Target camouflage system and method based on multi-modal reconnaissance and closed-loop optimization

Abstract

The present application relates to a target camouflage system and method based on multi-modal reconnaissance and closed-loop optimization, wherein the target camouflage system comprises: a multi-modal reconnaissance unit for synchronously collecting multi-spectral data of a target and an environment background; a data preprocessing unit for fusing the multi-spectral data to generate an environment feature map; a diagnosis unit for outputting a quantitative index representing a camouflage defect and a position heat map of the camouflage defect; an optimization decision unit for generating multiple sets of camouflage optimization schemes and screening an optimal camouflage scheme through a multi-objective evaluation model; an execution unit for deploying field equipment according to the optimal camouflage scheme; a verification unit for calling the multi-modal reconnaissance unit, the data preprocessing unit and the diagnosis unit to reevaluate the camouflage effect after the field equipment deployment and output a comprehensive score; if the comprehensive score after the deployment is lower than a preset threshold, the optimization decision unit and the execution unit are called to optimize the deployment and reevaluate the camouflage effect.

Description

Technical Field

[0001] This invention relates to the field of target camouflage technology, and in particular to a target camouflage system and method based on multimodal reconnaissance and closed-loop optimization. Background Technology

[0002] In the field of modern target camouflage, the effective application of camouflage technology is crucial for the concealment and security of targets. However, traditional camouflage techniques, such as camouflage nets and digital camouflage, have revealed a series of core flaws in practical applications, seriously affecting the camouflage effect and the concealment of targets.

[0003] First, the camouflage application is disconnected from the background environment, primarily due to a lack of matching guidance. Currently, existing camouflage equipment has developed into multi-spectral products, such as multispectral camouflage nets and infrared suppression materials, which possess certain capabilities in visible light, infrared, and radar stealth. However, in actual operational scenarios, operators often simply follow standard operating procedures, merely covering the target surface with camouflage nets, or relying on their own experience to choose a particular camouflage pattern. This approach lacks refined matching guidance for specific environmental backgrounds, resulting in operators lacking a clear understanding of key issues such as "how large is the gap between the current camouflage and the background environment," "which areas are at risk of exposure," and "what kind of convenient equipment needs to be added for local modification." Even when using multispectral camouflage nets, if their color, texture, and infrared emissivity deviate from the background environment, the target is still highly likely to be exposed. For example, in different seasons and regions, the color, texture, and infrared characteristics of the background will change significantly. If the camouflage equipment cannot accurately match these changes, it will be easily detected by increasingly comprehensive reconnaissance methods.

[0004] Secondly, the camouflage effect relies excessively on human experience and severely lacks scientific diagnosis. Currently, camouflage adjustments mainly depend on experienced personnel making on-site judgments. However, the accuracy of this judgment method is greatly affected by various factors, including the personnel's own experience level, their psychological state at the time, and ambient lighting conditions. When the camouflage net does not match the actual environment, operators often can only rely on experience to "guess," unsure of which parts need adjustment or what equipment needs to be added to improve the camouflage effect. Due to the lack of quantitative assessment methods, it is impossible to accurately locate defects in the camouflage process, leading to this "blind adjustment" method being not only inefficient but also difficult to control in terms of camouflage effect. For example, in complex and changing environments, lighting conditions are constantly changing, and different personnel may have significantly different judgments about the camouflage effect, making camouflage adjustments lack scientific rigor and accuracy.

[0005] Furthermore, traditional camouflage techniques lack rapid testing and closed-loop optimization mechanisms. After camouflage adjustments are made, there is a lack of quick and effective means to verify the results. Traditional testing methods mainly rely on visual observation or secondary reconnaissance for comparison, which makes it difficult to quantitatively assess whether the camouflage meets standards. More importantly, even if the camouflage effect is found to be unsatisfactory, there is a lack of a closed-loop optimization mechanism to promptly feed the "current effect" back to the adjustment decision-making process, thus hindering the formation of an iterative optimization cycle of "diagnosis → adjustment → verification → re-diagnosis." This makes it difficult to continuously improve and perfect the camouflage effect in response to environmental changes and actual needs.

[0006] Despite existing improvements such as multi-sensor reconnaissance and smart material camouflage, these solutions still have many shortcomings. In multi-sensor reconnaissance, the data acquired by multiple sensors is merely "overlayed" without feature-level fusion, making it impossible to accurately pinpoint camouflage defects. For example, the data acquired by visible light sensors, infrared sensors, and radar sensors are not deeply fused and analyzed, making it impossible to comprehensively and accurately identify problems when assessing camouflage effectiveness. Regarding intelligent diagnostic models, most are single-modal, such as those only diagnosing visible light, failing to cover multi-spectral camouflage requirements. With the development of modern reconnaissance technology, the enemy may conduct reconnaissance across multiple spectral bands, making single-modal diagnostic models insufficient for comprehensive camouflage needs. Furthermore, existing improvement solutions generally lack a closed-loop mechanism from "diagnosis" to "execution" to "verification." When camouflage mismatches are discovered, they cannot be scientifically and effectively remedied, significantly diminishing the improvement effect.

[0007] This demonstrates that traditional camouflage techniques suffer from core pain points such as "lack of matching guidance in application, reliance on human experience for adjustments, and inability to verify effects in a closed-loop manner," and existing improvement solutions have failed to effectively address these issues. Therefore, there is an urgent need to develop a full-process camouflage system capable of quantitatively diagnosing camouflage defects, intelligently generating optimization schemes, and verifying adjustment effects in a closed-loop manner. This system aims to meet the higher demands of camouflage technology in the modern environment and enhance the concealment and security of targets. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide a target camouflage system and method based on multimodal reconnaissance and closed-loop optimization.

[0009] To achieve the above-mentioned objectives, this invention provides a target camouflage system based on multimodal reconnaissance and closed-loop optimization, comprising: The multimodal reconnaissance unit, mounted on an airborne platform, is used to simultaneously acquire multispectral data of the target and the environmental background; the multimodal reconnaissance unit includes: a visible light camera, an infrared thermal imager, and radar; The data preprocessing unit is used to fuse multi-spectral data to generate environmental feature maps; The diagnostic unit analyzes the environmental feature map and the target's current state data based on a pre-trained conditional generative adversarial network model, and outputs a quantitative index representing the camouflage defects and a heat map of the location of the camouflage defects. The optimization decision unit is used to generate multiple camouflage optimization schemes based on quantitative indicators and location heatmaps, combined with a pre-set expert knowledge base and reinforcement learning algorithm, and select the optimal camouflage scheme through a multi-objective evaluation model. The execution unit is used to deploy on-site equipment according to the optimal camouflage plan; The verification unit is used to call the multimodal reconnaissance unit, data preprocessing unit, and diagnostic unit to re-evaluate the camouflage effect after the deployment of on-site equipment and output a comprehensive score after deployment. If the comprehensive score after deployment is lower than the preset threshold, the optimization decision unit and execution unit are called to perform iterative optimization until the comprehensive score reaches the preset threshold or the preset iteration limit is reached.

[0010] According to one aspect of the present invention, the data preprocessing unit performs time synchronization, spatial registration and noise removal on the multi-spectral data to generate a fused environmental feature map.

[0011] According to one aspect of the present invention, the pre-trained conditional generative adversarial network model in the diagnostic unit includes: a generator and a discriminator; The generator adopts a U-Net structure to map environmental feature maps and target current state data into virtual target images, and generates camouflage defect heatmaps based on intermediate layer features; The discriminator adopts the PatchGAN structure to judge the authenticity of the input image and outputs a quantitative index of the camouflage defects based on the intermediate layer feature regression; wherein, the input image is a real environment image or a virtual target image.

[0012] According to one aspect of the present invention, the pre-trained conditional generative adversarial network model in the diagnostic unit is obtained by training based on the following steps: A simulation environment is built based on a physics simulation engine, and an atmospheric radiation transfer model and radar scattering simulation module are integrated to generate a simulation training set of multimodal data. The conditional generative adversarial network is trained based on the simulation training set to generate a simulation pre-trained model. The conditional generative adversarial network model is obtained by fine-tuning the simulation pre-trained model using transfer learning technology and real multimodal data of a preset scale.

