data is transmitted to the onboard ARM Cortex-A53 low-power processing unit via a direct memory access channel. According to a multi-level engineering implementation, the raw high-resolution remote sensing image of 4096×4096 pixels is first divided into multiple 512×512 pixel image blocks. This embodiment uses one of the 512×512 pixel image blocks containing rich urban building features as an example: the downsampling module uses a bilinear interpolation algorithm to perform a downsampling operation on the 512×512 pixel image block to obtain a working resolution image, whose size is scaled down to 256×256 pixels.

[0157] The quality inspection module reads the RGB channel data of the working resolution image, converts the image to the HSV color space, performs a non-linear transformation, and extracts the pixel matrices of the luminance, saturation, and hue channels. It then maps the RGB color channel information into independent hue, saturation, and luminance pixel matrices. In this urban building area scene, the weather is clear and cloudless. The quality inspection module extracts continuous regions from the luminance pixel matrix where the luminance value is higher than a preset luminance threshold (e.g., 180) and from the saturation pixel matrix where the saturation value is lower than a preset saturation threshold (e.g., 50). It calculates the average gradient value of the pixels within these regions. When the average gradient value is lower than a preset gradient threshold, the corresponding candidate continuous region is deemed to meet the low gradient condition. The module also verifies the texture smoothness within the spatially continuous region using grayscale variance to determine if the conditions of high brightness, low saturation, low gradient, and smoothness are simultaneously met. Since no smooth, high-brightness regions matching cloud characteristics were extracted, a two-dimensional binarized cloud interference mask with all values ​​null (no cloud markers) was ultimately generated. .

[0158] The overall cloud coverage ratio was obtained by statistically analyzing the connected component pixels. The percentage of large-area cloud coverage is 0%. The value is 0%. The quality detection module will flag cloud interference. Set to 0. Due to the high image sharpness, the normalized sharpness parameter is calculated based on the global gradient mean. The image quality score is 0.94, calculated by the quality detection module. Rounded down, the output is 240. Because... Quality gate coefficient The value is assigned to 1.0.

[0159] Scene background assessment. The background judgment module calculates the scene background suppression factor for the working resolution image. This module extracts the hue and luminance pixel matrices from the working resolution image, extracts pixel sets within preset vegetation or water hue ranges using hard thresholding in the color space, and extracts pixel sets within preset dark background ranges using luminance judgment. Simultaneously, it assesses the texture smoothness of connected regions by combining the grayscale variance of local image patches. Pixels in regions that meet the conditions and whose grayscale variance is below a preset smooth texture threshold are summed and compared. The vegetation coverage ratio within the working resolution aerial image of the city is then statistically determined. The coverage rate of dark areas in rivers and other water bodies is 5%. 10%, The value is 10%, which does not exceed the system's preset background dominance threshold. (In this example, it's set to 50%). The background detection module will use the scene background suppression factor. Setting it to 0 indicates that the current field of view is not completely dominated by a monotonous background, and no additional environmental suppression penalty needs to be applied in the target calculation.

[0160] Target feature calculation and index output. The target estimation module calls multiple sets of basic computer vision operators on the 256×256 working resolution image to perform local feature extraction and statistically analyze the average gradient and gray-level variance response within a multi-scale window. Connected components of the working resolution image are extracted through local background gray-level difference comparison. The target estimation module calculates the morphological and intensity feature parameters of each connected component, including area, local contrast, aspect ratio of the bounding rectangle, fill rate, bounding box width and height, edge density, and shape constraint matching results. Due to the dense building edges and artificial structural corners in urban built-up areas, after screening, the target estimation module removes invalid noise points based on morphological and intensity feature parameters, retaining 45 valid candidate targets, i.e., the number of valid candidate targets. Among these connected components, the number of components with a bounding rectangle aspect ratio greater than 2.5 and a fill rate satisfying the characteristics of artificial buildings is 20. The target estimation module counts the number of structured targets based on morphological and strength characteristic parameters. .

