A feature camouflage method and device based on light and shadow projection in a star cluster game scenario

Abstract

The application discloses a feature camouflage method and device based on light and shadow projection in a star cluster game scene, and relates to the fields of artificial intelligence and machine vision. The method comprises the following steps: acquiring key visual features of a spacecraft of a home side, and constructing a digital target model; constructing a simulation model of an external observation side recognition system; generating a visual disturbance mode; based on the above model and the disturbance mode, performing static optimal feature confusion processing based on a covariance matrix adaptive evolution strategy, to obtain optimal static light and shadow disturbance baseline parameters; adjusting the optimal static light and shadow disturbance baseline parameters through a dynamic adaptive camouflage strategy based on multi-agent reinforcement learning, to obtain an optimal dynamic adaptive cooperative camouflage strategy; and inputting the strategy into a cooperative physical projection execution link, to be converted into a real light and shadow pattern. The application realizes the technical effects that a star cluster system can obtain an optimal adaptive cooperative camouflage strategy through dynamic adjustment, and a real light and shadow pattern based on the camouflage strategy.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and machine vision, and in particular to a feature camouflage method and device based on light and shadow projection in a star cluster game scenario. Background Technology

[0002] With the development of space technology, constellation systems (i.e., "star clusters") composed of a large number of satellites have become key infrastructures in fields such as communication, navigation, and remote sensing. At the same time, with the networked deployment of these systems, the competition in the star cluster game scenario is becoming increasingly fierce. As a key node of the constellation, the reliability of the on-orbit mission of a single spacecraft is directly related to the stable operation of the entire system.

[0003] Meanwhile, visual recognition technology based on deep learning, especially deep convolutional neural networks (CNNs), has been widely applied in the field of target detection and recognition. External observation systems can utilize this technology to process spacecraft image data acquired by optical or other electromagnetic waveband sensors. These algorithms can autonomously learn and extract abstract features from high-dimensional data, without relying on pre-defined, physical model-based feature engineering, to perform high-precision classification and recognition of spacecraft's three-dimensional geometric contours, optical properties of surface materials (such as the bidirectional reflection distribution function, BRDF), and attitude. This technology is highly robust to environmental changes (such as illumination and viewing angle), thus posing a significant technical challenge to existing spacecraft visual feature adjustment technologies that primarily rely on fixed physical properties.

[0004] Currently, visual feature adjustment technology for spacecraft primarily employs passive strategies, such as applying low-reflectivity surface coatings to reduce optical observability. However, such passive technologies have inherent technical limitations when dealing with the aforementioned intelligent recognition systems. The physical or electromagnetic properties of passive adjustment are fixed after the spacecraft is manufactured, and its preset, static characteristics cannot adapt to dynamically changing interactive environments. In actual operation, the relative position of the observer, the spacecraft's own attitude, and space illumination conditions (such as the solar radiation vector) are constantly changing, which can cause the preset optimal adjustment state to fail. Furthermore, the technical principle of passive adjustment is relatively fixed, and the characteristic signals it generates are predictable. The observer can accumulate long-term, multi-angle observation data to establish a detailed database of the spacecraft's characteristic signals and train its recognition model accordingly, thereby significantly reducing or even eliminating the effectiveness of passive adjustment.

[0005] In summary, existing passive feature adjustment technologies, due to their inherent static and predictable nature, are insufficient to meet the technical requirements for effective visual feature adjustment of spacecraft in modern intelligent space interaction environments. Therefore, there is an urgent need to develop a new technology capable of actively generating adjustment strategies and dynamically adjusting according to the real-time environment to solve the aforementioned technical problems. Summary of the Invention

[0006] The purpose of this application is to provide a feature camouflage method and device based on light and shadow projection in a star cluster game scenario, which can realize dynamic adjustment of camouflage strategy.

[0007] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a feature camouflage method based on light and shadow projection in a star cluster game scenario, comprising the following steps: The key visual features of one's own spacecraft are acquired, and digital modeling is performed based on the key visual features to obtain a digital target model; the key visual features include: three-dimensional geometric contour, light reflection distribution function of surface material, and albedo under a specific electromagnetic band.

[0008] Construct a simulation model of one or more external observer recognition systems. The simulation model is regarded as a black box function that can only be queried by input and return evaluation results. The input of the black box function is an image of a spacecraft to be identified, and the output of the black box function is the classification result of the spacecraft image and the corresponding confidence score.

[0009] Based on the distributed light and shadow projection units of multiple spacecraft within the constellation, a visual perturbation pattern is generated; the visual perturbation pattern includes: a stellar shadow perturbation model and a specular reflection glare perturbation model; the stellar shadow perturbation model is used to simulate complex lighting phenomena in the deep space environment; the specular reflection glare perturbation model is used to simulate complex lighting phenomena in the near-Earth orbit environment.

[0010] Based on the digital target model, the simulation model, and the visual perturbation mode, static optimal feature confusion processing based on the covariance matrix adaptive evolution strategy is performed to obtain the optimal static light and shadow perturbation baseline parameters.

[0011] By adjusting the optimal static light and shadow perturbation baseline parameters using a dynamic adaptive camouflage strategy based on multi-agent reinforcement learning, the optimal dynamic adaptive cooperative camouflage strategy is obtained.

[0012] The optimal dynamic adaptive cooperative camouflage strategy is input into the cooperative physical projection execution stage, and transformed into a real light and shadow pattern in physical space.

[0013] Optionally, the stellar shadow perturbation model simulates the complex umbra and penumbra effects formed by the occlusion of main-sequence stars, neighboring stars, or other friendly spacecraft in a star cluster by co-projecting a combination of shadows of specific shapes and intensities onto the surface of the target spacecraft, thereby changing the local illumination intensity of the target surface.

[0014] The expression for the stellar shadow perturbation model is as follows: ; in, The original target image; The image after shadows have been applied; For pixels; For a single perturbation mask function, ; The number of cooperating spacecraft in the constellation; It is the first Shadow parameter vectors of a cooperating spacecraft; The coordinates of the center of the ellipse; For major and minor axes; ( ) represents the rotation angle; ) represents the intensity of the umbra; () represents the penumbra softness.

[0015] Optionally, the specular reflection glare disturbance model simulates specular reflection or coordinated directional light interference caused by solar panels, optical windows, or metal components receiving direct starlight at a specific angle by coordinating the projection of high-intensity light spots onto the surface of the target spacecraft.