[0013] According to one aspect of the present invention, the conditional generative adversarial network model is trained by jointly optimizing adversarial loss and feature matching loss during the training process, and the loss function used is:

[0014]

[0015]

[0016] in, Represents the total loss function. Indicating resistance to loss, The feature matching loss is calculated based on the pre-trained ResNet-50 feature extractor. This represents the regularization coefficient that balances the adversarial loss and the feature matching loss. Represents the distribution of real-world environment images The mathematical expectation, This indicates that the discriminator recognizes the real-world image. The output probability, Represents the distribution of random noise. The mathematical expectation, This indicates that the discriminator recognizes the virtual target image. The output probability, This indicates the generator in the environmental feature map. Under the condition of random noise The generated virtual target image, Represents a random noise vector. Represents an environmental feature map. This represents the pre-trained ResNet-50 feature extractor.

[0017] According to one aspect of the present invention, in the optimization decision unit, the reinforcement learning algorithm adopts a deep deterministic policy gradient algorithm or a near-end policy optimization algorithm, wherein, in the process of generating multiple camouflage optimization schemes, the current quantitative index representing the camouflage defect, the target current state data and the environmental feature map are constructed in vector form to form the state space in the reinforcement learning process. The action space in the reinforcement learning process is constructed based on the type, quantity, deployment location coordinates, and deployment order of the equipment to be deployed; A reward function is constructed based on factors such as the degree of camouflage effect improvement, equipment cost, operation time, and penalty for plan failure. Using a pre-set expert knowledge base as the initial strategy and a pre-set sample for the experience replay pool, multiple camouflage optimization schemes are generated.

[0018] According to one aspect of the present invention, the multi-objective evaluation model comprehensively scores multiple camouflage optimization schemes, and selects the optimal camouflage scheme based on the best comprehensive score result; wherein, the comprehensive scoring formula is expressed as:

[0019] in, This indicates an improvement in camouflage effectiveness. Indicates equipment cost, The normalized baseline value representing the cost of equipment. Indicates the operation time. The normalized baseline value representing the operation time. , , Indicates the preset weight, and , .

[0020] According to one aspect of the present invention, the execution unit includes: a bidirectional communication interface module for communicating with an optimization decision unit, and a job module for communicating with the bidirectional communication interface module; The bidirectional communication interface module includes: a downlink port and an uplink port; The downlink port is used to receive the optimal camouflage scheme data generated by the optimization decision unit based on the optimal camouflage scheme and transmit it to the operation module; wherein, the optimal camouflage scheme data includes: task sequence, target location coordinates of each sub-task, equipment type and quantity, operation time window, accuracy requirements and tolerance range; The uplink port is used to receive the execution status data fed back by the operation module and feed it back to the optimization decision unit; wherein, the execution status data includes: the completion status of each sub-task, the deviation between the actual deployment location of the equipment and the target location, the operation time, and abnormal situation reports.

[0021] According to one aspect of the present invention, the preset expert knowledge base includes: a typical environmental feature library, a camouflage equipment characteristic library, a successful case library, and a camouflage standard library; The typical environment feature library is used to store the feature parameters of typical backgrounds in different environments, wherein the feature parameters include at least one of optical feature parameters, infrared feature parameters, and radar feature parameters. The camouflage equipment characteristic library is used to store the physical characteristic parameters of standard camouflage equipment and convenient equipment. The success case library is used to store successful solutions for historical camouflage operations; The camouflage standard library is used to store the specification parameters specified in the selected camouflage standards.

[0022] To achieve the aforementioned objectives, this invention provides a camouflage method for the aforementioned target camouflage system based on multimodal reconnaissance and closed-loop optimization, comprising the following steps: S1. Simultaneously acquire multi-spectral data of the target and environmental background, including: visible light image data, infrared image data, and radar image data; S2. Fuse multi-spectral data to generate an environmental feature map; S3. Based on the pre-trained conditional generative adversarial network model, analyze the environmental feature map and the target's current state data, and output quantitative indicators representing camouflage defects and a heat map of the location of camouflage defects; S4. Based on quantitative indicators and location heatmaps, and combined with a pre-set expert knowledge base and reinforcement learning algorithm, multiple camouflage optimization schemes are generated, and the optimal camouflage scheme is selected through a multi-objective evaluation model. S5. Deploy on-site equipment according to the optimal camouflage plan; S6. Re-evaluate the camouflage effect after the on-site equipment deployment and output the overall score after deployment; if the overall score after deployment is lower than the preset threshold, repeat steps S4 to S5 and re-evaluate the camouflage effect, and iterate and optimize accordingly until the preset conditions or the preset iteration limit is reached.

[0023] According to one aspect of the present invention, this approach fully and effectively achieves precise control over the camouflage effect. Specifically, through the diagnostic unit of the multimodal CMGAN, the camouflage defects are quantitatively assessed and precisely located, solving the pain point of traditional camouflage being "unquantifiable and blindly adjusted," thus making the camouflage effect measurable and verifiable.

[0024] According to one aspect of the present invention, this approach has the ability to adapt to complex environmental backgrounds. The multi-modal reconnaissance unit used can cover multiple spectral bands including visible light, infrared, and radar. Data preprocessing enables feature-level fusion, making it adaptable to various typical environments such as forests, deserts, and cities. It is highly versatile and has a wide range of applications.

[0025] According to one aspect of the present invention, the camouflage optimization scheme is scientific and efficient. Through the synergistic promotion of reinforcement learning and expert knowledge base, the generated optimization scheme takes into account the camouflage effect, equipment cost and operational efficiency. Furthermore, the comprehensive scoring of the multi-objective evaluation model of the bronze drum further ensures that the scheme is optimal and feasible, making the camouflage effect of the present scheme even better. According to one aspect of the present invention, this approach enables a hybrid collaborative operation of manual and mechanical processes during camouflage, solving the deployment challenges in areas difficult for personnel to access, shortening operation time, reducing the labor intensity of operators, and significantly improving operational efficiency.

[0026] According to one aspect of the present invention, this solution implements a closed-loop mechanism of "diagnosis-optimization-deployment-verification" for target camouflage optimization, enabling iterative optimization of camouflage effects, adapting to dynamic changes in the battlefield environment, and continuously improving target survivability; it also possesses dynamic adaptive capabilities. According to one aspect of the present invention, this approach effectively solves the problem of scarce real multimodal data by employing a combined model training strategy of simulation pre-training and experimental fine-tuning. The model has strong generalization ability and can be directly applied to actual camouflage operations, making the camouflage effect more fully meet the requirements of actual complex environments.

[0027] According to one aspect of the present invention, this solution utilizes a closed-loop mechanism of "diagnosis-optimization-deployment-verification" to upgrade the camouflage effect from "one-time adjustment" to "adaptive evolution." Each iteration is based on the verification results of the previous iteration for targeted optimization, until the target and environment are "multi-spectral fused." The closed-loop iteration ensures that the camouflage effect is measurable, verifiable, and optimizable, fundamentally solving the core pain points of traditional camouflage where "adjustments cannot be verified and effects are uncontrollable."

[0028] According to one aspect of the present invention, the solution, through a specially designed diagnostic unit, can output corresponding quantitative indicators and location heatmaps more comprehensively, thereby making the current camouflage and the environment more intuitive and accurate, and thus making the camouflage deployment more effective and targeted.

[0029] According to one aspect of the present invention, during the operation of the optimization decision-making unit, based on the diagnostic results, multiple optimization schemes are generated by combining a targeted preset expert knowledge base with reinforcement learning. Through a multi-objective evaluation model, the optimal camouflage scheme is selected under the comprehensive conditions of considering camouflage effect, cost, and efficiency, so that the scheme has a greater ability to fit the environmental background in terms of camouflage effect.

[0030] According to one aspect of the present invention, the diagnostic unit implements a dual-task collaborative CMGAN (Conditional Generative Adversarial Network) architecture that combines localization and quantization. In the localization task, the generator preserves spatial details through skip connections, outputting a virtual target image fused with the environment as a comparison benchmark. It also generates a pixel-level defect heatmap based on intermediate layer features, achieving precise spatial localization of defects and effectively solving the problem of "where is the defect?". In the quantization task, the discriminator, while determining the authenticity of the image, regresses and outputs multi-spectral quantization indicators based on intermediate layer features, achieving a quantitative assessment of the defect severity and effectively solving the problem of "how severe is the defect?". Furthermore, the two tasks share underlying features and reinforce each other. The localization results guide the quantization of the area of ​​interest, and the quantization results verify the accuracy of the localization, forming a collaborative optimization closed loop that ensures the accuracy and reliability of the output results.

[0031] According to one aspect of the present invention, a multi-spectral conditional input mechanism is constructed, wherein the environmental feature map of the three-channel tensor contains more multi-spectral environmental information, and the target current state data is used as the conditional information input. By comparing the target current state with the environmental features, the present invention can more accurately identify camouflage defects, that is, the gap between the current camouflage and the ideal fusion state, providing quantifiable and executable input for subsequent optimization decisions, and providing a reliable foundation for generating the optimal camouflage scheme.