[0161] The target estimation module integrates the basic feature scores of the above operators to calculate the basic feature score. Because the number of candidate targets or structured targets did not reach the preset judgment index, the condition accumulation mechanism for the dense small target reward item was not activated, and the dense small target reward item... =0; Current cloud interference flag =0 and scene background suppression factor =0, cloud interference suppression term With effective background suppression deduction value All values ​​are assigned the value 0. Substitute these values ​​into the target calculation formula: The target estimation module outputs a target index of 180.

[0162] Contextual multimodal semantic assessment. The contextual assessment module obtains real-time satellite telemetry parameters corresponding to the current remote sensing image, and uses these parameters to transform the four corner points of the current remote sensing image to a geographic coordinate system to construct the current field of view polygon. It also reads a pre-set region of interest configuration file containing the coordinates of known key observation areas. The overlap ratio between the current field of view polygon and the specified region of interest is calculated. The score is 0.9. This indicates the score for the intensity of task attention within a specified area of ​​interest. The configuration is set to 200. The context evaluation module calculates the task region matching score: That is, the task area matching is divided into 180.

[0163] The context evaluation module queries the onboard storage medium to obtain historical reference images of the same region. For spatial pixel-level alignment of two images, those skilled in the art can use a homography matrix transformation algorithm based on ORB feature point matching to perform image registration. This registration and resampling process is a well-known technique in the field and will not be elaborated here. After registration, grayscale difference comparison is performed to extract the rate of change of changed pixels as a percentage of total pixels. It is 15.7%. The change score is calculated as follows: That is, the changes are divided into 40.

[0164] The context evaluation module parses the urgency attribute of real-time satellite scheduling instructions, extracts the priority flag bit from the mission control message header that is annotated with the imaging mission instructions or synchronously carried by the onboard mission queue, determines whether the observation mission belongs to the routine patrol level, and maps and extracts the urgency attribute coefficient. The score is 0.56. Calculate the task time score: The rounded-down task timeframe is 143.

[0165] Linear weighted and gated scoring. The scoring gating module receives the above-mentioned quantization parameters and obtains the weight coefficients of each item from the system configuration file. ). (e.g., combined) Figure 4As shown, the multi-dimensional sub-indicator feature responses under different scenarios, including the aforementioned urban building area scenario, are displayed in a coordinate system with index scores (0-255). The scoring gating module generates the initial score according to the calculation formula. :

[0166] ;

[0167] .

[0168] like Figure 7 As shown, the gating restrictions and comprehensive scoring are confirmed. The output quality gating coefficient is then considered. The scoring gating module calculates the basic score: The scoring gating module performs weak task area interference flag determination. The structured target ratio is... .current A score of 0 indicates that the task region matching score of 180 is greater than or equal to the system's preset higher-level task region threshold of 180, meaning the current image is within the core task region and no interference risk flag has been triggered. The base score therefore skips any hard-truncation extreme value triggering conditions, and the scoring gating module will synthesize the scores. The assigned base score is 66.0.

[0169] Tiered physical resource scheduling and execution. Combined with... Figure 5 The diagram showing the comprehensive score and hierarchical scheduling results is based on a coordinate system of comprehensive score (0-100). The scheduling module receives a comprehensive score of 66.0 and performs logical judgment based on the preset high-value threshold (80) and medium-value threshold (60). Since... The system categorizes it into a medium-value level. Based on the dynamic scheduling priority calculation formula, when the image data enters the candidate buffer or lightweight processing queue, its cumulative queuing time is calculated. Seconds, preset wait time gain coefficient Waiting time compensation cap When the priority is set to 20, the dynamic scheduling priority is calculated as follows: .