[0016] The expression for the specular reflection glare perturbation model is as follows: ; in, The original target image; The image after applying glare; For pixels; For a single spot function, ; The number of cooperating spacecraft in the constellation; The intensity superposition function follows an exponential radial decay law; For pixels Euclidean distance to the center of the light spot; parameter vector For the first Glare parameter vectors of the cooperating spacecraft; The coordinates of the center of the light spot; The radius of action; Central strength; This represents the attenuation rate.

[0017] Optionally, based on the digital target model, the simulation model, and the visual perturbation pattern, static optimal feature obfuscation processing based on a covariance matrix adaptive evolution strategy is performed to obtain the optimal static light and shadow perturbation baseline parameters, specifically including the following steps: Based on the digital target model, the simulation model, and the visual perturbation mode, for a preset, relatively static interactive scenario, the light and shadow perturbation parameters of all cooperating spacecraft within the constellation are obtained; the light and shadow perturbation parameters include shadow parameters and glare parameters.

[0018] The light and shadow perturbation parameters are combined into a joint parameter vector of a preset dimension to form a joint parameter space of a preset dimension.

[0019] A covariance matrix adaptive evolution strategy is used to perform black-box optimization on the joint parameter space of the preset dimension to obtain the optimal static light and shadow perturbation baseline parameters; the objective function of the covariance matrix adaptive evolution strategy is defined as maximizing the classification error rate of the external recognition system or minimizing its confidence in correct classification.

[0020] Optionally, a covariance matrix adaptive evolution strategy is used to perform black-box optimization on the joint parameter space of the preset dimension to obtain the optimal static lighting perturbation baseline parameters, specifically including the following steps: Define a black-box fitness function; the input of the black-box fitness function is a set of joint parameter vectors of preset dimensions, and the output is a scalar evaluation value; the black-box fitness function generates a cooperative light and shadow pattern defined by the joint parameter vectors of preset dimensions through internal simulation, and submits the cooperative light and shadow pattern to the simulation model for evaluation to obtain a scalar evaluation value; the scalar evaluation value quantifies the degree of confusion caused by the cooperative light and shadow pattern to the recognition system.

[0021] Based on the black-box fitness function, the optimal solution is found in the joint parameter space of the preset dimension through the covariance matrix adaptive evolution strategy, and the optimal static light and shadow perturbation baseline parameters are obtained.

[0022] Optionally, the optimal static lighting perturbation baseline parameters are adjusted using a dynamic adaptive camouflage strategy based on multi-agent reinforcement learning to obtain the optimal dynamic adaptive cooperative camouflage strategy, specifically including the following steps: It can acquire real-time dynamic information from the observer and the coordinated state within the constellation.

[0023] Based on the dynamic information of the observer and the cooperative state within the constellation, the optimal static light and shadow perturbation baseline parameters are continuously adjusted at the millisecond level through multi-agent Markov decision-making to generate the optimal dynamic adaptive cooperative camouflage strategy.

[0024] Optionally, the multi-agent Markov decision-making process specifically includes the following steps: Each agent's state input vector is input to the policy network of the agent cluster, and a joint action vector is output. The state input vector includes the agent's current optimal static light and shadow perturbation baseline parameters, attitude information, state summary of other cooperating spacecraft in the cluster, and external sensing information such as relative distance, velocity vector, attitude angle, and estimated sensor orientation of the observing spacecraft obtained through lidar, passive optics, or radio sensors. The joint action vector contains the adjustment amount for the optimal static light and shadow perturbation baseline parameters of each cooperating spacecraft in the cluster.

[0025] A reward function is established based on the state input vector and the joint action vector; the positive reward term of the reward function includes the instantaneous decrease in the observer's confidence level and the duration of maintaining the observer in a state below the preset confidence level; the negative penalty term of the reward function includes the amplitude and frequency of the changes in the light and shadow pattern of the entire star cluster and the event of the target being successfully identified.

[0026] Based on the reward function, the policy network of the agent cluster is fine-tuned online in a lightweight manner to obtain the optimal dynamic adaptive cooperative camouflage strategy.

[0027] In a second aspect, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the feature camouflage method based on light and shadow projection in the star cluster game scenario described above.

[0028] Thirdly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the feature camouflage method based on light and shadow projection in the star cluster game scenario described above.

[0029] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the feature camouflage method based on light and shadow projection in the star cluster game scenario described above.

