Target physical stealth methods, devices, and media based on dynamic adaptive projection

Abstract

This invention discloses a target physical camouflage method, device, and medium based on dynamic adaptive projection. The method includes: acquiring an image dataset and training a target detection model based on the image dataset; acquiring a first image sample of the target object; projecting a random perturbation onto the target object and acquiring a second image sample of the target object after adding the random perturbation; inputting the first image sample and the second image sample into the trained target detection model, iteratively optimizing the random perturbation by minimizing a second joint loss function, which is constructed based on an alignment loss function and a camouflage loss function, and projecting the optimized perturbation as a camouflage texture onto the target object in real time until the target detection model can no longer correctly detect the target object.

Description

Technical Field

[0001] This invention belongs to the field of machine learning technology, and in particular relates to a method, device, and medium for target physical cloaking based on dynamic adaptive projection. Background Technology

[0002] In recent years, artificial intelligence technology has developed rapidly, especially deep neural networks, which have made breakthroughs in the field of computer vision due to their outstanding capabilities in feature extraction and high-dimensional data processing. Because of their excellent performance in target recognition tasks such as object detection, image classification, and object tracking, deep learning models are being rapidly and extensively adopted and deployed in key areas such as modern surveillance and security, autonomous driving, and smart cities.

[0003] However, as research deepens, academia and industry have revealed a critical issue concerning system security: deep neural networks, which heavily rely on data-driven approaches, suffer from severe physical stealth / adversarial camouflage in target recognition tasks. This means that deep neural networks exhibit significant vulnerability to specially designed adversarial perturbations, failing to correctly detect or identify camouflaged targets. Attackers can add specific, seemingly harmless camouflage textures to the surface of target objects, causing deep neural networks to severely skew feature extraction and ultimately rendering the target detector ineffective.

[0004] In real-world physical applications, this flaw presents extremely serious hidden dangers. It means that the detection capabilities of modern surveillance systems built on deep learning are highly susceptible to being compromised by physical stealth technology. If maliciously exploited, unidentified objects, unauthorized vehicles, or dangerous individuals could easily become "invisible" within the field of view of surveillance cameras, silently penetrating existing physical and digital security networks, posing a substantial and severe threat to public safety, personal safety, and the security of national and corporate property.

[0005] In the fields of cybersecurity and AI security, proactively exposing and studying system vulnerabilities is the first step in building a robust defense system. Therefore, developing a stable and reliable physical stealth technology is a crucial prerequisite for comprehensively evaluating the robustness of existing AI vision models and discovering potential algorithmic vulnerabilities in the surveillance and security field. By simulating real-world physical stealth attacks, high-value adversarial examples can be provided for deep learning models, which can then be used to compensate for potential deficiencies in deep learning models through adversarial training and other defensive measures.

[0006] Existing physical cloaking techniques typically follow four steps: (1) creating cloaking textures in digital space; (2) physically shaping the cloaking textures, such as laser printing or 3D printing; (3) having the camera recapture the target object; and (4) inputting the recaptured image into the target detector to trick it into outputting incorrect results. However, most physical cloaking techniques currently face numerous bottlenecks in practical applications: First, step (1) lacks consideration for dynamic changes, such as target rotation and displacement, and changes in ambient light, which makes the cloaking effect prone to failure in dynamic scenes. Second, in step (2), existing physical shaping methods are highly invasive, difficult to modify, and costly to implement. Finally, in step (3), due to the inevitable influence of transformations in physical space during image acquisition, such as changes in ambient light intensity and camera angle shifts, the recaptured image loses its original digital domain cloaking properties, which severely limits the stability and applicability of physical cloaking techniques.

[0007] In summary, there is a lack of a low-cost, non-invasive, and dynamically adaptable target physical stealth technology in the existing technology to overcome the technological gap caused by the aforementioned spatial transformation. Summary of the Invention

[0008] To address the shortcomings of existing technologies, embodiments of the present invention provide a method, device, and medium for target physical stealth based on dynamic adaptive projection.