[0170] Hierarchical physical resource scheduling is implemented. After calculating a dynamic scheduling priority of 69.0, the scheduling module sorts the data in descending order based on this priority, generating a real-time scheduling execution sequence. For the already segmented mid-value image data, the scheduling module schedules its task execution path to a low-power processing unit, controlling the mid-value image data to enter a lightweight classification model for coarse-grained classification and screening. Based on instruction set addressing, the scheduling module does not activate the NVIDIA Jetson graphics processor with a power consumption exceeding 15W, but instead guides the image data to a lightweight classification model running on an ARM processing unit with a power consumption of less than 2W. The lightweight classification model performs forward inference by calling lightweight networks with depthwise separable convolutional architectures, such as MobileNetV2, to confirm the existence of corresponding structures within the region with low hardware resource overhead. It then generates coarse-grained labels, performs conventional JPEG compression and encapsulation, and stores the data in solid-state storage for later transmission.

[0171] Combination Figure 4 and Figure 5 The comparison of the three sets of data presented and the specific execution logic of the above embodiments demonstrate that the method provided by this invention exhibits clear resource protection and allocation efficiency in a spaceborne environment with extremely limited resources.

[0172] For images located in open sea areas (low value) (i.e.) Figure 8 The scene depicts an open sea area, with the main subject being the deep sea background and only a few small targets. The system extracted a target-specific index of only 40, a task region matching score of 50, a variation score of 15, an image quality score of 120, and a task timeliness score of 100. After linear weighted calculation, the weighted sum is 53.0. Mapped to a percentage, the system directly calculates a comprehensive score of 20.78, retaining one decimal place, which is an extremely low value range of 20.8. The scheduling module directly triggers a discard action or high-rate data compression based on the low score. This mechanism utilizes low-power processing units (such as ARM cores) to perform invalid data judgment, which previously required idle processing in high-performance loads. This effectively prevents image data contaminated by large areas of cloud cover or completely underwater backgrounds from consuming valuable GPU memory or dedicated AI acceleration chip computing cycles.

[0173] For urban built-up areas (medium value) images (i.e.) Figure 7 In the urban building area scenario described in the above embodiments, the system evaluates the target area using structured target proportion constraints and task area matching score (180), locking its base score at 66.0. This evaluation mode avoids blindly classifying conventional building backgrounds as high-value core targets. The scheduling module triggers lightweight processing logic based on this value score. During processing, the dynamic scheduling priority mechanism relies on the following waiting time compensation item: It compensates for the time-consuming loss caused by image queuing, allowing this part of the data to be roughly screened in low-power hardware without being missed due to long-term backlog.

[0174] Combined with appendix Figure 6 This figure shows an example of a core port area (high-value) image captured in a real-world on-orbit application. The core port area scene input is characterized by a clear, cloudless near-shore coastline containing densely packed ships (high-value structured targets). Due to the high levels of edge density, MSER stable extreme region response, and target spatiotemporal context information in this scene, the sub-indicators exhibit excellent performance: targetability (220), task region matching score (255), variation score (79), image quality score (240), and task timeliness score (200).

[0175] These feature parameters, after weighted fusion, are as shown in the attached figure. Figure 6 As shown by the visual identifier directly output by the system in the upper left corner of the screen, its overall score jumped to 82.8. Since 82.8 far exceeds the high-value threshold of 80, the system added a HIGH-level (high-value) scheduling label to it. (Regarding the attached...) Figure 6 For high-value images tagged with "HIGH" as shown, the scheduling module acquires system physical state parameters in real time and, within hardware thermal control limits, precisely routes this portion of image data containing high-density semantic entities to the high-performance computing platform. Through the combined effect of the aforementioned lightweight pre-detection and hard truncation constraints, this invention ensures that every watt of onboard energy is channeled into remote sensing image processing tasks that are physically effective, have stable target features, and are timely, thereby achieving dynamic balance of multi-level resources on the onboard heterogeneous computing platform.