[0032] According to one aspect of the present invention, a standardized data interface and status feedback mechanism are designed between the diagnosis, decision-making, and execution stages to realize task parsing, automatic allocation, exception handling, and status reporting, forming an organically coordinated closed-loop system.

[0033] According to one aspect of the present invention, during the on-site equipment deployment phase, the method of operation based on the manual and mechanical collaborative operation of the execution unit makes the operation mode of the present invention closer to the actual complex scenario, making the present invention easier to promote and implement. Attached Figure Description

[0034] Figure 1 This is a structural block diagram of the target camouflage system based on multimodal reconnaissance and closed-loop optimization of the present invention. Detailed Implementation

[0035] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The embodiments cannot be described in detail here, but the embodiments of the present invention are not limited to the following embodiments.

[0036] like Figure 1 As shown, according to one embodiment of the present invention, a target camouflage system based on multimodal reconnaissance and closed-loop optimization includes: a multimodal reconnaissance unit, a data preprocessing unit, a diagnostic unit, an optimization decision-making unit, an execution unit, and a verification unit; wherein, the multimodal reconnaissance unit is mounted on an aerial platform and is used to simultaneously collect multispectral data of the target and environmental background; in this embodiment, the aerial platform can be a device capable of hovering in the air, such as an unmanned aerial vehicle (UAV), a manned aircraft, or a satellite; wherein, the aerial platform is preferably implemented as an UAV, which has the advantages of miniaturization and ease of operation. Furthermore, after the aerial platform takes off, it performs flight missions according to a preset flight path, simultaneously collecting multispectral data of the target and background. The flight path settings include: Flight altitude: Set according to the target type and environmental characteristics, usually between 50m and 150m; Incident angle: controlled within ≤45°, used to simulate other aerial reconnaissance perspectives; Navigation direction: within 30° to the left or right of the direction of sunlight, with a solar altitude angle > 30°; Atmospheric conditions: Visibility ≥ 5km, avoiding the impact of severe weather conditions such as fog, haze, rain, and snow on data quality.

[0037] In this embodiment, the multimodal reconnaissance unit includes: a visible light camera, an infrared thermal imager, and a radar; furthermore, the radar may be synthetic aperture radar (SAR). The multimodal reconnaissance unit collects the following data from each module to construct multispectral data of the target and environmental background, which are: Visible light camera: Acquires visible light images of the target and its surrounding environment, with a resolution ≥0.1m; Infrared thermal imager: Collects thermal radiation data of the target and the environmental background, with a sensitivity of ≤0.05℃; Radar: Collects radar scattering data of the target and its surrounding environment, with a resolution of ≤0.5m; In this embodiment, an environmental sensor can be further installed in the multimodal reconnaissance unit to synchronously record environmental parameters such as light intensity, temperature, humidity, and wind speed.

[0038] Furthermore, all data collected by the multimodal reconnaissance unit is supplemented with GPS / INS pose information to ensure the accuracy of subsequent spatiotemporal registration.

[0039] In this embodiment, the data preprocessing unit is communicatively connected to the multimodal reconnaissance unit. The communication is achieved through wired or wireless communication to receive multi-spectral data, thereby enabling the fusion of multi-spectral data and generating an environmental feature map.

[0040] In this embodiment, the diagnostic unit and the data preprocessing unit are communicatively connected, using either wired or wireless communication to receive environmental feature maps. Furthermore, the diagnostic unit also needs to receive target current state data. Based on a pre-trained conditional generative adversarial network model, the environmental feature maps and target current state data are analyzed, outputting quantitative indicators representing camouflage defects and a heatmap showing the location of these defects. In this embodiment, the target current state data includes: a color distribution histogram of the existing camouflage pattern on the target surface, an infrared emissivity distribution map, and a radar reflectivity distribution map. The color distribution histogram is generated from target surface images acquired by a visible light camera through color space conversion and histogram statistics. The infrared emissivity distribution map is generated by inverting thermal radiation data acquired by an infrared thermal imager combined with prior information on the target surface material's emissivity. The radar reflectivity distribution map is generated from radar images acquired by a SAR radar through scattering feature extraction.

[0041] In this embodiment, the communication connection between the decision-making unit and the diagnostic unit is optimized. The communication is achieved through wired or wireless communication to receive quantitative indicators and location heatmaps. Then, the decision-making unit generates multiple camouflage optimization schemes based on the quantitative indicators and location heatmaps, combined with a preset expert knowledge base and reinforcement learning algorithm, and selects the optimal camouflage scheme through a multi-objective evaluation model.

[0042] In this embodiment, the execution unit and the optimization decision unit are communicatively connected. The communication is achieved through wired or wireless communication to receive the optimal camouflage scheme and then deploy on-site equipment according to the optimal camouflage scheme.

[0043] In this embodiment, the verification unit is communicatively connected to the multimodal reconnaissance unit, the data preprocessing unit, the diagnostic unit, and the optimization decision unit, respectively. It is used to call the multimodal reconnaissance unit, the data preprocessing unit, and the diagnostic unit to re-evaluate the camouflage effect after the deployment of on-site equipment and output the comprehensive score after deployment. If the comprehensive score after deployment is lower than the preset threshold, the optimization decision unit and the execution unit are called to perform iterative optimization until the comprehensive score reaches the preset threshold or the preset iteration limit is reached.

[0044] According to one embodiment of the present invention, a data preprocessing unit performs time synchronization, spatial registration, and noise removal on multi-spectral data to generate a fused environmental feature map. In this embodiment, during time synchronization, time alignment of multi-sensor data is achieved through a hardware trigger signal, with a synchronization accuracy of less than or equal to 10 ms. During spatial registration, based on the Scale Invariant Feature Transform (SIFT) feature matching algorithm and combined with GPS / INS pose information, visible light image data, infrared image data, and radar image data are projected onto a unified coordinate system, with the registration error controlled within a preset range, for example, less than or equal to 0.1 m. During noise removal, Gaussian filtering is applied to the infrared image data to suppress thermal noise; speckle suppression is applied to the radar image data to improve image quality. Thus, the fused environmental feature map provides high-quality input for the subsequent diagnostic unit.

[0045] In this embodiment, the environmental feature map is generated by feature-level fusion of preprocessed visible light image data, infrared image data, and radar image data. It is a three-channel tensor: channel 1 represents visible light texture features (encoded after SIFT feature extraction), channel 2 represents infrared thermal features (normalized temperature distribution values), and channel 3 represents radar scattering features (normalized reflectivity values). In this embodiment, the environmental feature map is key data connecting the data preprocessing unit and the diagnostic unit, and it is represented as follows:

[0046] in, Represents an environmental feature map. This indicates a splicing operation along the channel dimension. Representing a three-dimensional tensor, Image height, Image width; Represents the visible light image The single-channel texture feature map (i.e., visible light texture feature) obtained by SIFT feature extraction and encoding is used as channel 1 to represent the geometric structure and texture details of the background, with a size of [size missing]. ; Indicates infrared thermal image Temperature normalization (i.e., infrared thermal characteristics) is performed, and the calculation formula is as follows: ,in Infrared thermal image The thermal radiation temperature value corresponding to the middle pixel. The minimum and maximum scene temperatures are used as channel 2 to characterize the background thermal radiation distribution, with a size of [value missing]. ; Indicates the SAR radar image Perform reflectivity normalization processing Or logarithmic compression That is, radar scattering characteristics, in which This is the radar cross section (RCS) value. The minimum and maximum reflectance values ​​of the scene are used as channel 3 to characterize the radar scattering properties of the background, with a size of [value missing]. .

[0047] By normalizing sensor data of three different physical quantities into the same tensor space through the environmental feature map set above, the problem of heterogeneous multimodal data is solved, providing standardized input for subsequent deep learning processing. The environmental feature map effectively represents the multispectral reconnaissance characteristics, clearly embeds the data flow of the entire conditional generative adversarial network model, and highlights its four major technical features: "clear physical meaning, conditional constraint generation, comparable diagnostic benchmarks, and unified multispectral expression".

[0048] According to one embodiment of the present invention, the pre-trained conditional generative adversarial network model in the diagnostic unit includes a generator and a discriminator. In this embodiment, the generator adopts a U-Net structure to map environmental feature maps and target current state data into a virtual target image, and generates a camouflage defect heatmap based on intermediate layer features. Specifically, the encoder part of the generator contains 6 convolutional layers (3×3 kernel size, stride 2), each followed by a batch normalization layer and a LeakyReLU activation function to extract input features. The decoder part contains 6 deconvolutional layers (3×3 kernel size, stride 2), each followed by a batch normalization layer and a ReLU activation function, and is concatenated with the corresponding encoder layer features through skip connections to preserve spatial detail information. The final output layer is a Tanh activation function, outputting a virtual target image. The generator undertakes the localization task and generates a camouflage defect heatmap based on intermediate layer features.

