A method and system for adversarial camouflage detection and cleanup based on zero-shot and texture-independent methods

By employing a zero-sample and texture-independent approach, and combining semantic response maps and texture anomaly maps with geometric constraints, adversarial camouflage detection and cleanup are achieved. This solves the problem of locating and repairing large-area camouflage in monocular depth estimation, thereby improving the system's security and stability.

CN122906667BActive Publication Date: 2026-11-17HARBIN INSTITUTE OF TECHNOLOGY (SHENZHEN) (INSTITUTE OF SCIENCE AND TECHNOLOGY INNOVATION HARBIN INSTITUTE OF TECHNOLOGY SHENZHEN)
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611383411.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-08
Publication Date
2026-11-17
Estimated Expiration
2046-09-08

AI Technical Summary

Technical Problem

Existing technologies struggle to accurately locate and remove anti-camouflage covering large areas of a vehicle's surface, especially in the field of monocular depth estimation. Direct repair may damage the vehicle's structure, and existing methods are ill-suited to unknown camouflage types and different models.

Method used

Semantic response maps and texture anomaly maps are extracted by a pre-trained target detector. Geometric constraints are used for camouflage localization and reconstruction. Diffusion-based texture reconstruction is used to preserve the vehicle structure. Multiple anomaly analysis methods and edge structure extraction are used for accurate localization and repair.

Benefits of technology

It improves the safety and stability of the monocular depth estimation system in complex environments, enabling it to adapt to various camouflage types and maintain the structural integrity of the vehicle.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122906667B_ABST
    Figure CN122906667B_ABST
Patent Text Reader

Abstract

This invention relates to a zero-shot and texture-independent adversarial camouflage detection and purification method, system, and medium. The method includes: target candidate region detection in a scene image; semantically aware adversarial camouflage localization; multi-domain texture anomaly localization; anomaly texture mask fusion; adversarial texture removal and connected component decomposition; geometric condition construction and diffusion-based texture reconstruction; and finally, boundary-aware fusion to obtain the purified image. This invention improves the security, stability, and generalization ability of monocular depth estimation systems in complex physical environments.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of deep learning target classification and detection technology, and in particular to an adversarial camouflage detection and purification method and system based on zero-shot and texture-independent methods. Background Technology

[0002] In recent years, physical domain adversarial camouflage has gradually become a significant security threat to autonomous driving perception systems, particularly interfering with monocular depth estimation modules. Compared to traditional imperceptible disturbances or localized 2D adversarial patches, adversarial camouflage typically covers most of the vehicle's visible surface and exhibits large-scale, dispersed, and discontinuous anomalous textures that vary with viewing angle, distance, lighting, and weather conditions. This induces problems in monocular depth estimation models, such as missing vehicle contours, local depth distortion, and misjudgment of target distances. However, existing adversarial defense methods mainly target small disturbances or locally compact adversarial patches, typically relying on low-level texture features, fixed thresholds, or large-area occlusion strategies, making it difficult to accurately locate distributed adversarial camouflage across the entire surface. Furthermore, directly removing or unconstrainedly repairing suspicious areas may damage vehicle edges, body contours, and local geometry, introducing new depth estimation errors. In addition, some existing methods rely on training with specific attack samples or target models, making it difficult to adapt to unknown camouflage types and different monocular depth estimation models.