Claims

1. A lightweight pre-detection and hierarchical scheduling method for remote sensing images for onboard AI, characterized in that, Includes the following steps: The working resolution image is obtained using the original high-resolution remote sensing image, and image quality evaluation parameters and background interference evaluation parameters are extracted. Based on the image quality evaluation parameters, the image quality score and quality gating coefficient are calculated. Based on the working resolution image, morphological and intensity feature parameters are obtained, and target indicators are calculated and output by combining background interference evaluation parameters and image quality evaluation parameters. Read the configuration file of the region of interest to generate a task region matching score, obtain historical reference images of the same region to perform comparison and generate a change score, and parse the urgency attribute of the satellite scheduling instructions to generate a task timeliness score; The target index, image quality score, task region matching score, change score and task timeliness score are linearly weighted to generate an initial score. The initial score is then attenuated using a quality gating coefficient to generate a base score. The weak task region interference suppression logic is then activated for the base score, triggering a hard truncation mechanism for the upper limit of the score, and outputting a comprehensive score after pressure reduction. Based on the comprehensive score, remote sensing images are divided into multiple value levels, and then hierarchical physical resource scheduling is performed.

2. The lightweight pre-detection and hierarchical scheduling method for remote sensing images oriented towards on-board AI according to claim 1, characterized in that, To obtain a working resolution image from the original high-resolution remote sensing image, a downsampling module is used to perform a downsampling operation on the original high-resolution remote sensing image. Image quality evaluation parameters include cloud cover evaluation parameters, and the quality detection module generates cloud interference flags based on these parameters. Before calculating the image quality score and quality gating coefficient, the quality detection module also generates a cloud interference mask based on the working resolution image, specifically including: Convert the working resolution image to the HSV color space and extract the pixel matrix of the luminance channel, saturation channel and hue channel; In the brightness pixel matrix, pixels with brightness values ​​higher than a preset brightness threshold are selected as pixels that meet the high brightness condition. In the saturation pixel matrix, pixels with saturation values ​​lower than a preset saturation threshold are selected as pixels that meet the low saturation condition. Pixels that meet both the high brightness and low saturation conditions are extracted to form a candidate continuous region. Obtain the gradient value and gray value of each pixel within the candidate continuous region, and then calculate the average gradient value and gray variance of the candidate continuous region; When the average gradient value is lower than the preset gradient threshold, the corresponding candidate continuous region is determined to meet the low gradient condition, and the texture smoothness inside the candidate continuous region is verified by combining the gray-scale variance to determine whether the smoothness condition is met. A continuous region that simultaneously satisfies the conditions of high brightness, low saturation, low gradient, and smoothness is marked as a cloud interference region, and a cloud interference mask is generated based on the cloud interference region.

3. The lightweight pre-detection and hierarchical scheduling method for remote sensing images oriented towards on-board AI according to claim 2, characterized in that, The cloud cover evaluation parameters include the proportion of large-area cloud coverage and the overall cloud coverage proportion. The specific steps involved in generating cloud cover evaluation parameters using a quality detection module based on a cloud interference mask, generating cloud interference indicators based on these parameters, and then calculating the image quality score and quality gating coefficient include: The quality detection module determines the overall cloud coverage ratio based on the ratio of the total number of pixels marked as cloud interference areas in the cloud interference mask to the total number of pixels in the working resolution image. Perform connected component analysis on the cloud interference area. After filtering out fragmented cloud patches with an area smaller than the area threshold, determine the large-area cloud coverage ratio based on the ratio of the total pixel area of ​​the remaining connected components to the total number of pixels in the working resolution image. The quality detection module combines the global gradient mean of the working resolution image with the gradient reference value read from the system configuration file to calculate the normalized sharpness parameter. It then subtracts a penalty term proportional to the overall cloud coverage ratio from the normalized sharpness parameter to output the image quality score, and truncates the result to zero when the calculation result is below zero. When the proportion of large-area cloud coverage exceeds the preset cloud occlusion tolerance threshold, the quality gating coefficient undergoes nonlinear attenuation and weight reduction calculation as the proportion of large-area cloud coverage increases, and the output quality gating coefficient has a value range between zero and one. When the proportion of cloud coverage over a large area exceeds the preset cloud occlusion tolerance threshold, a cloud interference flag is generated in a triggered state.