[0049] In this embodiment, during the input of the environmental feature map and the target's current state data to the generator, a random noise vector (which follows a Gaussian distribution) is also included to increase the diversity of the generated results. Then, the target's current state data is concatenated with the noise vector as conditional information and mapped to the environmental feature map to generate a corresponding virtual target image, represented as follows: ,in, Represents a random noise vector. Represents environmental feature maps. In virtual target images. In the middle, environmental feature map This approach provides multispectral prior information about the current environmental background, enabling the generation of virtual target images that blend seamlessly with the specific environmental background. This virtual target image serves as a reference benchmark for "ideal camouflage," achieving conditional generation that adapts to the environment. Traditional multimodal generative adversarial networks (GANs) use only random noise as input, and the content of their generated images depends on the statistical distribution of the training dataset. They cannot conditionally generate images based on the specific environmental background detected in real-time (such as visible light texture, infrared thermal features, and radar scattering features). Therefore, images generated by traditional methods are irrelevant to the specific environment in the current reconnaissance scene and cannot serve as a benchmark for "ideal camouflage," thus lacking the advantages of this approach.

[0050] Furthermore, the intermediate layer features (i.e., intermediate layer feature maps) output by the generator during the generation of virtual target images are used to obtain a camouflage defect heatmap through a convolutional layer. The value of each pixel in the camouflage defect heatmap represents the probability that a camouflage defect exists at that location.

[0051] In this embodiment, the discriminator employs a PatchGAN structure to determine the authenticity of the input image and outputs a quantitative index of camouflage defects based on intermediate layer feature regression. The input image can be a real environment image or a virtual target image. Specifically, the discriminator's PatchGAN structure contains 5 convolutional layers (4×4 kernel size, stride 2), each followed by a batch normalization layer and a LeakyReLU activation function. The final output is an N×N feature map, where each feature point corresponds to a receptive field region of the input image, and the output value represents the probability that the region is a real image. The discriminator performs the quantization task, outputting a quantitative index of camouflage defects based on intermediate layer feature regression.

[0052] Furthermore, quantification metrics are obtained by regressing the features from the discriminator's intermediate layer through a fully connected layer. These quantification metrics include multiple dimensions, namely: Color similarity: Calculate the color difference between the target and the ambient background based on the CIE Lab color space. Color difference The smaller the value, the higher the color blending. Infrared feature similarity: Calculates the infrared emissivity deviation between the target surface infrared emissivity and the average emissivity of the environmental background. Infrared emissivity deviation The smaller the value, the better the heat fusion effect; Radar feature similarity: Calculates the radar reflectivity deviation between the target's radar cross-section (RCS) and the average reflectivity of the environmental background. Radar reflectivity deviation The smaller the value, the better the radar stealth effect.

[0053] In this embodiment, to further improve the accuracy of the quantitative indicators, the dimensions of the quantitative indicators can be selectively increased, that is: Texture similarity: Based on the gray-level co-occurrence matrix (GLCM), the texture features of the target and the background are extracted, and the similarity of indicators such as contrast, correlation, energy, and homogeneity is calculated.

[0054] In this embodiment, the discriminator receives real environment images or virtual target images to perform real / fake identification and defect quantification. In real / fake identification, the result is output for adversarial training. Defect quantification, based on intermediate layer features, regresses and outputs a quantitative index of the faking defects. Throughout the discriminator's inference phase, the environmental feature map serves as shared conditional information, guiding the discriminator to focus on multi-spectral differences and ensuring that the quantification index aligns with the specific environmental background.

[0055] According to one embodiment of the present invention, the pre-trained conditional generative adversarial network model in the diagnostic unit is obtained by training based on the following steps: A simulation environment was built based on a physics simulation engine, integrating an atmospheric radiative transfer model and a radar scattering simulation module to generate a multimodal simulation training set. Specifically, a high-fidelity simulation environment was constructed based on Unreal Engine 5.1, integrating the MODTRAN atmospheric radiative transfer model and the CST microwave studio radar scattering simulation module to generate a multimodal simulation dataset covering typical scenarios such as forests, deserts, cities, and snowfields, totaling 500,000 samples. Each sample includes: an environmental background image of the target area (visible light image data, infrared image data, and radar image data), the original target image (including camouflage), camouflage defect annotation information (automatically generated by the simulation engine), and the corresponding optimized target image.

[0056] The conditional generative adversarial network is trained based on the simulation training set to generate a simulation pre-trained model. A conditional generative adversarial network (GAN) model is obtained by fine-tuning a simulation pre-trained model using transfer learning technology and real multimodal data of a preset scale. In this embodiment, real multimodal data of a preset scale (no less than 500 sets) can be collected in a selected training field, and the simulation pre-trained model can be fine-tuned using transfer learning technology to adapt the model to the optical / electromagnetic characteristics of the real physical environment.

[0057] According to one embodiment of the present invention, the conditional generative adversarial network model is trained by jointly optimizing adversarial loss and feature matching loss during the training process, and the loss function used is:

[0058]

[0059]

[0060] in, Represents the total loss function. Indicating resistance to loss, The feature matching loss is calculated based on the pre-trained ResNet-50 feature extractor. This represents the regularization coefficient that balances the adversarial loss and the feature matching loss. Represents the distribution of real-world environment images The mathematical expectation, This indicates that the discriminator recognizes the real-world image. The output probability, that is, the probability of , represents For the possibility of a real image, Represents the distribution of random noise. The mathematical expectation, This indicates that the discriminator recognizes the virtual target image. The output probability, This indicates the generator in the environmental feature map. Under the condition of random noise The generated virtual target image, Represents a random noise vector. Represents an environmental feature map. This represents the pre-trained ResNet-50 feature extractor.

[0061] According to one embodiment of the present invention, the diagnostic unit may optionally be further equipped with a continuous update module with a continuous learning mechanism. After the target camouflage system is deployed, the "diagnosis-deployment-verification" data generated in each actual camouflage operation is manually reviewed and included in the training set. The conditional generative adversarial network model is incrementally updated periodically to fully ensure its accuracy in analyzing actual scenarios.

[0062] According to one embodiment of the present invention, in the optimization decision unit, the reinforcement learning algorithm employs a deep deterministic policy gradient algorithm or a proximal policy optimization algorithm. Specifically, during the generation of multiple camouflage optimization schemes, the quantitative index representing the camouflage defect, the target's current state data, and the environmental feature map are constructed in vector form to form the state space in the reinforcement learning process. The state space is defined as follows:

[0063] in, Color difference (multi-region), represented in vector form. Infrared emissivity deviation, expressed in vector form. Let be the radar reflectivity deviation vector, expressed in vector form. The state represented by the target's current state data, i.e., the target's current camouflage state encoding, is represented in vector form. The environment type encoding obtained from the environment feature map is represented in vector form. This is a record of historical operations.

[0064] Furthermore, an action space for the reinforcement learning process is constructed based on the type, quantity, deployment location coordinates, and deployment order of the equipment to be deployed; specifically, the action space is defined as:

[0065] in, The type of equipment to be deployed (such as camouflage nets, infrared shielding blankets, spray paint, branches, fallen leaves, sand, etc.). The quantity or amount of equipment to be deployed. The coordinates of the deployment location of the equipment to be deployed. The deployment order.

[0066] Furthermore, a reward function is constructed based on factors such as the degree of camouflage effect enhancement, equipment cost, operation time, and penalty for plan failure; specifically, the reward function is defined as:

[0067] in, To improve the camouflage effect, For equipment costs, For operation time, This is a penalty item for a failed plan. , , , Preset weights.

[0068] Furthermore, a pre-set expert knowledge base is used to provide initial strategies and pre-set samples for the experience replay pool in the reinforcement learning process, thereby generating multiple camouflage optimization schemes. In this embodiment, the pre-set expert knowledge base includes: a typical environment feature library, a camouflage equipment characteristic library, a successful case library, and a camouflage standard library; wherein, the typical environment feature library is used to store the feature parameters of typical backgrounds in different environments, wherein the feature parameters include at least one of optical feature parameters, infrared feature parameters, and radar feature parameters; in this embodiment, the different environments involved are typical backgrounds such as forests, deserts, snowfields, cities, and oceans; and the specific feature parameters can be set as environmental factors that affect camouflage, such as background main color, texture type, infrared emissivity distribution, and radar reflectivity range. In this embodiment, the camouflage equipment characteristic library is used to store the physical characteristic parameters of standard camouflage equipment and camouflage equipment. Standard camouflage equipment can be camouflage nets, infrared shielding blankets, camouflage paints, etc., while camouflage equipment can be tree branches, fallen leaves, sand, gravel, etc. The specific physical characteristic parameters can be set as physical factors that affect camouflage, such as visible light reflectivity curves, infrared emissivity, radar cross section, geometric dimensions, mass, connection method, etc.