[0030] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a feature camouflage method and apparatus based on light and shadow projection in a star cluster game scenario. It obtains optimal static light and shadow perturbation baseline parameters by performing static optimal feature obfuscation processing based on a covariance matrix adaptive evolution strategy, using the digitized target model, the simulation model, and the visual perturbation mode. Then, it adjusts the optimal static light and shadow perturbation baseline parameters using a dynamic adaptive camouflage strategy based on multi-agent reinforcement learning to obtain the optimal dynamic adaptive cooperative camouflage strategy. This solves the technical problem that traditional passive camouflage strategies cannot adapt to dynamic adversarial environments, and enables star cluster systems to obtain the optimal adaptive cooperative camouflage strategy through dynamic adjustment. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is an application environment diagram of a feature camouflage method based on light and shadow projection in a star cluster game scenario according to an embodiment of this application; Figure 2 A flowchart illustrating a feature camouflage method based on light and shadow projection in a star cluster game scenario provided in an embodiment of this application; Figure 3 This is a schematic diagram of the distribution of key visual features of a Chinese spacecraft provided in an embodiment of this application; Figure 4 A schematic diagram of a dual-branch decoding network structure for visual perturbation modeling provided in an embodiment of this application; Figure 5 This is a schematic diagram of the visual perturbation pattern calculation process provided in an embodiment of this application; Figure 6 A schematic diagram illustrating the generation of a static optimal feature obfuscation scheme based on CMA-ES, provided in an embodiment of this application; Figure 7 A schematic diagram of a dynamic adaptive adjustment framework based on multi-agent reinforcement learning provided in an embodiment of this application; Figure 8 A schematic diagram illustrating the implementation structure and processing of a collaborative physical projection execution stage provided in an embodiment of this application; Figure 9 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] The feature camouflage method based on light and shadow projection in the star cluster game scenario provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on other servers. Terminal 102 can send key visual features of its own spacecraft to server 104. These key visual features include three-dimensional geometric contours, the light reflection distribution function of the surface material, and the albedo under a specific electromagnetic band. Server 104 receives the key visual features of its own spacecraft and performs digital modeling based on these features to obtain a digital target model. It then constructs simulation models of one or more external observer recognition systems. The simulation model is considered a black-box function that can only be queried and return evaluation results. The input of the black-box function is an image of the spacecraft to be identified, and the output is the classification result of the spacecraft image and the corresponding confidence score. Based on multiple spacecraft distributed light and shadow projection units within the constellation, a visual perturbation pattern is generated. The visual perturbation models include a stellar shadow perturbation model and a specular reflection glare perturbation model. The stellar shadow perturbation model is used to simulate complex lighting phenomena in deep space. The specular reflection glare perturbation model is used to simulate complex lighting phenomena in near-Earth orbit. Based on the digitized target model, the simulation model, and the visual perturbation models, static optimal feature obfuscation processing based on a covariance matrix adaptive evolution strategy is performed to obtain the optimal static light and shadow perturbation baseline parameters. The optimal static light and shadow perturbation baseline parameters are adjusted using a dynamic adaptive camouflage strategy based on multi-agent reinforcement learning to obtain the optimal dynamic adaptive cooperative camouflage strategy. The optimal dynamic adaptive cooperative camouflage strategy is input into the cooperative physical projection execution stage to transform it into a real light and shadow pattern in physical space.Furthermore, in some embodiments, the feature camouflage method based on light and shadow projection in the star cluster game scenario can also be implemented independently by the server 104 or the terminal 102. For example, the terminal 102 can directly acquire the key visual features of its own spacecraft, and perform digital modeling based on the key visual features to obtain a digital target model. The key visual features include three-dimensional geometric contours, the illumination reflection distribution function of the surface material, and the albedo under a specific electromagnetic band. A simulation model of one or more external observer identification systems is constructed. The simulation model is regarded as a black box function that can only query by input and return evaluation results. The input of the black box function is an image of the spacecraft to be identified, and the output is the classification result of the spacecraft image and the corresponding confidence score. Based on multiple spacecraft distributed light and shadow projection units within the star cluster, A visual perturbation pattern is generated, comprising a stellar shadow perturbation model and a specular reflection glare perturbation model. The stellar shadow perturbation model is used to simulate complex lighting phenomena in deep space, while the specular reflection glare perturbation model is used to simulate complex lighting phenomena in near-Earth orbit. Based on the digitized target model, the simulation model, and the visual perturbation pattern, static optimal feature obfuscation processing based on a covariance matrix adaptive evolution strategy is performed to obtain the optimal static light and shadow perturbation baseline parameters. The optimal static light and shadow perturbation baseline parameters are adjusted using a dynamic adaptive camouflage strategy based on multi-agent reinforcement learning to obtain the optimal dynamic adaptive cooperative camouflage strategy. The optimal dynamic adaptive cooperative camouflage strategy is input into the cooperative physical projection execution stage to transform it into a real light and shadow pattern in physical space.Alternatively, server 104 can retrieve key visual features of its own spacecraft from the data storage system, and perform digital modeling based on these key visual features to obtain a digital target model. The key visual features include three-dimensional geometric contours, the light reflection distribution function of the surface material, and the albedo under a specific electromagnetic band. A simulation model of one or more external observation identification systems is constructed. This simulation model is considered a black-box function that can only be queried by input and return evaluation results. The input of the black-box function is an image of the spacecraft to be identified, and the output is the classification result of the spacecraft image and the corresponding confidence score. Based on multiple spacecraft distributed light and shadow projection units within the constellation, a visual perturbation pattern is generated. The visual perturbation pattern includes: stellar shadow... The model includes a stellar shadow perturbation model and a specular reflection glare perturbation model. The stellar shadow perturbation model is used to simulate complex lighting phenomena in deep space. The specular reflection glare perturbation model is used to simulate complex lighting phenomena in near-Earth orbit. Based on the digitized target model, the simulation model, and the visual perturbation pattern, static optimal feature obfuscation processing based on a covariance matrix adaptive evolution strategy is performed to obtain the optimal static light and shadow perturbation baseline parameters. The optimal static light and shadow perturbation baseline parameters are adjusted using a dynamic adaptive camouflage strategy based on multi-agent reinforcement learning to obtain the optimal dynamic adaptive cooperative camouflage strategy. The optimal dynamic adaptive cooperative camouflage strategy is input into the cooperative physical projection execution stage to transform it into a real light and shadow pattern in physical space.

[0036] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, and spacecraft; the spacecraft can be a satellite; the server 104 can be implemented by an independent server or a server cluster composed of multiple servers, or it can be a cloud server.

[0037] In one exemplary embodiment, such as Figure 2 As shown, a feature camouflage method based on light and shadow projection is provided for a star cluster game scenario. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps S201 to S206.

[0038] S201: Obtain key visual features of the spacecraft, and perform digital modeling based on these key visual features to obtain a digital target model; the key visual features include, but are not limited to, three-dimensional geometric contours, the light reflectance distribution function (BRDF) of surface materials, and albedo in specific electromagnetic bands (such as visible light and infrared); based on the digital target model, obtain a schematic diagram of the distribution of key visual features of the spacecraft, such as... Figure 3 As shown.

[0039] S202: Construct one or more simulation models of an external observer identification system. Each simulation model is strictly defined as a black-box function whose internal structural information is unknown, capable only of receiving input queries and returning evaluation results. In subsequent steps, this model will serve as a computable evaluation function, providing a quantitative basis for optimizing camouflage strategies. This is used to quantitatively evaluate the effectiveness of feature adjustment strategies in subsequent modules. The simulation model simulates the operation of a real-world observer identification system. The input to the simulation model is an image of a spacecraft to be identified, and the output is the classification result of the spacecraft image and the corresponding confidence score. This model can be a white-box model with known network topology and parameters for in-depth adversarial analysis; or it can be a black-box model that can only be used to input queries and observe its classification results and confidence scores to simulate more realistic adversarial scenarios.

[0040] S203: Based on the distributed light and shadow projection units of multiple spacecraft within the constellation, a visual perturbation mode is generated; the visual perturbation mode includes: a stellar shadow perturbation model and a specular reflection glare perturbation model; the stellar shadow perturbation model is used to simulate complex lighting phenomena in the deep space environment; the specular reflection glare perturbation model is used to simulate complex lighting phenomena in the near-Earth orbit environment.

[0041] S204: Based on the digital target model, the simulation model, and the visual perturbation mode, perform static optimal feature confusion processing based on the covariance matrix adaptive evolution strategy to obtain the optimal static light and shadow perturbation baseline parameters.