4. A lightweight pre-detection and hierarchical scheduling method for remote sensing images for onboard AI as described in claim 1, characterized in that, The specific steps for generating background interference evaluation parameters by processing working resolution images using the background judgment module include: The background judgment module extracts the set of pixels in the preset vegetation tone range or preset water tone range by dividing the color space by hard threshold, and extracts the set of pixels in the preset dark background range by brightness judgment. For pixels in the color tone range that meet the vegetation characteristics and whose grayscale variance is lower than the smooth texture threshold, the summation and comparison are performed to calculate the vegetation coverage ratio. For pixels in the color tone range that meet the preset water body color tone range or brightness range that meet the dark background characteristics and whose grayscale variance is lower than the smooth texture threshold, the summation and comparison are performed to calculate the water body dark area coverage ratio. When the maximum value of the vegetation coverage ratio and the water body dark area coverage ratio exceeds the preset background dominant threshold, the difference between the maximum value and the background dominant threshold is calculated as the excess ratio amplitude value, and the background interference evaluation parameter is generated using the excess ratio amplitude value. When neither the vegetation coverage ratio nor the water body dark area coverage ratio exceeds the background dominant threshold, the background interference evaluation parameter is set to zero.

5. A lightweight pre-detection and hierarchical scheduling method for remote sensing images for onboard AI according to claim 2, characterized in that, It also includes using a target estimation module to perform feature extraction on the working resolution image to generate various local feature detection scores, specifically including: Multiple sets of basic computer vision operators are invoked to perform local feature extraction on the working resolution image; Based on the pixel scale range of typical targets in the working resolution image, the target coverage requirements, and the computing power budget of the low-power processing unit, the number of windows to be extracted is jointly determined. Local windows whose combined ranking of average gradient and gray-level variance response is within the range of the number of windows to be extracted are selected. The average response of the selected local windows is calculated and mapped to generate a gradient target detection score. The edge map of the working resolution image is extracted by combining the edge extraction operator in the basic computer vision operator to generate candidate connected regions. The number of edge pixels in the candidate connected regions is counted. The edge density is determined according to the ratio of the number of edge pixels to the total number of pixels in the candidate connected regions. The local intensity features of the candidate connected regions are extracted. The edge candidate detection score is generated by combining the edge density and the local intensity features. Find regions that maintain stable shape under multiple grayscale thresholds, remove noise points with an area lower than the preset lower area threshold or background blocks with an area higher than the preset upper area threshold, as well as regions with an aspect ratio exceeding the preset aspect ratio abnormal threshold. Also, exclude regions when the proportion of overlapping pixels with the cloud interference mask is higher than the preset cloud mask overlap threshold. Integrate the remaining effective regions that maintain stable shape to generate a stable bright and dark region detection score. Identify local corner points in the working resolution image where the grayscale change exceeds a preset neighborhood contrast threshold, count the density of local corner points within the effective region, and map them to generate a local structure point detection score. The visual salient region response is calculated based on the color and brightness differences of the working resolution image. The areas covered by cloud interference masks are subtracted, and the visual salient region response of the uncovered areas is normalized to generate a salient feature analysis score.

6. A lightweight pre-detection and hierarchical scheduling method for remote sensing images for onboard AI according to claim 5, characterized in that, It also includes using a target estimation module to statistically determine the number of valid candidate targets and the number of structured targets based on morphological and intensity feature parameters, specifically including: The morphological and intensity characteristic parameters include area, local contrast, aspect ratio of the bounding rectangle, fill rate, width and height of the bounding box, edge density, and shape constraint matching results. The target estimation module removes invalid noise points based on morphological and intensity feature parameters, and calculates and generates candidate target analysis scores based on the feature parameters of the retained candidate connected regions, and counts the number of valid candidate targets. Based on morphological and intensity feature parameters, geometric structure matching analysis is performed on the retained candidate connected regions to count the number of structured targets with preset geometric structure features.