[0009] In a first aspect, embodiments of the present invention provide a target physical stealth method based on dynamic adaptive projection, the method comprising: Obtain an image dataset and train an object detection model based on the image dataset; Acquire a first image sample of the target object; project a random perturbation onto the target object and acquire a second image sample of the target object after adding the random perturbation; The first image sample and the second image sample are input into the trained target detection model. By minimizing the second joint loss function, which is constructed based on the alignment loss function and the camouflage loss function, the random perturbation is iteratively optimized. The optimized perturbation is then projected as a camouflage texture onto the target object in real time until the target detection model can no longer detect the target object correctly.

[0010] In a second aspect, embodiments of the present invention provide an electronic device, comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores one or more computer programs that can be executed by the at least one processor, and the one or more computer programs are executed by the at least one processor to enable the at least one processor to perform the above-described target physical stealth method based on dynamic adaptive projection.

[0011] Thirdly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described target physical stealth method based on dynamic adaptive projection.

[0012] Fourthly, embodiments of the present invention provide a computer program product, including a computer program / instruction, which, when executed by a processor, implements the above-described target physical stealth method based on dynamic adaptive projection.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a target physical camouflage method based on dynamic adaptive projection. Through a continuous cycle of real-time camera acquisition, joint loss iteration, and real-time projection, the generated camouflage texture is no longer a static, fixed pattern, but a dynamic optical flow capable of millisecond-level adaptive reconstruction based on changes in the target's own state and the external environment. This technique completely breaks through the spatial and media limitations of traditional physical countermeasures, achieving an unexpected technical effect of transitioning from static countermeasures to dynamic adaptability. It achieves highly robust camouflage in the physical domain at extremely low cost, enabling the discovery of potential defects in more deep learning models in the surveillance and security field.

[0014] Meanwhile, after inputting the first and second image samples into the target detection model, this invention iteratively optimizes the model by combining a second joint loss function that includes alignment loss and cloaking loss. In this closed loop, the alignment loss function (covering feature matching, geometric alignment, and color consistency) forcibly bridges the feature gap between the perturbations generated in the digital domain and the physically captured images by the camera on a macroscopic scale. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a target physical stealth method based on dynamic adaptive projection provided in an embodiment of the present invention; Figure 2A flowchart illustrating the target physical stealth method based on dynamic adaptive projection provided in this embodiment of the invention; Figure 3 A schematic diagram illustrating an application scenario of the target physical stealth method based on dynamic adaptive projection provided in this embodiment of the invention; Figure 4 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] It should be noted that, unless otherwise specified, the features in the following embodiments and implementation methods can be combined with each other.

[0019] like Figure 1 and Figure 2 As shown, this embodiment of the invention provides a target physical stealth method based on dynamic adaptive projection, the method comprising: Step S1: Obtain the image dataset and train the target detection model based on the image dataset.

[0020] Specifically, step S1 includes the following steps: Step S101: Obtain the image dataset.

[0021] For example, the image dataset can be the COCO (Common Objects in Context) dataset, the PASCAL VOC dataset, or the ImageNet dataset.

[0022] Furthermore, in this example, the image dataset is denoted as... Its label dataset ,in, , and These are the predicted center coordinates of the bounding box. and It predicts the length and width of the bounding box. It is target classification information.

[0023] Step S102: Convert the image dataset from the RGB color space to the LAB color space.

[0024] Specifically, the image dataset Convert from RGB color space to LAB color space The expression is as follows: in, in, .

[0025] It's important to note that the RGB (Red, Green, Blue) color space is a hardware-oriented color model and is not visually uniform. Mathematical distances calculated in the RGB space (such as Euclidean distance) do not accurately represent the visual color differences captured by humans or camera lenses. The LAB color model is designed to approximate human visual perception. In the LAB space, color continuity is better, and adjacent color values ​​appear visually closer. This means that when processing images, the smooth transitions in color values ​​accurately reflect gradations in the physical world, facilitating more precise calculations of differences in image content.

[0026] In physical cloaking scenarios, dynamic changes in ambient lighting are one of the core pain points leading to cloaking failure. The three RGB channels (red, green, and blue) are highly coupled with luminance information; a change in lighting causes drastic fluctuations in the values ​​of all three channels simultaneously. The LAB color space completely decouples luminance information (L channel) from color channels (A and B channels). Through this conversion, the target model can focus more on the essential color characteristics of the target, thus exhibiting stronger stability and robustness when dealing with uncontrollable changes in lighting intensity in physical space.