[0042] S205: By using a dynamic adaptive camouflage strategy based on multi-agent reinforcement learning, the optimal static light and shadow perturbation baseline parameters are adjusted to obtain the optimal dynamic adaptive cooperative camouflage strategy.

[0043] S206: Input the optimal dynamic adaptive cooperative camouflage strategy into the cooperative physical projection execution stage to transform it into a real light and shadow pattern in physical space.

[0044] By implementing steps S201 to S206 above, this application obtains the optimal static light and shadow perturbation baseline parameters by performing static optimal feature obfuscation processing based on the digital target model, the simulation model, and the visual perturbation mode using a covariance matrix adaptive evolution strategy. Then, by adjusting the optimal static light and shadow perturbation baseline parameters using a dynamic adaptive camouflage strategy based on multi-agent reinforcement learning, the optimal dynamic adaptive cooperative camouflage strategy is obtained. This solves the technical problem that traditional passive camouflage strategies cannot adapt to dynamic adversarial environments, and achieves the technical effect of obtaining the optimal adaptive cooperative camouflage strategy for constellation systems through dynamic adjustment.

[0045] As an optional implementation, in step S203, a visual perturbation pattern is generated based on the distributed light and shadow projection units of multiple spacecraft within the constellation. The visual perturbation pattern includes a stellar shadow perturbation model and a specular reflection glare perturbation model. The stellar shadow perturbation model is used to simulate complex lighting phenomena in deep space; the specular reflection glare perturbation model is used to simulate complex lighting phenomena in near-Earth orbit. Specifically, step S203 is responsible for mathematizing and parameterizing the lighting phenomena in the physical world, making them calculable and optimizable. Step S203 defines a set of visual perturbation patterns collaboratively generated by multiple spacecraft within the constellation through their onboard distributed light and shadow projection units, aiming to simulate complex dynamic lighting phenomena in deep space and near-Earth orbit environments. This is specifically achieved through two parameterizable models: one is the stellar shadow perturbation model, which collaboratively projects one or more combined shadow patterns of specific shapes, intensities, and softness onto the surface of the target spacecraft. Each shadow is precisely controlled by a set of parameters (including center position, major and minor axes, rotation angle, umbra intensity, and penumbra softness). The first method involves the coordinated projection of star clusters to create more complex and larger-scale dynamic shadows than a single projection unit. This simulates the complex umbra and penumbra effects caused by main-sequence stars, nearby celestial bodies, or other friendly spacecraft within the cluster. The second method is the specular reflection glare perturbation model. This model coordinates the projection of one or more high-intensity light spots onto the surface of the target spacecraft. Each spot is precisely controlled by a set of parameters (including center position, radius of influence, center intensity, attenuation rate, etc.). Through multi-point coordinated projection, multi-angle, distributed glare interference can be achieved on the observation system to simulate specular reflection or directional light interference caused by solar panels, optical windows, or metal components receiving direct stellar light at specific angles.

[0046] like Figure 4 The diagram illustrates a two-branch decoding network structure used for visual perturbation modeling. Input image. After multi-scale features are extracted by the encoder (E), they are fed into the image decoder (Di) and the shadow decoder (Ds), respectively. The image decoder introduces noise perturbations into the multi-layer features and reconstructs the output image step-by-step using transport blocks. The shadow decoder then fuses encoded features with cross-layer information and generates a shadow mask through a cross-attention mechanism. This structure enables joint modeling of the original observation image and the shadow perturbation distribution, providing a fundamental characterization for the subsequent parameterized generation of stellar shadows and specular glare.

[0047] The calculation process of the visual perturbation pattern in step S203 is as follows: Figure 5As shown in the flowchart, this method illustrates a collaborative simulation processing method for stellar shadows and specular glare in a target image. The system inputs the original image's shadow parameter set and glare parameter set, then proceeds to two parallel computation branches. The left branch generates a shadow intensity mask based on an elliptic distance function and forms a shadow perturbation image through pixel-by-pixel multiplicative superposition, characterizing the umbra and penumbra effects. The right branch generates multi-source light spot intensities based on a radial attenuation model and performs linear superposition and normalization constraints to obtain the glare enhancement result. After the two branches complete their calculations, the glare is superimposed onto the shadow result in the fusion stage, finally outputting the synthesized image. This structure achieves decoupled modeling and efficient parallel computation of shadow occlusion and specular reflection interference. The specific process is as follows: S2031: The stellar shadow perturbation model simulates the complex umbra and penumbra effects formed by the occlusion of main-sequence stars (such as the Sun), nearby stars, or other friendly spacecraft in a star cluster by coordinating the projection of a combination of shadows of specific shapes and intensities onto the surface of the target spacecraft, thereby changing the local illumination intensity of the target surface.

[0048] The stellar shadow perturbation model described uses a parameterized perturbation mask function. The final composite shadow effect is achieved. In the star cluster A single perturbation mask function generated by a cooperative spacecraft The result of combined effects.

[0049] The expression for the stellar shadow perturbation model is as follows: ; in, The original target image; The image after shadows have been applied; For pixels; It is an intensity mask generated based on the elliptic distance function; For a single perturbation mask function, ; The number of cooperating spacecraft in the constellation; It is the first The shadow parameter vectors of the cooperating spacecraft control the coordinates of the ellipse centers respectively. Major and minor axes Rotation angle ( ), umbra intensity ( ) and penumbra softness ( This model simulates the physical properties of celestial shadows with high fidelity. Through multi-point collaborative projection, it can be combined to form more complex and larger-scale shadow patterns.

[0050] S2032: The specular reflection glare disturbance model simulates specular reflection or coordinated directional light interference caused by solar panels, optical windows or metal components receiving direct starlight at a specific angle by coordinating the projection of high-intensity light spots onto the surface of the target spacecraft.

[0051] The aforementioned specular reflection glare perturbation model uses an intensity superposition function. The final composite glare effect is achieved. In the star cluster A single spot function generated by a cooperative spacecraft The result of linear superposition.

[0052] The expression for the specular reflection glare perturbation model is as follows: ; in, The original target image; The image after applying glare; For pixels; For a single spot function, ; The number of cooperating spacecraft in the constellation; The intensity superposition function follows an exponential radial decay law; For pixels Euclidean distance to the center of the light spot; parameter vector For the first The glare parameter vectors of each cooperating spacecraft control the coordinates of the light spot center. Radius of action Central strength and attenuation rate This model aims to simulate the strong light area generated by material reflection or active light source illumination. Through multi-point collaborative projection, it can achieve multi-angle and distributed glare interference to the observation system.