[0003] Existing adversarial defense efforts primarily target imperceptible adversarial perturbations and localized two-dimensional adversarial patches. For imperceptible perturbations, existing methods include: adding diffusion noise to the input image and performing a back-diffusion process to remove the perturbation; and proposing the Adversarial Diffusion Bridge (ADBM) model to establish a recovery process between the adversarial sample and the normal image, thereby improving the stability of input cleanup. However, these methods mainly target adversarial perturbations with small amplitudes distributed across the entire image, making it difficult to directly handle physical adversarial camouflage covering large areas of a vehicle's surface. For localized adversarial patches, existing research has proposed the Segment and Complete method, which first segments suspected adversarial regions and then restores occluded content through image completion; utilizes local entropy differences to locate adversarial patches and employs autoencoders to reconstruct suspicious regions; analyzes the semantic independence and spatial heterogeneity between image regions to locate and remove anomalous regions without training for specific patches; improves the robustness of the target detector against patch-hiding attacks through multiple occlusions and consistency analysis of detection results; and combines zero-shot anomalous region localization with structure-guided image inpainting to detect and reconstruct adversarial patches. The aforementioned methods have achieved some success in defending against localized, compact adversarial patches, but they typically assume that anomalous regions have limited areas, concentrated spatial locations, or significant local statistical differences. In the field of monocular depth estimation, existing research on 2D adversarial patch attacks targeting monocular depth estimation networks has demonstrated that local textures can significantly interfere with depth prediction results; further, 3D2Fool has been proposed to map adversarial textures onto the vehicle's 3D surface, achieving physical domain 3D adversarial camouflage attacks targeting monocular depth estimation models. Compared to 2D patches, 3D adversarial camouflage typically covers most of the vehicle's visible surface and forms multiple scattered and discontinuous anomalous regions depending on the viewing angle, distance, lighting, and weather conditions, thus failing to satisfy the local compact region assumption of existing patch defense methods. Furthermore, monocular depth estimation is sensitive to vehicle contours, edges, and local geometry; directly occluding or unconstrained repair of suspicious areas easily disrupts depth-related structures and introduces new depth prediction errors. Therefore, existing imperceptible perturbation cleanup methods and local patch defense methods are difficult to directly apply to full-surface adversarial camouflage for monocular depth estimation, and accurate localization and geometry-preserving cleanup techniques for this type of attack still require further improvement. Summary of the Invention

[0004] This invention provides a zero-sample and texture-independent adversarial camouflage detection and purification method, system, and medium. It aims to construct an adversarial camouflage purification method that does not require obtaining attack priors or modifying the target model. It can accurately locate targets by jointly utilizing high-level semantic inconsistencies and multi-domain texture anomalies, and preserve the target structure through geometrically constrained texture reconstruction. This will improve the security, stability, and generalization ability of monocular depth estimation systems in complex physical environments.

[0005] According to a first aspect of the present disclosure, a method for adversarial spoofing detection and cleanup based on zero-sample and texture-independent methods is provided, comprising the following steps: The scene image is input into the pre-trained target detector. A set of target candidate boxes is obtained by passing a low confidence threshold. Based on the obtained target candidate boxes, the target local image is cropped from the scene image. For each target local image, a dense semantic response map is extracted using a pre-trained encoder. A binary localization mask is obtained by applying a percentile-based threshold to the semantic response map. Each binary localization mask is then projected back to the original scene image coordinate system and merged to form a global semantic localization mask. For each target local image, a texture anomaly score map of the target candidate box is calculated based on multiple complementary anomaly analysis methods. By applying a percentile-based threshold to the texture anomaly score map, a local binary texture anomaly mask is obtained. Each local binary texture anomaly mask is projected back to the original scene image coordinate system and merged to form a global texture anomaly localization mask. The global semantic localization mask is fused with the global texture anomaly localization mask to obtain the fused mask; The fusion mask is decomposed into connected components. For each component, its bounding box is extended by a padding margin, and the edge structure of the target local image is extracted. Based on the target local image, the components decomposed by the fusion mask, edge structure, positive and negative sample prompts, geometric conditions are constructed and diffusion-type texture reconstruction is performed to obtain the reconstructed target local image; The components of the target local image and the reconstructed target local image are smoothed separately and then fused to obtain the final purified image.

[0006] A further technical solution of the present invention is as follows: The specific expression of the binary positioning mask is: , , among which, This represents the operation of removing connected components whose area is less than a given threshold. It is an indicator function. Represents a semantic response graph. Indicates the spatial location within the target candidate box. Indicates the first Percentile operators.

[0007] A further technical solution of the present invention is: calculating the texture anomaly score map of the target candidate box based on multiple complementary anomaly analysis methods, specifically expressed as: ,in, This represents a set of anomaly analysis methods. Indicates control of the first The contribution of each anomaly analysis method, and , This indicates a normalization operation. This represents the quantization value of the anomaly analysis method. Represents the scene image Obtained target candidate boxes , This indicates the number of target candidate regions detected.

[0008] A further technical solution of the present invention is: the anomaly analysis method includes bilateral residuals, Laplace response, local standard deviation, and color saturation.