[0027] This invention relates to generating camouflage textures by jointly minimizing alignment loss and camouflage loss. This process is highly dependent on backpropagation of gradients. Because the LAB color space has better color continuity, the loss surface encountered by the optimization algorithm when searching for the optimal camouflage perturbation parameters is relatively smoother. This not only helps the algorithm converge more efficiently but also ensures that the generated camouflage projection rays appear smoother and more natural on the physical surface, reducing abrupt high-frequency noise and thus improving the success rate of physical camouflage shaping. Therefore, converting image datasets from RGB color space to LAB color space is not merely simple data preprocessing but lays the underlying data foundation for overcoming illumination and color distortions caused by physical space conversion and ensuring stable operation of closed-loop real-time optimization.

[0028] Step S103: Train the target detection model based on the first joint loss function using the image dataset in the LAB color space; wherein, the first joint loss function is a weighted sum of the position regression loss function, the cross-entropy loss function, and the tracking loss function.

[0029] Furthermore, in this example, the target detection model uses a deep convolutional neural network as a feature module and constructs two output branches for target detection and target tracking.

[0030] Object detection includes two parts: regression and classification.

[0031] For the regression part, the expression for the location regression loss function is as follows: In the formula, It is the position of the ground truth bounding box corresponding to the i-th sample. It is the predicted detection box position corresponding to the i-th sample. and Represents the weighting coefficient, ( ) represent the center coordinates, width, and height of the detection box, respectively.

[0032] For the classification part, the cross-entropy loss function is usually used, as shown in the following expression: In the formula, This represents the category label of the i-th sample. This represents the one-hot encoding of the true class label of the i-th sample. This represents the predicted class label of the i-th sample.

[0033] Furthermore, the expression for the tracking loss function is as follows: In the formula, This represents the i-th sample. The template image data corresponding to the i-th sample. This represents the predicted output of the object detection model corresponding to the i-th sample. This is the Siamese feature extraction subnetwork.

[0034] It should be noted that the purpose of the tracking loss function is to train the feature extraction subnetwork ψ so that it can extract features based on template image data. In the search area The generated predicted response map is made as close as possible to the true response map centered on the actual target location. By minimizing this loss, the feature extraction model learns how to accurately locate the initial target (template) in subsequent frames (search region).

[0035] Construct a first joint loss function, which is a weighted sum of the position regression loss function, the cross-entropy loss function, and the tracking loss function, as shown in the following expression: In the formula, , and These are weight parameters.

[0036] Step S2: Acquire a first image sample of the target object; project a random perturbation onto the target object and acquire a second image sample of the target object after adding the random perturbation.

[0037] Specifically, such as Figure 3 As shown, this example constructs a target physical cloaking device, which includes a camera, a projector, and a computer. The camera is responsible for capturing real-time image data of the target object, while the projector projects specific patterns or light (i.e., cloaking textures) onto the target object according to the computer's instructions. The computer acts as the brain, responsible for processing the image data captured by the camera, analyzing the characteristics of the target object, generating corresponding cloaking textures, and controlling the projector to execute these cloaking textures.

[0038] Step S3: Input the first image sample and the second image sample into the target detection model trained in step S1. Iteratively optimize the random perturbation by minimizing the second joint loss function. The optimized perturbation is projected as a camouflage texture onto the target object in real time through a projector until the target detection model can no longer detect the target object correctly. Specifically, a second joint loss function is constructed. The sum of the alignment loss function and the stealth loss function is expressed as follows: In the formula, Represents the alignment loss function. This represents the stealth loss function.

[0039] Furthermore, the process of constructing the alignment loss function includes: In the formula, Represents feature matching loss, This indicates the alignment error penalty term. Indicates loss of color consistency. , , All are weighting coefficients. This indicates a feature extraction operation. This represents the optical flow field or registration field. Indicates channel The average color value, This represents the image sample of the i-th frame. This represents the image sample of the (i+1)th frame. This represents a random perturbation, and a and b represent the coordinate indices of pixels in the image.

[0040] It should be noted that the core technical effect of the alignment loss function lies in eliminating the spatial transformation error caused by digital domain perturbation when projected into the physical domain, thereby achieving high-precision alignment between digital images and physically captured images.