[0053] S2033: Overlays the combined glare onto the image that has already been shadowed, and finally outputs the composite image. The expression is as follows: .

[0054] As an optional implementation, in step S204, based on the digital target model, the simulation model, and the visual perturbation mode, static optimal feature obfuscation processing based on the covariance matrix adaptive evolution strategy is performed to obtain the optimal static light and shadow perturbation baseline parameters. Specifically, this includes the following process: S2041: Based on the digital target model, the simulation model, and the visual perturbation mode, for a preset, relatively static interactive scenario, obtain the light and shadow perturbation parameters of all cooperating spacecraft within the constellation; the light and shadow perturbation parameters include shadow parameters and glare parameters.

[0055] S2042: Combine the light and shadow perturbation parameters into a joint parameter vector of a preset dimension to form a joint parameter space of a preset dimension.

[0056] S2043: The covariance matrix adaptive evolution strategy is used to perform black-box optimization on the joint parameter space of the preset dimension to obtain the optimal static light and shadow perturbation baseline parameters; the objective function of the covariance matrix adaptive evolution strategy is defined as maximizing the classification error rate of the external recognition system or minimizing its confidence in correct classification.

[0057] Step S2043 above is replaced by steps S301 to S302, as follows: S301: Define a black-box fitness function; the input of the black-box fitness function is a set of joint parameter vectors of preset dimensions, and the output is a scalar evaluation value; the black-box fitness function generates a cooperative light and shadow pattern defined by the joint parameter vectors of preset dimensions through internal simulation, and submits the cooperative light and shadow pattern to the simulation model for evaluation to obtain a scalar evaluation value; the scalar evaluation value quantifies the degree of confusion caused by the cooperative light and shadow pattern to the recognition system.

[0058] S302: Based on the black-box fitness function, the optimal solution is found in the joint parameter space of the preset dimension through the covariance matrix adaptive evolution strategy to obtain the optimal static light and shadow perturbation baseline parameters.

[0059] Specifically, a static optimal feature obfuscation scheme is generated based on the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) to calculate an optimal set of static lighting and shadow perturbation baseline parameters offline for a pre-defined, relatively static interaction scenario (such as a specific orbital rendezvous window or a fixed monitoring angle). First, based on the digitized target model, the simulation model, and the visual perturbation mode, for a pre-defined, relatively static interaction scenario, the lighting and shadow perturbation parameters of all cooperating spacecraft within the constellation are obtained. These parameters are then concatenated into a single, higher-dimensional joint parameter vector, forming a high-dimensional joint parameter space. Next, the CMA-ES algorithm is used to perform black-box optimization on this high-dimensional parameter space. CMA-ES is an advanced evolutionary algorithm, particularly adept at handling non-convex, ill-conditioned, and high-dimensional complex optimization problems. The objective function (fitness function) of this algorithm is defined as maximizing the classification error rate of the external black-box recognition system, or minimizing its confidence in correct classification. In each iteration, the CMA-ES algorithm samples a batch of candidate joint parameter vectors from a multidimensional Gaussian distribution. It then evaluates the confusion effect of each candidate scheme using the simulation model defined in step S202, and adaptively adjusts the mean, step size, and covariance matrix of the Gaussian distribution based on the evaluation results, thereby guiding the search towards a more optimal parameter region. This method requires no acquisition of the internal gradient or structural information of the recognition model; it only iteratively optimizes by evaluating the final effect, ultimately converging efficiently to a set of static lighting and shadow perturbation baseline parameters that achieve the optimal feature confusion effect.

[0060] Further details include the following processes: Joint parameter space definition: within the star cluster All light and shadow perturbation parameters (shadow parameters) of each cooperating spacecraft and glare parameters Connect them into a higher-dimensional joint parameter vector. And find the optimal solution in this joint parameter space. .

[0061] Black-box fitness function: Define a black-box fitness function. Its input is a set of joint parameter vectors The output is a scalar evaluation value. This function generates the result through internal simulation. The defined cooperative light and shadow pattern is submitted to the simulation model defined in step S202 for evaluation to obtain a scalar evaluation value, which quantifies the degree of confusion caused by the pattern to the recognition system.

[0062] The CMA-ES optimization process is as follows: (a) In the joint parameter space, initialize and maintain a multidimensional Gaussian distribution. ,in It is a mean vector. It's the step length. It is the covariance matrix.

[0063] (b) In each iteration, a population is sampled from the Gaussian distribution. (number of) candidate parameter vectors .

[0064] (c) For each candidate parameter vector Call the black-box fitness function An evaluation is conducted to obtain its fitness value; the fitness value is the scalar evaluation value.

[0065] (d) Update the parameters of the Gaussian distribution based on the fitness ranking of the population. , and This update process aims to bias the distribution of the next generation of samples towards the parameter regions that can produce higher fitness values.

[0066] (e) Repeat steps (b) to (d) until the convergence condition is met. This algorithm can adaptively adjust the shape and orientation of the search distribution to efficiently handle complex optimization problems that are high-dimensional, non-convex, and ill-conditioned.

[0067] In this embodiment, the generated static optimal feature obfuscation scheme based on CMA-ES is as follows: Figure 6 As shown.

[0068] As an optional implementation, in step S205, the optimal static lighting and shadow perturbation baseline parameters are adjusted using a dynamic adaptive camouflage strategy based on multi-agent reinforcement learning to obtain the optimal dynamic adaptive cooperative camouflage strategy, specifically including the following process: S2051: Real-time acquisition of dynamic information from the observer and the coordinated state within the constellation.

[0069] S2052: Based on the dynamic information of the observer and the cooperative state within the constellation, the optimal static light and shadow perturbation baseline parameters are continuously adjusted at the millisecond level through multi-agent Markov decision-making to generate the optimal dynamic adaptive cooperative camouflage strategy.

[0070] The multi-agent Markov decision-making process in step S2052 above is replaced by the following steps S401 to S403, as follows: S401: Input the state input vector of each agent into the policy network of the agent cluster and output a joint action vector; the state input vector includes the agent's current optimal static light and shadow perturbation baseline parameters, attitude information, state summary of other cooperating spacecraft in the cluster, and external sensing information such as relative distance, velocity vector, attitude angle, and estimated sensor orientation of the observing spacecraft obtained by lidar, passive optics, or radio sensors; the joint action vector contains the adjustment amount of the optimal static light and shadow perturbation baseline parameters for each cooperating spacecraft in the cluster.