[0009] A further technical solution of the present invention is: fusing the global semantic positioning mask with the global texture anomaly positioning mask to obtain a fused mask. The specific expression is: ,in, This represents a global texture anomaly localization mask. This represents a global semantic location mask. This represents element-wise multiplication. This indicates the removal of small components.

[0010] A further technical solution of the present invention is to use Canny to extract the edge structure of the target local image.

[0011] A further technical solution of the present invention is as follows: The components of the target local image and the reconstructed target local image are smoothed separately and then fused, specifically expressed as: , ,in, Represents convolution. The standard deviation is Gaussian kernel, This represents the reconstructed local image of the target. Represents a local image of the target.

[0012] According to a second aspect of the present disclosure, a zero-sample and texture-independent adversarial camouflage detection and defacement system is provided, the system comprising: The target candidate region detection module is used to input the scene image into the pre-trained target detector, obtain a set of target candidate boxes through a low confidence threshold, and crop the target local image from the scene image based on the obtained target candidate boxes; The semantic-aware adversarial camouflage localization module is used to extract a dense semantic response map for each target local image using a pre-trained encoder, and obtain a binary localization mask by applying a percentile-based threshold to the semantic response map. The binary localization masks are then projected back to the original scene image coordinate system and merged to form a global semantic localization mask. The multi-domain texture anomaly localization module is used to calculate the texture anomaly score map of the target candidate box based on multiple complementary anomaly analysis methods for each target local image. By applying a percentile-based threshold to the texture anomaly score map, a local binary texture anomaly mask is obtained. Each local binary texture anomaly mask is projected back to the original scene image coordinate system and merged to form a global texture anomaly localization mask. The abnormal texture mask fusion module is used to fuse the global semantic positioning mask with the global texture abnormal positioning mask to obtain the fused mask; The adversarial texture removal and connected component decomposition module is used to decompose the fusion mask into connected components, extend the bounding box of each component by a padding margin, and extract the edge structure of the target local image. The geometric condition construction and diffusion texture reconstruction module is used to construct geometric conditions and perform diffusion texture reconstruction based on the target local image, the components decomposed by the fusion mask, the edge structure, and positive and negative example prompts to obtain the reconstructed target local image. The boundary-aware fusion module is used to smooth the components of the target local image and the reconstructed target local image separately and then fuse them to obtain the final purified image.

[0013] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described zero-sample and texture-independent anti-camouflage detection and cleanup method.

[0014] According to a fourth aspect of the present disclosure, a non-transitory computer-readable storage medium is provided, the storage medium storing computer instructions that, when executed by a processor, implement the steps of the above-described zero-sample and texture-independent anti-spoofing detection and cleanup method.

[0015] This invention provides a zero-sample and texture-independent adversarial camouflage detection and cleanup method, system, and medium. The method addresses the lack of adversarial camouflage detection and cleanup methods in existing technologies, as well as the poor detection effect of directly applying adversarial patch defense methods to adversarial camouflage. This invention constructs an adversarial camouflage cleanup method that does not require obtaining attack priors or modifying the target model. It can accurately locate targets by jointly utilizing high-level semantic inconsistencies and multi-domain texture anomaly information, and preserve the target structure through geometrically constrained texture reconstruction. This improves the security, stability, and generalization ability of monocular depth estimation systems in complex physical environments.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0018] Figure 1 This is a flowchart of the anti-camouflage detection and purification method based on zero-sample and texture-independent methods in this invention embodiment; Figure 2 This is a schematic diagram of the anti-camouflage detection and purification system based on zero samples and texture independence in an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present invention are shown in the drawings, not the entire structure.