[0069] In this embodiment, the success case library is used to store successful schemes of historical camouflage operations; each successful scheme may include corresponding environmental features (or environmental feature map), diagnostic results, selected optimal camouflage scheme, deployment effect and comprehensive score, which are used for pre-training of the initial strategy of reinforcement learning.

[0070] In this embodiment, the camouflage standard library is used to store the specification parameters specified in the selected camouflage standards; for example, the specification parameters such as camouflage patterns, color ratios, and spot sizes specified in camouflage standards (such as GJB 453-88 and GJB 4004-2000).

[0071] Therefore, the process of generating multiple camouflage optimization schemes by the optimization decision-making unit can be divided into the following steps: Initialization phase: Successful solutions from a pre-set expert knowledge base are used for imitation learning pre-training of the reinforcement learning policy network, enabling the initial policy to have basic decision-making capabilities and solving the cold start problem in reinforcement learning; Experience replay pool: Successful solutions from the pre-set expert knowledge base are stored as pre-set experience samples in the experience replay pool, which, together with experience samples generated from actual system operations, are used to train the reinforcement learning policy network. Solution generation stage: The reinforcement learning policy network outputs the corresponding action space A based on the current state space S, thereby generating multiple candidate camouflage optimization solutions; at the same time, the expert knowledge base retrieves cases similar to the current state space S and outputs similar solutions as a supplement to the candidate camouflage optimization solutions; thus, the multiple camouflage optimization solutions generated by the reinforcement learning policy network and the camouflage optimization solutions retrieved from the expert knowledge base are merged to form a candidate solution set, which forms the final multiple camouflage optimization solutions. Furthermore, the multi-objective evaluation model comprehensively scores multiple camouflage optimization schemes, and selects the optimal camouflage scheme based on the best comprehensive score; the comprehensive score formula is expressed as:

[0072] in, This indicates an improvement in camouflage effectiveness. Indicates equipment cost, The normalized baseline value representing the cost of equipment. Indicates the operation time. The normalized baseline value representing the operation time. , , Indicates the preset weight, and , .

[0073] In this embodiment, the camouflage effect is improved. Represented as:

[0074] in, This indicates the overall score under the current camouflage status. This represents the overall score in the previous disguise state.

[0075] According to one embodiment of the present invention, after each actual camouflage operation, the relevant operation data (such as diagnostic results, selected optimal camouflage scheme, deployment effect) are stored as experience samples in the experience replay pool. The reinforcement learning policy network is incrementally updated periodically using Deep Deterministic Policy Gradient (DDPG) or Proximal Policy Optimization (PPO) algorithms, so that the reinforcement learning policy network gradually adapts to new environmental changes and operation constraints, thereby being able to better generate camouflage optimization schemes that conform to different actual environments.

[0076] According to one embodiment of the present invention, the execution unit includes: a bidirectional communication interface module for communicating with an optimization decision unit, and a work module for communicating with the bidirectional communication interface module; in this embodiment, the bidirectional communication interface module includes: a downlink port and an uplink port; wherein, the downlink port is used to receive optimal camouflage scheme data generated by the optimization decision unit based on the optimal camouflage scheme and transmit it to the work module; wherein, the optimal camouflage scheme data includes: task sequence (such as including subtask ID, task type, and dependency relationship), target location coordinates (UTM coordinate system) of each subtask, equipment type and quantity, work time window, accuracy requirements, and tolerance range; the uplink port is used to receive execution status data fed back by the work module and feed it back to the optimization decision unit; wherein, the execution status data includes: completion status of each subtask (not started / in execution / completed / failed), deviation between the actual deployment location of the equipment and the target location, work time, and abnormal situation report.

[0077] In this embodiment, the optimal camouflage scheme data for the operation of the task module can be obtained by the optimization decision unit or the execution unit, so that the task module can perform the corresponding actions.

[0078] In this embodiment, the operation module can employ manual tools, mechanically assisted laying tools, drone-assisted deployment tools, handheld command terminals, etc. Specifically, manual operation can be performed using manual tools and / or handheld command terminals. The handheld command terminal is used to display corresponding deployment instructions to facilitate accurate completion of the corresponding operations, as follows: Camouflage net adjustment and installation: If the target has been camouflaged, make local adjustments according to the requirements of the optimal camouflage scheme: add folds to simulate terrain undulations, adjust the hanging angle to adapt to the direction of light, and interweave vegetation on the net surface to break the regular texture. If additional camouflage netting is needed, the workers will unfold the netting and cover the target surface, securing it with magnets or hooks to ensure that the netting fits the target outline and the edges transition naturally to the ground.

[0079] Infrared shielding equipment coverage: For areas with diagnosed infrared defects (such as engine compartments and exhaust pipes), operators cover the high-temperature surfaces with infrared shielding blankets or infrared suppression covers, ensuring complete coverage and secure fixation to prevent heat leakage.

[0080] Rapid spraying: For small areas with significant color differences (such as parts of armor plates), workers use portable spraying equipment to spray temporary camouflage paint that matches the main color of the background. When spraying, use irregular patches to mimic the background texture (such as mottled leaves or sandy smudges) and avoid regular edges.

[0081] Convenient equipment installation and modification: Vegetation materials: Workers insert branches and clumps of grass into the mesh of the camouflage net or fix them to protruding parts of the vehicle body to simulate the three-dimensional layers of natural vegetation; leaf materials are spread manually or by drones to cover flat areas such as armor plates and turret tops. Terrain-related: Workers evenly spread sand and gravel on the target surface and surrounding ground to blend the target outline into the terrain; the thickness of the layer is dynamically adjusted according to environmental characteristics (e.g., 3-5cm in desert environments, 1-2cm in forest humus). Assisted shaping: Based on real-time guidance displayed on the drone or handheld terminal, operators can make fine adjustments to the deployed equipment, such as adjusting the angle of tree branches, replenishing fallen leaves, and trimming the edges of camouflage nets, to minimize human traces.

[0082] During operation, the drone hovers to provide real-time positioning guidance from a top-down perspective, while the handheld terminal displays a heat map of the target area and the expected coverage range, allowing operators to complete precise operations based on the guidance.

[0083] In this embodiment, the mechanically assisted laying tool can be mounted on ground vehicles (such as manned or unmanned vehicles) and perform one or more of the following operations according to the scheme requirements: unfolding and partially adjusting camouflage nets, covering and fixing infrared shielding equipment, temporary spraying of color difference areas by rapid spraying equipment, and uniform spreading of granular materials such as sand and gravel; the machine can switch different execution ends through quick-change tool interface to adapt to the laying needs of multiple types of equipment.

[0084] In this embodiment, the drone-assisted delivery tool is mounted on the drone platform. The drone delivers bundles of branches, grass clumps, fallen leaves, or lightweight camouflage equipment to the top and surrounding area of ​​the target based on the received location coordinates. During the delivery process, visual guidance is used to achieve point delivery, and onboard sensors monitor the trajectory of the equipment in real time. If the deviation from the preset landing point exceeds the tolerance range, a second delivery is triggered or ground personnel are notified to make adjustments.

[0085] In this embodiment, after the execution unit completes the on-site equipment deployment, it feeds back the actual deployment location (obtained via RTK-GPS or visual positioning), operation time, and any abnormal situations to the optimization decision module via the uplink port. Furthermore, if a subtask fails or the deployment deviation exceeds the tolerance range, the execution unit automatically triggers supplementary operations or notifies the optimization decision module to adjust subsequent tasks.

[0086] In this embodiment, the execution unit may optionally be further equipped with a self-testing module. After the on-site equipment deployment is completed, the self-testing module triggers the multimodal reconnaissance unit to collect visible light images of the deployment location. By comparing whether the deployment area in the collected visible light images completely covers the corresponding area of ​​the optimal camouflage scheme, the uncovered or insufficiently covered areas are identified. Furthermore, instructions are given through a handheld instruction terminal, and commands are issued to mechanically assisted laying tools and drone-assisted delivery tools to correct the uncovered areas. After the self-test is passed, the verification unit can be triggered to execute the verification process.