[0071] S402: Establish a reward function based on the state input vector and the joint action vector; the positive reward term of the reward function includes the instantaneous decrease in the observer's confidence level and the duration of maintaining the observer in a state below the preset confidence level; the negative penalty term of the reward function includes the amplitude and frequency of the change in the light and shadow pattern of the entire star cluster and the event of the target being successfully identified.

[0072] S403: Based on the reward function, perform online lightweight fine-tuning of the policy network of the agent cluster to obtain the optimal dynamic adaptive cooperative camouflage strategy.

[0073] Specifically, a dynamic adaptive adjustment strategy based on Multi-Agent Reinforcement Learning (MARL) is used to address dynamic scenarios where observers make highly maneuverable and unpredictable orbital or attitude adjustments, endowing static baseline parameters with real-time adaptive capabilities. This process is modeled as a multi-agent Markov decision process, employing the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm. Within a framework of Centralized Training with Decentralized Execution (CTDE), a decision-making agent is trained for each spacecraft in the constellation. Through offline training and online fine-tuning of a multi-agent deep reinforcement learning system, it is enabled to continuously adjust the static baseline parameters at the millisecond level based on real-time dynamic information from onboard sensors and the collaborative state within the constellation, thereby generating an optimal dynamic adaptive cooperative camouflage strategy. During training, each agent's state input includes its current optimal static lighting and shadow perturbation baseline parameters, attitude information, a state summary of other cooperating spacecraft within the constellation, and external sensing information such as relative distance, velocity vector, attitude angle, and estimated sensor orientation of the observing spacecraft obtained through lidar, passive optics, or radio sensors. Based on this, all agents collaboratively output a joint action, which is a small, continuous adjustment to the optimal static lighting and shadow perturbation baseline parameters for each cooperating spacecraft. The effectiveness of this joint action is evaluated by a global reward function directly linked to the simulation model evaluation results defined in step S202, aiming to incentivize collaborative adjustment behaviors that continuously reduce the observer's identification confidence. The use of the MAPPO algorithm ensures the stability and convergence of policy updates, making it particularly suitable for safety-critical multi-agent continuous control tasks, while the CTDE framework guarantees that the learned cooperative policy possesses global optimality and real-time response capabilities. Ultimately, the system can be deployed to various spacecraft after sufficient offline training, and perform lightweight online fine-tuning during on-orbit operation, continuously and millisecond-level coordinated adjustments to static baseline parameters, thereby generating an optimal dynamic adaptive adjustment strategy.

[0074] Further, the following processes are included: State space definition: The state input of each agent. It is a high-dimensional vector that contains the current optimal static light and shadow perturbation baseline parameters and attitude information of our own spacecraft, the state summary of other cooperating spacecraft in the constellation, and external sensing information such as the relative distance, velocity vector, attitude angle, and estimated sensor orientation of the observing spacecraft obtained through lidar, passive optical or radio sensors.

[0075] Action space definition: The actions output by the policy network of an agent swarm. It is a joint action vector that contains the optimal static lighting perturbation baseline parameters for each cooperating spacecraft in the constellation. The adjustment amount enables a smooth and rapid collaborative continuous transformation of the light and shadow pattern, rather than generating entirely new parameters at each time step.

[0076] Reward function design: Reward function It is a scalar reward signal applied to the entire star cluster, the value of which is obtained by querying the black-box observer identification model. The function is a weighted combination, with its positive reward term including the instantaneous decrease in the observer's identification confidence and the duration of maintaining the observer in a low confidence state; its negative penalty term includes the amplitude and frequency of changes in the light and shadow pattern of the entire star cluster (to conserve energy and avoid being identified as active adjustment behavior due to drastic changes) and the event of the target being successfully identified.

[0077] Policy learning and execution: The policy network is trained using the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm. This type of algorithm is particularly suitable for safety-critical multi-agent continuous control tasks due to its stability of policy updates and reliability of convergence. The network is fully trained in an offline simulation environment under the framework of centralized training and distributed execution (CTDE) and then deployed to various spacecraft. During on-orbit execution, it is fine-tuned online based on real sensor data.

[0078] like Figure 7 The diagram illustrates a dynamic adaptive adjustment framework based on multi-agent reinforcement learning. Each spacecraft acts as an agent, integrating its own state, collaborative information, and the dynamic input strategy network of the observer to continuously adjust the static lighting baseline parameters. A global reward is generated through simulation evaluation, achieving collaborative optimization and real-time response under a centralized training and distributed execution mechanism.

[0079] As an optional implementation, in step S206, the optimal dynamic adaptive cooperative camouflage strategy is input into the cooperative physical projection execution stage to be transformed into a real light and shadow pattern in physical space, specifically including the following process: The Cooperative Physical Projection Execution Unit is responsible for receiving and executing the cooperative parameter instructions output by the static or dynamic strategy generation module. Through the configurable light and shadow projection system (such as laser array, digital micromirror device DMD, or manipulable optical components) carried by each spacecraft in the constellation, under communication and clock synchronization, it accurately generates corresponding combined shadow and glare patterns on the physical surface of the target spacecraft to complete the final feature camouflage.

[0080] Specifically, this step is responsible for the physical implementation of the calculated adjustment strategy. Its internal workflow begins with a parameter analysis and digital pattern generation unit, which converts the received cooperative parameter instructions into a high-resolution digital light field image in real time. Before projection, a distortion correction unit integrating a 3D model of the spacecraft performs reverse geometric pre-compensation on this image to ensure that the 2D pattern projected onto the curved target accurately matches the expected effect in 3D space. Subsequently, the corrected digital light field is sent to a hybrid optical projection assembly, where a Digital Micromirror Device (DMD) renders continuously varying grayscale shadow patterns, while a scannable laser diode generates high-energy-density glare spots. The entire process is precisely controlled by a drive controller under high-speed communication and clock synchronization, achieving a seamless, real-time cooperative response to the observer's highly maneuverable behavior.

[0081] like Figure 8 The diagram illustrates the implementation structure and processing steps of the collaborative physical projection execution stage. Figure 8 (a) is a schematic diagram of the overall optical projection system, including an exposure light source, a digital micromirror device, an objective lens, a substrate and a stage. The digital micromirror device spatially modulates the incident light, which is then focused by the objective lens and projected onto the surface of the substrate to achieve precise shaping of the light field pattern. Figure 8 (b) in the figure represents the initial stacked structure of the substrate and the photocurable resin; Figure 8 (c) in the image represents the striped exposure pattern formed under the action of a digital light field; Figure 8 In the diagram, (d) represents the structured surface obtained after selective curing. This process embodies the execution mechanism for transforming digital light field images into physical spatial light and shadow distributions.