[0041] Extracting digit field images using a pre-trained deep convolutional neural network. ) and Physically captured images ( By using high-dimensional feature maps and calculating the Euclidean distance or cosine similarity between them, pixel-level surface noise differences can be overcome, and the differences in image content can be quantified from a deep feature semantic level, ensuring that the image content after projection perturbation does not undergo severe visual distortion.

[0042] In real physical space, geometric deformation inevitably occurs due to the angle of the projection device, the camera's viewpoint, and the dynamic displacement of the target object. This loss term introduces optical flow field or image registration techniques to measure the pixel-level correspondence between two images, and uses L2 loss to penalize alignment errors, which can achieve accurate geometric alignment and effectively suppress spatial displacement distortions such as stretching and twisting that occur during physical projection.

[0043] Different capture devices and varying ambient lighting conditions can lead to uncontrollable color differences. By introducing color moments or color histograms for comparison, the mean color value of each channel can be calculated. By using the differences in color temperature and lighting conditions as constraints, color balance and lighting conditions can be dynamically adjusted to counteract color distortion caused by changes in the physical lighting environment.

[0044] By weighting and combining the three loss functions mentioned above, the alignment loss function can simultaneously constrain the three core physical domain variables—content, space, and color—during the perturbation optimization process. By minimizing the alignment loss function, the ideal digital domain image can be effectively aligned with the camera-captured image containing environmental noise, achieving high-quality image fusion and enhancement, thereby removing physical space transformation obstacles for the accurate projection of stealth perturbations.

[0045] Furthermore, the stealth loss function The expression is as follows: In the formula, These represent the weight coefficients corresponding to the cloaking loss function, and N represents the number of image samples. This represents the image sample of the i-th frame. This indicates a random perturbation. This represents the prediction result of the object detection model for the i-th frame image sample after adding random perturbations.

[0046] It should be noted that the stealth loss function directly disrupts the detection mechanism of the target detection model, allowing the perturbated physical target to become invisible within the target detection model. This stealth loss function explicitly hides the image captured by the hidden camera. For the purpose. In the formula This characterizes the confidence output of the target model for the detection box. During the perturbation optimization process, by... As a minimization penalty term, it can force the direction of gradient descent, making the probability of the object detection model recognizing the projected target approach zero, thereby achieving the direct technical effect of reducing the confidence of the predicted detection box.

[0047] The stealth loss function and the alignment loss function work together as a second joint loss function to construct a game and balance mechanism that seeks to maximize attack gains under physical constraints. In perturbation optimization, alignment loss aims to align the digital domain image with the camera-captured image. It filters out lighting variations, camera viewpoint shifts, and geometric deformations in physical space, ensuring that the generated digital perturbation can be realistically reproduced in the physical world. Stealth loss aims to directly reduce the confidence of the target model's detection box to hide the camera-captured image. With only stealth loss, the generated perturbation might be highly aggressive in the digital domain, but once projected onto a complex physical environment, it will instantly become ineffective due to lighting and deformation (i.e., overfitting to the digital domain). With only alignment loss, the system becomes merely a perfect image restoration tool, lacking any offensive capability. Minimizing both together ensures that when optimizing adversarial perturbations, the system must find the feature solution that best blinds the detector (stealth) while ensuring that the perturbation is physically reproducible by the projector (alignment).

[0048] Furthermore, in real-world physical adversarial attacks, the biggest technical barrier is the domain gap. The second joint loss function mandates that the direction of gradient descent must simultaneously satisfy both digital adversarialism and physical consistency, effectively eliminating the discrepancy between the ideal perturbation calculated digitally and the actual physical projection effect. This ensures that the generated camouflage texture is no longer fragile digital code, but rather a robust physical light spot capable of resisting interference from the real-world physical environment and being accurately projected onto the target surface by the projector.

[0049] In the dynamic cycle of camera capture, joint loss calculation, and projection optimization of perturbations, the alignment loss continuously provides negative feedback from the current ambient lighting and physical displacement to correct deviations, while the stealth loss continuously provides positive traction to evaluate the current stealth gains. This deep synergy enables stealth technology to possess dynamic adaptability, making perturbation optimization a dynamic optimization process capable of millisecond-level adaptive reconstruction as the target's own state and external environment change.