7. A lightweight pre-detection and hierarchical scheduling method for remote sensing images for onboard AI as described in claim 6, characterized in that, The specific steps for calculating and outputting target indicators by combining background interference evaluation parameters and cloud cover evaluation parameters include: The target estimation module multiplies the gradient target detection score, edge candidate detection score, stable bright and dark region detection score, candidate target analysis score, local structure point detection score, and saliency feature analysis score by their respective weight coefficients and sums them to calculate the basic feature score. When the cloud interference flag is triggered and there is large-area cloud obstruction interference, and the number of effective candidate targets is insufficient to reach the lower limit threshold of the number of effective candidates, the target estimation module will assign the preset penalty score as the cloud interference suppression term. When the background interference evaluation parameter is greater than zero and the number of effective candidate targets is insufficient to reach the lower limit threshold of the number of effective candidates, the target estimation module sets the effective background suppression deduction value to the value of the background interference evaluation parameter. The basic feature score is supplemented with a dense small target reward, and the cloud interference suppression term and effective background suppression deduction value are subtracted. The maximum and minimum value functions are used to limit the calculation results to a preset effective value range, and the target index is output.

8. A lightweight pre-detection and hierarchical scheduling method for remote sensing images for onboard AI according to claim 6, characterized in that, The process of activating the weak task area interference suppression logic includes: The proportion of structured targets is calculated based on the ratio between the number of structured targets and the number of valid candidate targets. When the number of valid candidate targets is higher than the preset interference risk assessment index, the proportion of structured targets is lower than the preset lower limit threshold of the structured proportion, the cloud interference flag is in the triggered state or the overall cloud coverage ratio reaches the preset fragmented cloud interference threshold, and the task area matching score is in a weak task-related area, the interference risk flag is triggered.

9. A lightweight pre-detection and hierarchical scheduling method for remote sensing images oriented towards on-board AI, as described in claim 1, is characterized in that... The specific steps of classifying remote sensing images into multiple value levels based on a comprehensive score using the scheduling module include: The scheduling module reads the computing load rate, available video memory, on-chip storage or external storage margin, current chip junction temperature and real-time power consumption of each hardware computing node, and classifies the value level into high value level, medium value level and low value level. When the comprehensive score is greater than or equal to the high value threshold, the remote sensing image is classified as a high value level. When the comprehensive score is greater than or equal to the medium value threshold and less than the high value threshold, the remote sensing image is classified as medium value. When the overall score is less than the medium value threshold, the remote sensing image is classified as low value. For image data that enters the fidelity buffer and waits to be processed, the accumulated queuing time is multiplied by a preset waiting time gain coefficient to generate a calculation result. The minimum value between the calculation result and the waiting time compensation limit is taken to generate a waiting time compensation term. The dynamic scheduling priority value is calculated by adding the waiting time compensation term to the comprehensive score. The execution sequence of the scheduling is generated by arranging the dynamic scheduling priority values ​​in descending order.

10. A lightweight pre-detection and hierarchical scheduling method for remote sensing images oriented towards on-board AI, as described in claim 9, is characterized in that... The process of performing hierarchical physical resource scheduling includes: Image data of high value level is selected sequentially according to the execution sequence. Under the condition that the computing power load rate is lower than the full load threshold and the available video memory, on-chip storage margin or external storage margin can meet the inference requirements, an early warning mark signal is generated, the image data of high value level is guided to the high-performance processing payload on the satellite, and the storage physical address pointer is sent. The task execution path of image data with medium value level is scheduled to low power processing unit, and the image data with medium value level is controlled to enter the lightweight classification model for coarse classification and coarse screening, or the image data with medium value level is placed in the candidate buffer for queuing. When the available video memory, on-chip storage balance, or external storage balance is detected to be lower than the preset capacity warning line, the image data with low value level is directly discarded to release storage space.

Images