[0082] In summary, the overall technical solution comprises five steps: 1) Target visual feature modeling and recognition system assumptions: constructing a digital target model and a simulation model; 2) Cooperative spatial light and shadow perturbation pattern definition: generating visual perturbation patterns; 3) Generation of a static optimal feature confusion scheme based on a covariance matrix adaptive evolution strategy (CMA-ES); 4) A dynamic adaptive adjustment strategy based on multi-agent reinforcement learning (MARL); and 5) Cooperative physical projection execution. Logically, these steps constitute a progressive relationship of "basic modeling—mechanism definition—static optimization—dynamic evolution—physical execution." First, the target feature modeling and recognition system assumes the game environment and black-box evaluation feedback as the basis, and the collaborative spatial light and shadow perturbation pattern defines the physical boundary of feature tampering. Second, the static optimal feature obfuscation scheme based on CMA-ES uses the above model and mechanism to optimize the camouflage baseline parameters in a relatively static scene. Subsequently, the dynamic adaptive camouflage strategy based on MARL uses this baseline to make real-time continuous collaborative adjustments for the high-maneuverability behavior of the observer. Finally, the decision command is input to the collaborative physical projection execution stage and transformed into a real light and shadow pattern in physical space.

[0083] This application aims to address the detection and classification challenges posed by advanced external visual recognition systems in multi-spacecraft interaction scenarios. Traditional passive adjustment strategies struggle to adapt to dynamic interactive environments. This application proposes a two-tiered strategy combining offline global optimization and online dynamic adaptation, where multiple spacecraft within a constellation collaboratively generate and project highly dynamic light and shadow patterns. First, for relatively static interaction scenarios, this application employs the Covariance Matrix Adaptive Evolution (CMA-ES) algorithm for offline baseline adjustment scheme calculation. This algorithm treats the observer's recognition model as a black box, eliminating the need to acquire its internal gradients. Through efficient global search in a high-dimensional joint parameter space, it converges to a set of optimal light and shadow parameter combinations that maximize the classification error rate or minimize the confidence of the recognition system. Second, to address the high maneuverability of the observer, this application introduces a Multi-Agent Reinforcement Learning (MARL) framework to train a distributed decision-making system. Each agent continuously and adaptively adjusts the static baseline parameters at the millisecond level based on real-time sensor data and constellation collaboration information. This system adopts a centralized training and distributed execution (CTDE) model to ensure the global optimality and real-time response capability of the strategy. Ultimately, the collaborative parameter commands generated by this two-layer strategy are sent to the physical projection system onboard the spacecraft to accurately generate dynamically changing combined shadows and glare, thereby achieving continuous feature obfuscation of external visual recognition systems and significantly improving the mission reliability of the spacecraft in complex dynamic environments.

[0084] This application proposes an active, dynamic camouflage framework that fundamentally solves the technical problem that traditional passive camouflage strategies cannot adapt to dynamic adversarial environments. The physical properties of traditional passive camouflage adjustments are fixed and cannot be changed, and their effectiveness is highly dependent on preset, specific observation conditions. This invention can utilize airborne sensors to acquire key variables of the interactive environment (such as spatial illumination vectors, relative position and attitude of the observer, etc.) in real time and use these variables as decision inputs. Based on these inputs, the constellation system can collaboratively and dynamically adjust and generate light and shadow projection patterns, continuously modulating the visual features of the spacecraft. This dynamic adaptability ensures that the feature adjustment strategy maintains high efficiency in various complex interactive scenarios, thereby significantly improving the adaptability and effectiveness of the strategy.

[0085] This application introduces an advanced optimization and learning mechanism based on Covariance Matrix Adaptive Evolutionary Strategy (CMA-ES) and Multi-Agent Proximal Policy Optimization (MAPPO), enabling the adjustment strategy to possess high complexity and unpredictability, effectively countering the modeling and prediction capabilities of the observer. The technical principle of passive adjustment is relatively well-defined, and the characteristic signals it generates are predictable, making it easy for the observer to build a feature database through long-term observation and specifically target and decipher it. This invention utilizes the CMA-ES algorithm for black-box optimization of a high-dimensional, nonlinear space composed of the joint light and shadow parameters of the entire star cluster, and combines it with the MAPPO algorithm to adjust the strategy in real time through multi-agent collaborative dynamic learning. The inherent laws of these advanced algorithms are difficult to decipher through conventional physical modeling or signal analysis methods, greatly increasing the difficulty for the observer to identify and counteract them.

[0086] This application employs an optimization algorithm that does not require gradients in the target model, achieving universal adjustment capabilities for black-box recognition models and significantly enhancing the practical value and long-term viability of the technology. The CMA-ES algorithm and multi-agent reinforcement learning method used in this invention both belong to the category of gradient-free optimization. They iteratively optimize their strategies simply by querying the black-box observation model's output responses to different inputs, without needing to understand its internal workings. This black-box interaction capability means that the technical solution of this invention has extremely strong versatility and can be applied to visual recognition systems of any type and structure.

[0087] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 9As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores data used in the process of generating the optimal dynamic adaptive cooperative camouflage strategy. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a feature camouflage method based on light and shadow projection in a star cluster game scenario.

[0088] Those skilled in the art will understand that Figure 9 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0089] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0090] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0091] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of the relevant data are carried out in compliance with the relevant data protection laws and policies of the country where the location is located, and with the authorization granted by the owner of the corresponding device.

[0092] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0093] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0094] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0095] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A feature camouflage method based on light and shadow projection in a star cluster game scenario, characterized in that, The feature camouflage method based on light and shadow projection in the star cluster game scenario includes: The key visual features of one's own spacecraft are acquired, and digital modeling is performed based on the key visual features to obtain a digital target model; the key visual features include: three-dimensional geometric contour, light reflection distribution function of surface material, and albedo under a specific electromagnetic band; Construct a simulation model of one or more external observer identification systems. The simulation model is regarded as a black box function that can only input queries and return evaluation results. The input of the black box function is an image of a spacecraft to be identified, and the output of the black box function is the classification result of the spacecraft image and the corresponding confidence score. Based on the distributed light and shadow projection units of multiple spacecraft within the constellation, a visual perturbation pattern is generated; the visual perturbation pattern includes: a stellar shadow perturbation model and a specular reflection glare perturbation model; the stellar shadow perturbation model is used to simulate complex lighting phenomena in the deep space environment; the specular reflection glare perturbation model is used to simulate complex lighting phenomena in the near-Earth orbit environment; Based on the digital target model, the simulation model, and the visual perturbation mode, static optimal feature confusion processing based on the covariance matrix adaptive evolution strategy is performed to obtain the optimal static light and shadow perturbation baseline parameters. The optimal dynamic adaptive cooperative camouflage strategy is obtained by adjusting the optimal static light and shadow perturbation baseline parameters through a dynamic adaptive camouflage strategy based on multi-agent reinforcement learning. The optimal dynamic adaptive cooperative camouflage strategy is input into the cooperative physical projection execution stage, and transformed into a real light and shadow pattern in physical space.