[0050] Meanwhile, in complex and ever-changing physical spaces (such as sudden target rotation or abrupt changes in ambient light), the optimization of adversarial attacks on a single target is highly susceptible to gradient explosion or divergence during backpropagation due to high-frequency environmental noise. The introduction of alignment loss terms (especially feature matching and color consistency) in the second joint loss function is mathematically equivalent to adding a physical prior regularization term to the optimization objective of the stealth attack. This significantly smooths the loss surface of the joint optimization, effectively suppresses the drastic gradient oscillations caused by environmental noise, and enables the system to converge quickly and stably and generate high-quality stealth textures even in harsh dynamic physical coupling environments.

[0051] Furthermore, in this example, the Sharpness-Aware Minimization (SAM) algorithm is used to minimize the perturbation during the optimization iteration process. The optimization process includes the following steps: For images captured by the camera , Independently and identically distributed Disturbance The optimization loss satisfies This invention minimizes... Find parameter values ​​that have a consistent low training loss value across the entire neighborhood (i.e., the neighborhood has both low loss and low curvature).

[0052] Considering The right-hand side of the inequality can be rewritten as: The items in square brackets are measured from... The speed at which the training loss is increased by moving to nearby parameter values ​​is used to calculate this. exist The sharpness at that point. Then, this sharpness term is multiplied by the training loss value itself and... The regularizers are added together. Therefore, this example selects the optimal value by solving the following sharpness-perceived minimization problem: in, hyperparameters .

[0053] To minimize This example derives the following by maximizing the internal derivative, i.e., stochastic gradient descent. Approximate to .

[0054] pass Approximating the inner maximization problem with a first-order Taylor expansion near 0, we obtain: It should be noted that existing physical cloaking technologies, when facing the real world, are prone to failure in the physical domain due to the influence of transformations in the physical space during camera acquisition (such as changes in light intensity and camera perspective changes caused by target movement). This invention acquires a second image sample with random perturbations in real time and compares it with the first image sample for closed-loop calculation. Relying on the strong constraints of the aforementioned multidimensional alignment loss and SAM optimization, high-frequency noise interference in the physical environment can be filtered out. This substantially solves the technical problem that when the geometric deformation of the target object is highly coupled with the complex and ever-changing dynamic lighting of the environment, the adversarial perturbation is prone to divergence and non-convergence during gradient backpropagation, ensuring high-quality fusion of the projection camouflage. In this example, the Sharpness-Aware Minimization (SAM) algorithm is used in the optimization iteration process to find a flat parameter neighborhood in the micro-parameter space that has both low training loss and low curvature, achieving a synergistic effect of macro-domain alignment and micro-flat minima finding.

[0055] In summary, this invention provides a target physical camouflage method based on dynamic adaptive projection. Traditional physical camouflage technologies often rely on solid shaping methods such as 3D printing and stickers, which suffer from drawbacks such as high invasiveness, high cost, and lack of consideration for dynamic changes (e.g., inability to adapt to target rotation and displacement). This example, through continuous iteration of real-time camera acquisition, joint loss iteration, and real-time projection, generates a camouflage texture that is no longer a static, fixed pattern, but a dynamic optical flow capable of millisecond-level adaptive reconstruction based on changes in the target's own state and the external environment. This technique completely breaks through the spatial and media limitations of traditional physical countermeasures, achieving an unexpected technical effect of transitioning from static countermeasures to dynamic adaptability, and realizing highly robust camouflage in the physical domain at extremely low cost.

[0056] Meanwhile, after inputting the first and second image samples into the target detection model, this invention iteratively optimizes the model by combining a second joint loss function that includes alignment loss and cloaking loss. In this closed loop, the alignment loss function (covering feature matching, geometric alignment, and color consistency) forcibly bridges the feature gap between the perturbations generated in the digital domain and the physically captured images by the camera on a macroscopic scale.

[0057] like Figure 4 As shown, this application provides an electronic device including a memory 101 for storing one or more programs and a processor 102. When the one or more programs are executed by the processor 102, they implement the method as described in any of the first aspects above.

[0058] The system also includes a communication interface 103. The memory 101, processor 102, and communication interface 103 are electrically connected directly or indirectly to each other to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more communication buses or signal lines. The memory 101 can be used to store software programs and modules, and the processor 102 executes various functional applications and data processing by executing the software programs and modules stored in the memory 101. The communication interface 103 can be used for signaling or data communication with other node devices.