[0020] Before discussing the exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. The process can be terminated when its operation is complete, but may also have additional steps not included in the figures. The process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0021] like Figure 1 As shown, the embodiment is an adversarial spoofing detection and cleanup method based on zero-shot and texture-independent methods, including the following steps: S1. Input the scene image into the pre-trained target detector, obtain a set of target candidate boxes by passing a low confidence threshold, and crop the target local image from the scene image based on the obtained target candidate boxes; S2. For each target local image, a dense semantic response map is extracted using a pre-trained encoder. A binary localization mask is obtained by applying a percentile-based threshold to the semantic response map. Each binary localization mask is projected back to the original scene image coordinate system and merged to form a global semantic localization mask. S3. For each target local image, calculate the texture anomaly score map of the target candidate box based on multiple complementary anomaly analysis methods, and obtain the local binary texture anomaly mask by applying a percentile-based threshold to the texture anomaly score map. Project each local binary texture anomaly mask back to the original scene image coordinate system and merge them to form a global texture anomaly localization mask. S4. Fuse the global semantic location mask with the global texture anomaly location mask to obtain the fused mask; S5. Decompose the fusion mask into connected components. For each component, extend its bounding box by a padding margin and extract the edge structure of the target local image. S6. Based on the target local image, geometric conditions are constructed and diffusion-type texture reconstruction is performed using components decomposed by the fusion mask, edge structure, positive samples and negative sample prompts to obtain the reconstructed target local image. S7. After smoothing the components of the target local image and the reconstructed target local image respectively, they are fused to obtain the final purified image.

[0022] The specific implementation process uses vehicles as the target detection object, including: In the specific implementation of S1, scene images Input pre-trained target detector Passing with a confidence threshold lower than that of conventional detection processes Obtain the set of target candidate boxes: , in, This represents the number of target candidate regions detected, the th Each vehicle candidate box is represented as In the formula, and They represent the first The coordinates of the top left and bottom right corners of the candidate boxes. 1 target candidate bounding box, cropped from the scene image to obtain the target local image:

[0023] in, This indicates an image cropping operation. The purpose of using a low confidence threshold is to reduce the probability of missed vehicle detections when the confidence level of target detection decreases due to adversarial camouflage, while simultaneously eliminating a large amount of background content using target candidate regions.

[0024] In the specific implementation of S2, for each detected vehicle area The pre-trained DINOv3 encoder is used. To extract a dense semantic response graph: , While adversarial camouflage may contain visually complex colors and patterns, natural vehicles possess coherent semantic features in contrast, while the local representations of adversarial camouflage often exhibit anomalous semantic responses. Therefore, by analyzing... By applying a percentile-based threshold, a binary localization mask is obtained: , , in express Internal spatial location, Indicates the first Percentile operators It is an indicator function, and This represents the operation of removing connected components with an area smaller than a given threshold. The resulting local binary mask... The coordinates are projected back to the original image coordinate system and merged to form a global semantic localization mask. .

[0025] In the specific implementation of S3, multi-domain texture anomaly detection locates adversarial camouflage by analyzing the abnormal appearance and texture distribution within each detected vehicle region. For the first... For each vehicle region, its corresponding texture anomaly score map can be uniformly represented as: , in This represents a set of anomaly analysis methods. Control the first The contribution of each anomaly analysis method, and In one specific embodiment, four complementary anomaly analysis methods were employed, including bilateral residuals, Laplace response, local standard deviation, and color saturation. Specifically, Calculate bilateral residuals to capture local appearance deviations. Extracting high-frequency intensity changes, Measure the local texture dispersion, and It can then identify abnormal color saturation. This indicates a normalization operation because the response graphs have different numerical ranges. The larger the value in the value range, the greater the deviation from the natural appearance and texture distribution of the target vehicle. Subsequently, a percentile-based threshold is applied to obtain a local binary texture anomaly mask: , , in express A spatial location within, Indicates the first Percentile operators. Finally, the resulting local mask. Projected back to the original image coordinate system and merged to form a global texture anomaly localization mask. .

[0026] In its implementation, S4 uses step S2 to capture semantic inconsistencies, while step S3 captures local texture anomalies. To preserve regions supported by both cues and reduce false positives, S4 employs intersection as the default fusion strategy. , in This represents element-wise multiplication. This indicates the removal of small components.

[0027] In the specific implementation of S5, the fused mask is decomposed into connected components. , Smaller noisy components were discarded, and at most, only a few were retained. Each mask component is processed individually, reducing memory consumption and preventing unrelated vehicle regions from interfering with each other during diffusion-based reconstruction. For the first component... Each component is used to expand its bounding box by a padding margin p to preserve sufficient contextual information. To prevent StableDiffusion from altering the vehicle shape, Canny is used to extract the edge structure of the local clipping regions.