2. The feature camouflage method based on light and shadow projection in the star cluster game scenario according to claim 1, characterized in that, The stellar shadow perturbation model simulates the complex umbra and penumbra effects formed by the occlusion of main-sequence stars, neighboring stars, or other friendly spacecraft in a star cluster by coordinating the projection of a combination of shadows of specific shapes and intensities onto the surface of the target spacecraft, thereby changing the local illumination intensity of the target surface. The expression for the stellar shadow perturbation model is: ； in, The original target image; The image after shading has been applied; For pixels; For a single perturbation mask function, ; The number of cooperating spacecraft in the constellation; It is the first Shadow parameter vectors of a cooperating spacecraft; The coordinates of the center of the ellipse; For major and minor axes; ( ) represents the rotation angle; ) represents the intensity of the umbra; () represents the penumbra softness.

3. The feature camouflage method based on light and shadow projection in the star cluster game scenario according to claim 1, characterized in that, The specular reflection glare disturbance model simulates specular reflection or coordinated directional light interference caused by solar panels, optical windows or metal components receiving direct starlight at a specific angle by coordinating the projection of high-intensity light spots onto the surface of the target spacecraft. The expression for the specular reflection glare perturbation model is: ； in, The original target image; The image after applying glare; For pixels; For a single spot function, ; The number of cooperating spacecraft in the constellation; The intensity superposition function follows an exponential radial decay law; For pixels Euclidean distance to the center of the light spot; parameter vector For the first Glare parameter vectors of the cooperating spacecraft; The coordinates of the center of the light spot; The radius of action; Central strength; This represents the attenuation rate.

4. The feature camouflage method based on light and shadow projection in the star cluster game scenario according to claim 1, characterized in that, Based on the digital target model, the simulation model, and the visual perturbation pattern, static optimal feature obfuscation processing based on a covariance matrix adaptive evolution strategy is performed to obtain the optimal static lighting perturbation baseline parameters, specifically including: Based on the digital target model, the simulation model, and the visual perturbation mode, for a preset, relatively static interactive scenario, the light and shadow perturbation parameters of all cooperating spacecraft within the constellation are obtained; the light and shadow perturbation parameters include shadow parameters and glare parameters. The light and shadow perturbation parameters are combined into a joint parameter vector of a preset dimension to form a joint parameter space of a preset dimension. A covariance matrix adaptive evolution strategy is used to perform black-box optimization on the joint parameter space of the preset dimension to obtain the optimal static light and shadow perturbation baseline parameters; the objective function of the covariance matrix adaptive evolution strategy is defined as maximizing the classification error rate of the external recognition system or minimizing its confidence in correct classification.

5. The feature camouflage method based on light and shadow projection in the star cluster game scenario according to claim 4, characterized in that, A covariance matrix adaptive evolution strategy is used to perform black-box optimization on the joint parameter space of the preset dimension to obtain the optimal static lighting perturbation baseline parameters, specifically including: Define a black-box fitness function; the input of the black-box fitness function is a set of joint parameter vectors of a preset dimension, and the output is a scalar evaluation value; the black-box fitness function generates a cooperative lighting pattern defined by the joint parameter vectors of the preset dimension through internal simulation, and submits the cooperative lighting pattern to the simulation model for evaluation to obtain a scalar evaluation value; the scalar evaluation value quantifies the degree of confusion caused by the cooperative lighting pattern to the recognition system; Based on the black-box fitness function, the optimal solution is found in the joint parameter space of the preset dimension through the covariance matrix adaptive evolution strategy, and the optimal static light and shadow perturbation baseline parameters are obtained.

6. The feature camouflage method based on light and shadow projection in the star cluster game scenario according to claim 1, characterized in that, By employing a dynamic adaptive camouflage strategy based on multi-agent reinforcement learning, the optimal static lighting and shadow perturbation baseline parameters are adjusted to obtain the optimal dynamic adaptive cooperative camouflage strategy, which specifically includes: Real-time acquisition of dynamic information from the observer and the coordinated state within the constellation; Based on the dynamic information of the observer and the cooperative state within the constellation, the optimal static light and shadow perturbation baseline parameters are continuously adjusted at the millisecond level through multi-agent Markov decision-making to generate the optimal dynamic adaptive cooperative camouflage strategy.

7. The feature camouflage method based on light and shadow projection in the star cluster game scenario according to claim 6, characterized in that, The process of the multi-agent Markov decision-making specifically includes: Each agent's state input vector is input to the policy network of the agent cluster, outputting a joint action vector. The state input vector includes the agent's current optimal static light and shadow perturbation baseline parameters, attitude information, state summary of other cooperating spacecraft in the cluster, and external sensing information such as relative distance, velocity vector, attitude angle, and estimated sensor orientation of the observing spacecraft obtained through lidar, passive optical, or radio sensors. The joint action vector includes the adjustment amount for the optimal static light and shadow perturbation baseline parameters of each cooperating spacecraft in the cluster. A reward function is established based on the state input vector and the joint action vector; the positive reward term of the reward function includes the instantaneous decrease in the observer's confidence level and the duration of maintaining the observer in a state below the preset confidence level; the negative penalty term of the reward function includes the amplitude and frequency of the changes in the light and shadow pattern of the entire star cluster and the event of the target being successfully identified. Based on the reward function, the policy network of the agent cluster is fine-tuned online in a lightweight manner to obtain the optimal dynamic adaptive cooperative camouflage strategy.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the feature camouflage method based on light and shadow projection in the star cluster game scenario according to any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the feature camouflage method based on light and shadow projection in the star cluster game scenario as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the feature camouflage method based on light and shadow projection in the star cluster game scenario as described in any one of claims 1-7.

Images