[0059] The memory 101 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.

[0060] The processor 102 can be an integrated circuit chip with signal processing capabilities. The processor 102 can be a general-purpose processor 102, 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.

[0061] In the embodiments provided in this application, it should be understood that the disclosed methods and systems can also be implemented in other ways. The method and system embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0062] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0063] On the other hand, embodiments of this application provide a computer-readable storage medium storing a computer program thereon. When executed by processor 102, the computer program implements the methods described in any of the first aspects above. If the functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0064] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only.

[0065] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A target physical stealth method based on dynamic adaptive projection, characterized in that, The method includes: Obtain an image dataset and train an object detection model based on the image dataset; Acquire a first image sample of the target object; project a random perturbation onto the target object and acquire a second image sample of the target object after adding the random perturbation; The first image sample and the second image sample are input into the trained target detection model. By minimizing the second joint loss function, which is constructed based on the alignment loss function and the camouflage loss function, the random perturbation is iteratively optimized. The optimized perturbation is then projected as a camouflage texture onto the target object in real time until the target detection model can no longer detect the target object correctly.

2. The target physical stealth method based on dynamic adaptive projection according to claim 1, characterized in that, The process of training an object detection model based on the aforementioned image dataset includes: Obtain the image dataset; Convert the image dataset from the RGB color space to the LAB color space; An object detection model was trained using an image dataset in the LAB color space based on a first joint loss function.

3. The target physical stealth method based on dynamic adaptive projection according to claim 2, characterized in that, The first joint loss function is a weighted sum of the location regression loss function, the cross-entropy loss function, and the tracking loss function; The expression for the location regression loss function is as follows: ； In the formula, It is the position of the ground truth bounding box corresponding to the i-th sample. It is the predicted detection box position corresponding to the i-th sample. and Represents the weighting coefficient, ( These represent the center coordinates, width, and height of the detection box, respectively. The expression for the cross-entropy loss function is as follows: ； In the formula, This represents the category label of the i-th sample. This represents the one-hot encoding of the true class label of the i-th sample. This represents the predicted class label for the i-th sample; The expression for the tracking loss function is as follows: ； In the formula, This represents the i-th sample. The template image data corresponding to the i-th sample. This represents the predicted output of the object detection model corresponding to the i-th sample. This is the feature extraction subnetwork.

4. The target physical stealth method based on dynamic adaptive projection according to claim 1 or 2, characterized in that, The image datasets selected are the COCO dataset, the PASCAL VOC dataset, or the ImageNet dataset.

5. The target physical stealth method based on dynamic adaptive projection according to claim 1 or 2, characterized in that, The target detection model uses a deep convolutional neural network as a feature module and constructs two output branches for target detection and target tracking.

6. The target physical stealth method based on dynamic adaptive projection according to claim 1, characterized in that, The process of constructing the alignment loss function includes: ； ； ； ； In the formula, Represents feature matching loss, This indicates the alignment error penalty term. Indicates loss of color consistency. Represents the alignment loss function. , , All are weighting coefficients. This indicates a feature extraction operation. This represents the optical flow field or registration field. Indicates channel The average color value, This represents the image sample of the i-th frame. This represents the image sample of the (i+1)th frame. This represents a random perturbation, and a and b represent the coordinate indices of pixels in the image.

7. The target physical stealth method based on dynamic adaptive projection according to claim 1, characterized in that, The expression for calculating the stealth loss function is as follows: ； In the formula, This represents the weight coefficients corresponding to the stealth loss function. This represents the cloaking loss function, where N represents the number of image samples. This represents the image sample of the i-th frame. This indicates a random perturbation. This represents the prediction result of the object detection model for the i-th frame image sample after adding random perturbations.

8. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores one or more computer programs that can be executed by the at least one processor, the one or more of the computer programs being executed by the at least one processor to enable the at least one processor to perform the target physical stealth method based on dynamic adaptive projection as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the target physical stealth method based on dynamic adaptive projection as described in any one of claims 1-7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the target physical stealth method based on dynamic adaptive projection as described in any one of claims 1-7.

Images