[0028] Indicates the first i A local image of the target, the local image, camouflage mask, and edge conditions are all adjusted to the standard diffusion input resolution. .

[0029] In the specific implementation of S6, the square clipping region is reconstructed using a Stable Diffusion-based filling model under the dual constraints of a camouflage mask and a Canny edge map:

[0030] in, This represents the original local image to be cleaned. This represents the mask of the local image obtained in step S4. This represents the structural information diagram obtained in step S5. and These represent the text embeddings of positive and negative sample cue words, respectively. Positive cue words aim to generate clean, smooth, and factory-original vehicle body surfaces, while negative cue words suppress camouflage, stripes, and graffiti. Parameter set This includes the number of denoising steps, classifier-free guided scaling, ControlNet conditional scaling, inpainting intensity, and random seeds. During backdiffusion, the ControlNet branch injects multi-scale edge features into the denoising U-Net:

[0031] in It is the noisy latent variable at step t. It is a potential denoising U-Net. It's a ControlNet encoder. Boundary constraints require the generated content to follow the original vehicle outline, while the inpainting mask limits the reconstruction to the detected camouflage area. The generated result is then resized to restore the original cropping resolution. In the formula ( ) indicates the cropping process. The original image is of unknown shape and needs to be cropped to a specified square, such as 512*512, before being input into U-net. ( ) restores the image from a square to its original shape. This is the image before cropping.

[0032] In the actual implementation of S7, directly replacing the image may produce obvious seams at the mask boundaries. Therefore, a Gaussian kernel is used to smooth the local mask:

[0033] in Represents convolution. The standard deviation is Gaussian kernel, This represents the mask obtained in step S4.

[0034] The reconstructed cropped area is blended with the original cropped area in the following manner: , Each purified region was pasted back to its original image location. After processing all connected components, the final purified image was obtained. .

[0035] Another embodiment provides an adversarial camouflage detection and defacement system based on zero-sample and texture-independent methods, such as Figure 2 As shown, system 200 includes: The target candidate region detection module 210 is used to input the scene image into the pre-trained target detector, obtain a set of target candidate boxes through a low confidence threshold, and crop the target local image from the scene image based on the obtained target candidate boxes. The semantic-aware adversarial camouflage localization module 220 is used to extract a dense semantic response map for each target local image using a pre-trained encoder, and obtain a binary localization mask by applying a percentile-based threshold to the semantic response map. The binary localization masks are then projected back to the original scene image coordinate system and merged to form a global semantic localization mask. The multi-domain texture anomaly localization module 230 is used to calculate the texture anomaly score map of the target candidate box based on multiple complementary anomaly analysis methods for each target local image, and obtain a local binary texture anomaly mask by applying a percentile-based threshold to the texture anomaly score map. The local binary texture anomaly masks are then projected back to the original scene image coordinate system and merged to form a global texture anomaly localization mask. The abnormal texture mask fusion module 240 is used to fuse the global semantic positioning mask with the global texture abnormal positioning mask to obtain the fused mask; The adversarial texture removal and connected component decomposition module 250 is used to decompose the fusion mask into connected components, extend the bounding box of each component by a padding margin, and extract the edge structure of the target local image. The geometric condition construction and diffusion texture reconstruction module 260 is used to construct geometric conditions and perform diffusion texture reconstruction based on the target local image, the components decomposed by the fusion mask, the edge structure, positive samples and negative sample prompts, to obtain the reconstructed target local image. The boundary-aware fusion module 270 is used to smooth the components of the target local image and the reconstructed target local image respectively and then fuse them to obtain the final purified image.

[0036] In addition to the modules described above, the zero-sample and texture-independent anti-camouflage detection and purification system 200 may also include other components. However, since these components are not relevant to the content of this disclosure, their illustrations and descriptions are omitted here.

[0037] Other specific working processes of the zero-sample and texture-independent anti-camouflage detection and purification method using the aforementioned zero-sample and texture-independent anti-camouflage detection and purification system 200 are described in the above-described embodiment of the zero-sample and texture-independent anti-camouflage detection and purification method, and will not be repeated here.