[0087] According to one embodiment of the present invention, after the execution unit completes the deployment of on-site equipment, the verification unit triggers the execution verification process. Specifically, the multimodal reconnaissance unit, data preprocessing unit, and diagnostic unit are invoked to re-evaluate the camouflage effect after the deployment of on-site equipment, and a comprehensive score after deployment is output. If the comprehensive score after deployment is lower than a preset threshold, the optimization decision unit and execution unit are invoked for iterative optimization until the comprehensive score reaches the preset threshold or the preset maximum number of iterations is reached. In this embodiment, the comprehensive score used in the process of re-evaluating the camouflage effect after the deployment of on-site equipment is consistent with the comprehensive score method of the aforementioned multi-objective evaluation model, and will not be repeated here.

[0088] In this implementation, the final iterative optimization is completed when the optimization decision-making unit and execution unit are invoked to optimize the deployment and re-evaluate the camouflage effect until one of the following convergence conditions is met: Convergence condition 1: The overall score reaches a preset threshold (e.g., ≥90%). Convergence condition 2: The improvement in the overall score between two consecutive iterations is ≤2%, and the current overall score is ≥85%; Convergence condition 3: Reaching the preset maximum number of iterations (set to 3 based on experimental statistics).

[0089] Once the convergence condition is met, the system outputs the current optimal solution and the final camouflage effect report, and stores the data of this operation (diagnostic results, selected solution, deployment effect, and comprehensive score) in the expert knowledge base for incremental training of subsequent reinforcement learning models.

[0090] According to one embodiment of the present invention, each module in the above-described target camouflage system can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0091] In this embodiment, the memory may be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc.

[0092] In this embodiment, the processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0093] like Figure 1 As shown, according to one embodiment of the present invention, the present invention provides a camouflage method for the aforementioned target camouflage system based on multimodal reconnaissance and closed-loop optimization, comprising the following steps: S1. Simultaneously acquire multi-spectral data of the target and environmental background, including: visible light image data, infrared image data, and radar image data; S2. Fuse multi-spectral data to generate an environmental feature map; S3. Based on the pre-trained conditional generative adversarial network model, analyze the environmental feature map and the target's current state data, and output quantitative indicators representing camouflage defects and a heat map of the location of camouflage defects; S4. Based on quantitative indicators and location heatmaps, and combined with a pre-set expert knowledge base and reinforcement learning algorithm, multiple camouflage optimization schemes are generated, and the optimal camouflage scheme is selected through a multi-objective evaluation model. S5. Deploy on-site equipment according to the optimal camouflage plan; S6. Re-evaluate the camouflage effect after the on-site equipment deployment and output the overall score after deployment; if the overall score after deployment is lower than the preset threshold, repeat steps S4 to S5 and re-evaluate the camouflage effect, and iterate and optimize accordingly until the preset conditions or the preset iteration limit is reached.

[0094] According to one embodiment of the present invention, in step S1, a multi-modal reconnaissance unit mounted on an airborne platform is used to collect multi-spectral data. The specific collection method and process have been clearly stated in the foregoing and will not be repeated here.

[0095] According to one embodiment of the present invention, in step S2, a data preprocessing unit is used to receive the collected multi-spectral data to construct an environmental feature map. The specific implementation method has been clearly stated in the foregoing content and will not be repeated here.

[0096] According to one embodiment of the present invention, in step S3, a diagnostic unit is used to generate quantitative indicators and location heatmaps. The specific implementation method has been clearly stated in the foregoing and will not be repeated here.

[0097] According to one embodiment of the present invention, in step S4, an optimization decision unit is used to generate multiple camouflage optimization schemes and select the optimal camouflage scheme; the specific implementation method has been clearly stated in the foregoing content and will not be repeated here.

[0098] According to one embodiment of the present invention, in step S5, the execution unit deploys on-site equipment according to the optimal camouflage scheme; and the specific implementation method has been clearly stated in the foregoing content, and will not be repeated here.

[0099] According to one embodiment of the present invention, in step S6, a verification unit is used to evaluate the camouflage effect after the deployment of on-site equipment and outputs a comprehensive score after deployment; if the comprehensive score after deployment is lower than a preset threshold, the optimization decision unit and the execution unit are invoked to optimize the deployment and re-evaluate the camouflage effect; the specific implementation method has been clearly stated in the foregoing content and will not be repeated here.

[0100] To further illustrate this plan, further examples will be provided.

[0101] Example 1: Tank camouflage in an autumn forest environment Environmental context: Autumn forest environment at a training ground; target is a certain type of tank (surface covered with multispectral camouflage netting, high infrared emissivity, and high radar reflectivity). Environmental characteristics include light green tree canopies, dappled sunlight in the forest, and a layer of fallen leaves covering the ground. The background infrared emissivity is approximately 0.3, and the radar reflectivity is approximately 18 dB.

[0102] Multimodal reconnaissance and diagnosis are performed using a multimodal reconnaissance unit, a data preprocessing unit, and a diagnostic unit. Specifically, the UAV takes off along a preset reconnaissance route (flight altitude 100m, incident angle 30°, heading aligned with the direction of illumination), simultaneously collecting visible light image data, infrared image data, and radar image data of the target and its surrounding environment to form multispectral data, and simultaneously recording environmental parameters (illuminance 4500 lux, temperature 18℃, humidity 65%). After time synchronization (accuracy ≤8ms), spatial registration (error ≤0.1m), and noise removal, the data generates a fused environmental feature map, which is input into the diagnostic unit to complete defect diagnosis. The output results are shown in Table 1 below.

[0103] Table 1

[0104] The diagnostic unit also outputs a heat map showing the location of the camouflaged defects, highlighting the top armor plate, engine compartment, and armor plate as the main defect areas, providing precise location for subsequent optimization.

[0105] The optimal camouflage scheme is generated based on the optimized decision-making unit. Specifically, based on the diagnostic results, the "Autumn Forest Environment Camouflage Case Library" in the preset expert knowledge base is invoked, and combined with the Proximal Policy Optimization (PPO) algorithm, three sets of camouflage optimization schemes are generated. These schemes are then evaluated using a multi-objective evaluation model (weighted...). , , Calculate the overall score to obtain the camouflage effect enhancement. As shown in Table 2.

[0106] Table 2

[0107] Based on comprehensive evaluation, Option A was selected as the optimal option. This option performed best in terms of camouflage effect improvement, equipment cost, and operation time. It also made full use of readily available materials (tree branches and fallen leaves), resulting in low cost and good naturalness.

[0108] Based on the execution unit, on-site equipment is deployed according to the optimal camouflage scheme. Specifically, taking into account the characteristics of the forest environment with abundant vegetation and soft ground, a hybrid operation mode of "manual operation as the main method and mechanical operation as the auxiliary method" is adopted. The operators cooperate with the execution module to complete the deployment according to the following process. The deployment method is shown in Table 3.

[0109] Table 3

[0110] During operation, the drone hovers to provide real-time positioning guidance from a top-down perspective, while the handheld terminal displays a heat map of the target area and the expected coverage range, assisting manual operations in precise control.

[0111] During deployment, each operational module provided real-time feedback on its status via the uplink interface: the drone's leaf delivery drop point deviation was ≤5cm, meeting the tolerance requirements; the radar absorbing material was laid with the required flatness; and all manual operations were completed according to the instructions without any abnormalities. The entire deployment took 45 minutes, less than the preset 60-minute time limit.

[0112] Based on the verification unit, the verification process is carried out. Specifically, the UAV takes off again to perform an independent reconnaissance mission (flight altitude 100m, incident angle 30°). By calling the multimodal reconnaissance unit, data preprocessing unit and diagnostic unit, the camouflage effect after the deployment of on-site equipment is re-evaluated, and the secondary evaluation results are output. The evaluation results are shown in Table 4.

[0113] Table 4

[0114] The overall score output by the diagnostic unit is 92.5 points (based on the overall score). (The result is obtained by converting ×100 to a percentage). If the target percentage (≥90%) is reached, convergence condition 1 is met, and the camouflage operation is considered successful. The system automatically outputs a final camouflage effect report, storing the diagnostic results, optimization plan, deployment details, and evaluation data of this operation into the expert knowledge base for incremental training of subsequent reinforcement learning models, providing a reference case for tank camouflage in similar autumn forest environments.

[0115] Example 2: Armored vehicle camouflage in desert environments Environmental Background: Summer desert environment at a desert training ground. The target is an 8x8 wheeled armored vehicle (no standard camouflage, only basic desert yellow paint, rapid infrared heat dissipation, and obvious radar reflectivity). Environmental characteristics include strong sunlight, large diurnal temperature range, and frequent sandstorms. The background consists mainly of yellow sand and low shrubs. The background infrared emissivity is approximately 0.35, and the radar reflectivity is approximately 19 dB.

[0116] Multimodal reconnaissance and diagnosis are performed using a multimodal reconnaissance unit, a data preprocessing unit, and a diagnostic unit. Specifically, the UAV takes off along a preset reconnaissance route (flight altitude 80m, incident angle 35°, heading aligned with the direction of illumination), simultaneously collecting visible light image data, infrared image data, and radar image data of the target and the surrounding desert environment to form multispectral data. Environmental parameters (illuminance 8000 lux, air temperature 38℃, wind speed 2.5m / s) are recorded simultaneously. After the data is preprocessed by the preprocessing module to complete time synchronization (accuracy 8ms), spatial registration (error ≤3cm), and noise removal, a fused environmental feature map is generated. This map is then input into the diagnostic unit to complete defect diagnosis, and the output results are shown in Table 5 below.

[0117] Table 5

[0118] The diagnostic unit also outputs a heat map of camouflage defects, highlighting the main defect areas on the entire vehicle surface, tires, wheels, and roof weapon station, providing precise location for subsequent optimization.

[0119] The optimal camouflage scheme is generated based on the optimized decision-making unit. Specifically, based on the diagnostic results, the "desert environment camouflage case library" in the expert knowledge base is invoked, and three candidate schemes are generated by combining the proximal strategy optimization (PPO) algorithm. These schemes are then evaluated using a multi-objective evaluation model (weighted...). , , Calculate the overall score to obtain the camouflage effect enhancement. As shown in Table 6.

[0120] Table 6

[0121] Based on comprehensive evaluation, Option A was selected as the optimal option. This option significantly outperforms other options in terms of camouflage effect, makes full use of readily available materials (fine yellow sand), and keeps costs under control.

[0122] Based on the execution unit, on-site equipment is deployed according to the optimal camouflage scheme. Specifically, considering the characteristics of the desert environment, such as high temperature and sandstorms, a hybrid operation mode of "manual operation as the main method and mechanical operation as the auxiliary method" is adopted. The operators cooperate with the execution module to complete the deployment according to the following process. The deployment method is shown in Table 7.

[0123] Table 7

[0124] During operation, the drone hovers to provide real-time positioning guidance from a top-down perspective, while the handheld terminal displays a heat map of the target area and the expected coverage range, assisting manual refinement.

[0125] During deployment, each execution unit provided real-time feedback on the operation status through the uplink interface: the spraying deviation was ≤2cm, the radar absorber was firmly installed, and the sand transportation and rough laying were completed; the entire operation took 55 minutes, which was less than the preset 70-minute time limit, and the heavy sand transportation was completed with mechanical assistance, which effectively reduced the labor intensity of the workers in the high-temperature environment.

[0126] The verification process is based on the verification unit. Specifically, the UAV takes off again to avoid the direct sunlight period (choosing 3 pm, with a solar altitude angle of 45°). By calling the multimodal reconnaissance unit, data preprocessing unit, and diagnostic unit, the camouflage effect after the deployment of on-site equipment is re-evaluated, and the secondary evaluation results are output. The evaluation results are shown in Table 8.

[0127] Table 8

[0128] The comprehensive score output by the diagnostic unit was 93.1 points, reaching the preset threshold (≥90%), satisfying convergence condition 1, and the camouflage operation was successful. The system automatically recorded all data from this operation, updated the desert environment camouflage cases in the expert knowledge base, provided standardized references for subsequent camouflage operations of similar wheeled armored vehicles in desert and arid environments, and provided real-world scene samples for incremental training of the reinforcement learning model.

[0129] Example 3: Armored vehicle camouflage in urban environments Environmental context: A training ground in the suburbs of a city. The target is an infantry fighting vehicle (its surface is painted with urban camouflage, but some areas deviate from the background). The environment is characterized by gray-white building walls, dark asphalt roads, and interspersed green belts and vegetation. The background is complex and varied, with uneven distribution of infrared and radar signatures.

[0130] Multimodal reconnaissance and diagnosis are performed using a multimodal reconnaissance unit, a data preprocessing unit, and a diagnostic unit. Specifically, the UAV is launched for reconnaissance (flying at an altitude of 80m), simultaneously collecting visible light image data, infrared image data, and radar image data of the target and the surrounding urban environment to form multispectral data. After the data is preprocessed by the module to complete time synchronization, spatial registration, and noise removal, a fused environmental feature map is generated. This map is then input into the diagnostic unit to complete defect diagnosis, and the output results are shown in Table 9 below.

[0131] Table 9

[0132] Based on the aforementioned comprehensive camouflage effect scoring formula (weighted by four dimensions: color, texture, infrared, and radar), the current camouflage effect score is calculated to be 72 points (the preset threshold for meeting the standard is 90 points), which is determined to be unqualified.

[0133] The optimal camouflage scheme is generated based on the optimized decision-making unit. Specifically, based on the diagnostic results, the "Urban Environment Camouflage Case Library" in the expert knowledge base is invoked, and three candidate schemes are generated by combining the Proximal Policy Optimization (PPO) algorithm. A multi-objective evaluation model (weight: , , Calculate the overall score to obtain the camouflage effect enhancement. As shown in Table 10.

[0134] Table 10

[0135] Based on comprehensive evaluation, Option A was selected as the optimal option, which is as follows: spraying a light gray temporary paint matching the building wall on the side armor area, covering the engine compartment with an infrared shielding blanket, and installing radar absorbing material on the turret.

[0136] Based on the execution unit, on-site equipment is deployed according to the optimal camouflage scheme. Specifically, considering the characteristics of dense buildings and limited space in urban environments, a hybrid operation mode of "manual operation as the main method and mechanical operation as the auxiliary method" is adopted. The operators cooperate with the execution module to complete the deployment according to the following process. The deployment method is shown in Table 11.

[0137] Table 11

[0138] During operation, the drone hovers to provide real-time positioning guidance from a top-down perspective, while the handheld terminal displays a heat map of the target area and the expected coverage range, assisting manual operations in precise control.

[0139] After deployment, the verification process is performed based on the verification unit. By calling the multimodal reconnaissance unit, data preprocessing unit and diagnostic unit, the camouflage effect of the equipment after deployment is re-evaluated and the secondary evaluation result is output. The comprehensive score is 86 points (still below the 90% threshold, but 14 points higher than the first time).

[0140] Based on the verification unit's call to the optimization decision unit, the remaining gap between the current camouflage's quantitative indicators and location heatmap and the environmental background is analyzed (see Table 12). Table 12

[0141] The optimization decision unit further generates new camouflage optimization schemes to supplement the previous schemes. Specifically, it adds a gradient transition to the edge of the side armor spraying area, lays additional radar absorbing materials, and uses drones to assist in dropping green vegetation bags to cover the exposed area on the top of the turret, simulating the effect of urban green belt vegetation.

[0142] Therefore, the execution unit is invoked to perform secondary field equipment deployment, and the deployment method is shown in Table 13.

[0143] Table 13

[0144] During operation, the drone hovers and provides real-time positioning guidance to assist manual edge reshaping.

[0145] The secondary verification process is performed based on the verification unit. Specifically, the UAV takes off again to perform an independent reconnaissance mission. By calling the multimodal reconnaissance unit, data preprocessing unit, and diagnostic unit, the camouflage effect after the deployment of on-site equipment is re-evaluated, and the re-evaluation results are output. The evaluation results are shown in Table 14.

[0146] Table 14

[0147] The comprehensive score output by the diagnostic unit is 91 points, which reaches the preset threshold (≥90%), meets convergence condition 1, and the camouflage operation is qualified.

[0148] The results of the two closed-loop optimization verification processes for the camouflage effect in this scheme can be summarized in Table 15.

[0149] Table 15

[0150] Therefore, it can be seen that during this urban environment armored vehicle camouflage process, the comprehensive score improved from 72 to 91 points through two iterations of optimization, verifying the effectiveness of the closed-loop iterative optimization mechanism of this invention and fully demonstrating the excellent dynamic adaptability and iterative optimization advantages of this solution. The two iterations took a total of 95 minutes, which is more than 50% more efficient than the traditional manual method of "repeated guessing and multiple adjustments".

[0151] Compared with the previous two embodiments, the urban environment is complex and diverse, and a single camouflage scheme is difficult to meet the requirements of multi-spectral stealth. However, this solution uses a closed-loop iterative process of "diagnosis-optimization-deployment-evaluation-re-diagnosis" to accurately capture every camouflage defect and gradually optimize and supplement the scheme. This not only ensures that the camouflage effect meets the standards, but also avoids the waste of equipment and operational redundancy caused by excessive camouflage.

[0152] After deployment, the system automatically stores the entire process data of this iteration optimization (including the results of two diagnostics, optimization schemes, deployment details, evaluation indicators, and iteration improvement data) into the expert knowledge base. It focuses on supplementing the case parameters of "multi-round iterative camouflage" in urban environments, providing a replicable and scalable iterative optimization template for subsequent camouflage operations of similar infantry fighting vehicles and armored vehicles in complex environments such as cities and suburbs. At the same time, it provides rich complex scenario samples for incremental training of reinforcement learning PPO algorithm, further improving the system's adaptability to complex environments and the efficiency of camouflage scheme generation.

[0153] The above description is merely an example of a specific solution of the present invention. For any devices and structures not described in detail herein, it should be understood that they are implemented using common devices and methods already available in the art.

[0154] The above description is merely one embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A target camouflage system based on multimodal reconnaissance and closed-loop optimization, characterized in that, include: The multimodal reconnaissance unit, mounted on an airborne platform, is used to simultaneously acquire multispectral data of the target and the environmental background; the multimodal reconnaissance unit includes: a visible light camera, an infrared thermal imager, and radar; The data preprocessing unit is used to fuse multi-spectral data to generate environmental feature maps; The diagnostic unit analyzes the environmental feature map and the target's current state data based on a pre-trained conditional generative adversarial network model, and outputs a quantitative index representing the camouflage defects and a heat map of the location of the camouflage defects. The optimization decision unit is used to generate multiple camouflage optimization schemes based on quantitative indicators and location heatmaps, combined with a pre-set expert knowledge base and reinforcement learning algorithm, and select the optimal camouflage scheme through a multi-objective evaluation model. The execution unit is used to deploy on-site equipment according to the optimal camouflage plan; The verification unit is used to call the multimodal reconnaissance unit, data preprocessing unit, and diagnostic unit to re-evaluate the camouflage effect after the deployment of on-site equipment and output a comprehensive score after deployment. If the comprehensive score after deployment is lower than the preset threshold, the optimization decision unit and execution unit are called to perform iterative optimization until the comprehensive score reaches the preset threshold or the preset iteration limit is reached.

2. The target camouflage system based on multimodal reconnaissance and closed-loop optimization according to claim 1, characterized in that, The data preprocessing unit performs time synchronization, spatial registration, and noise removal on the multi-spectral data to generate a fused environmental feature map.

3. The target camouflage system based on multimodal reconnaissance and closed-loop optimization according to claim 2, characterized in that, The pre-trained conditional generative adversarial network model in the diagnostic unit includes a generator and a discriminator; The generator adopts a U-Net structure to map environmental feature maps and target current state data into virtual target images, and generates camouflage defect heatmaps based on intermediate layer features; The discriminator adopts the PatchGAN structure to judge the authenticity of the input image and outputs a quantitative index of the camouflage defects based on the intermediate layer feature regression; wherein, the input image is a real environment image or a virtual target image.

4. The target camouflage system based on multimodal reconnaissance and closed-loop optimization according to claim 3, characterized in that, The pre-trained conditional generative adversarial network model in the diagnostic unit is obtained by training based on the following steps: A simulation environment is built based on a physics simulation engine, and an atmospheric radiation transfer model and radar scattering simulation module are integrated to generate a simulation training set of multimodal data. The conditional generative adversarial network is trained based on the simulation training set to generate a simulation pre-trained model. The conditional generative adversarial network model is obtained by fine-tuning the simulation pre-trained model using transfer learning technology and real multimodal data of a preset scale.

5. The target camouflage system based on multimodal reconnaissance and closed-loop optimization according to claim 4, characterized in that, The conditional generative adversarial network model is trained using a joint optimization approach of adversarial loss and feature matching loss, and the loss function used is: in, Represents the total loss function. Indicating resistance to loss, The feature matching loss is calculated based on the pre-trained ResNet-50 feature extractor. This represents the regularization coefficient that balances the adversarial loss and the feature matching loss. Represents the distribution of real-world environment images The mathematical expectation, This indicates that the discriminator recognizes the real-world image. The output probability, Represents the distribution of random noise. The mathematical expectation, This indicates that the discriminator recognizes the virtual target image. The output probability, This indicates the generator in the environmental feature map. Under the condition of random noise The generated virtual target image, Represents a random noise vector. Represents an environmental feature map. This represents the pre-trained ResNet-50 feature extractor.

6. The target camouflage system based on multimodal reconnaissance and closed-loop optimization according to claim 5, characterized in that, In the optimization decision unit, the reinforcement learning algorithm adopts the deep deterministic policy gradient algorithm or the near-end policy optimization algorithm. In the process of generating multiple camouflage optimization schemes, the quantitative index representing the camouflage defect, the target's current state data, and the environmental feature map are constructed in vector form to form the state space in the reinforcement learning process. The action space in the reinforcement learning process is constructed based on the type, quantity, deployment location coordinates, and deployment order of the equipment to be deployed; A reward function is constructed based on factors such as the degree of camouflage effect improvement, equipment cost, operation time, and penalty for plan failure. Using a pre-set expert knowledge base as a foundation, the reinforcement learning process is provided with initial strategies and pre-set samples for experience replay pools, thereby generating multiple sets of camouflage optimization schemes.

7. The target camouflage system based on multimodal reconnaissance and closed-loop optimization according to claim 6, characterized in that, The multi-objective evaluation model comprehensively scores multiple camouflage optimization schemes, and selects the optimal camouflage scheme based on the best comprehensive score; the comprehensive score formula is expressed as: in, This indicates an improvement in camouflage effectiveness. Indicates equipment cost, The normalized baseline value representing the cost of equipment. Indicates the operation time. The normalized baseline value representing the operation time. , , Indicates the preset weight, and , .

8. The target camouflage system based on multimodal reconnaissance and closed-loop optimization according to claim 7, characterized in that, The execution unit includes: a bidirectional communication interface module for communicating with the optimization decision unit, and a job module for communicating with the bidirectional communication interface module; The bidirectional communication interface module includes: a downlink port and an uplink port; The downlink port is used to receive the optimal camouflage scheme data generated by the optimization decision unit based on the optimal camouflage scheme and transmit it to the operation module; wherein, the optimal camouflage scheme data includes: task sequence, target location coordinates of each sub-task, equipment type and quantity, operation time window, accuracy requirements and tolerance range; The uplink port is used to receive the execution status data fed back by the operation module and feed it back to the optimization decision unit; wherein, the execution status data includes: the completion status of each sub-task, the deviation between the actual deployment location of the equipment and the target location, the operation time, and abnormal situation reports.

9. The target camouflage system based on multimodal reconnaissance and closed-loop optimization according to claim 1, characterized in that, The pre-set expert knowledge base includes: a typical environmental feature library, a camouflage equipment characteristic library, a successful case library, and a camouflage standard library; The typical environment feature library is used to store the feature parameters of typical backgrounds in different environments, wherein the feature parameters include at least one of optical feature parameters, infrared feature parameters, and radar feature parameters. The camouflage equipment characteristic library is used to store the physical characteristic parameters of standard camouflage equipment and convenient equipment. The success case library is used to store successful solutions for historical camouflage operations; The camouflage standard library is used to store the specification parameters defined in the selected camouflage standards.

10. A camouflage method for a target camouflage system based on multimodal reconnaissance and closed-loop optimization as described in any one of claims 1 to 9, characterized in that, Includes the following steps: S1. Simultaneously acquire multi-spectral data of the target and environmental background, including: visible light image data, infrared image data, and radar image data; S2. Fuse multi-spectral data to generate an environmental feature map; S3. Based on the pre-trained conditional generative adversarial network model, analyze the environmental feature map and the target's current state data, and output quantitative indicators representing camouflage defects and a heat map of the location of camouflage defects; S4. Based on quantitative indicators and location heatmaps, and combined with a pre-set expert knowledge base and reinforcement learning algorithm, multiple camouflage optimization schemes are generated, and the optimal camouflage scheme is selected through a multi-objective evaluation model. S5. Deploy on-site equipment according to the optimal camouflage plan; S6. Re-evaluate the camouflage effect after the on-site equipment deployment and output the overall score after deployment; if the overall score after deployment is lower than the preset threshold, repeat steps S4 to S5 and re-evaluate the camouflage effect, and iterate and optimize accordingly until the preset conditions or the preset iteration limit is reached.

Images