[0038] Another embodiment illustrates that the system of the present invention can also be achieved by means of... Figure 3 The architecture of the computing device shown is used to implement this. Figure 3 The architecture of the computing device is shown. For example... Figure 3 As shown, the system includes a computer system 310, a system bus 330, one or more CPUs 340, input / output 320, and a memory 350. The memory 350 may store various data or files used by the computer for processing and / or communication, as well as program instructions executed by the CPU, including zero-sample and texture-independent anti-camouflage detection and deconstruction methods. Figure 3 The architecture shown is merely exemplary and should be adjusted according to actual needs when implementing different devices. Figure 3One or more components are included. The memory 350, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the zero-sample and texture-independent anti-camouflage detection and cleanup method described above in this embodiment of the invention. One or more CPUs 340 execute various functional applications and data processing of the system of the present invention by running the software programs, instructions, and modules stored in the memory 350.

[0039] Of course, the server provided in the embodiments of the present invention is not limited to performing the method operations described above, but can also perform related operations in the above-described zero-sample and texture-independent anti-camouflage detection and purification method provided in any embodiment of the present invention.

[0040] The memory 350 may primarily include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on terminal usage. Furthermore, the memory 350 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 350 may further include memory remotely configured relative to one or more CPUs 340, these remote memories being connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0041] Input / output 320 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the device. Input / output 320 may also include a display device such as a display screen.

[0042] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, this computer program implements the zero-sample and texture-independent anti-spoofing detection and cleanup method described in the above embodiments. The computer-readable storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0043] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0044] The program code contained on the storage medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0045] Furthermore, other specific working processes of a non-temporary computer-readable storage medium are described in the above-described embodiments of the adversarial spoofing detection and cleanup method based on zero-sample and texture-independent methods, and will not be repeated here.

[0046] In this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a step or method that comprises a list of elements includes not only those elements but also other elements not expressly listed or inherent to such a step or method.

[0047] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A zero-shot and texture-agnostic adversarial camouflage detection and purification method, characterized in that, Includes the following steps: The scene image is input into the pre-trained target detector. A set of target candidate boxes is obtained by passing a low confidence threshold. Based on the obtained target candidate boxes, the target local image is cropped from the scene image. For each target local image, a dense semantic response map is extracted using a pre-trained encoder. A binary localization mask is obtained by applying a percentile-based threshold to the semantic response map. Each binary localization mask is then projected back to the original scene image coordinate system and merged to form a global semantic localization mask. For each target local image, a texture anomaly score map of the target candidate box is calculated based on multiple complementary anomaly analysis methods. By applying a percentile-based threshold to the texture anomaly score map, a local binary texture anomaly mask is obtained. Each local binary texture anomaly mask is projected back to the original scene image coordinate system and merged to form a global texture anomaly localization mask. The global semantic localization mask is fused with the global texture anomaly localization mask to obtain the fused mask; The fusion mask is decomposed into connected components. For each component, its bounding box is extended by a padding margin, and the edge structure of the target local image is extracted. Based on the target local image, the components decomposed by the fusion mask, edge structure, positive and negative sample prompts, geometric conditions are constructed and diffusion-type texture reconstruction is performed to obtain the reconstructed target local image; The components of the target local image and the reconstructed target local image are smoothed separately and then fused to obtain the final purified image.

2. The adversarial camouflage detection and cleanup method based on zero-shot and texture-independent methods according to claim 1, characterized in that, The binary localization mask is specifically expressed as: , , , wherein represents an operation of removing connected components with an area less than a given threshold, is an indicator function, represents a semantic response graph, represents a spatial position within a target candidate frame, represents the first percentile operator.

3. The zero-sample and texture-independent adversarial camouflage detection and cleanup method according to claim 1, characterized in that, The texture anomaly score map of the target candidate box is calculated based on a plurality of complementary anomaly analysis methods, and a specific expression is as follows: wherein, represents a set of anomaly analysis methods, represents a contribution of the first anomaly analysis method, and , represents a normalization operation, represents a quantization value of the anomaly analysis method, represents a target candidate box obtained from a scene image , represents a number of target candidate regions detected.

4. The zero-sample and texture-agnostic adversarial camouflage detection and purification method according to claim 3, characterized in that, Anomaly analysis methods include bilateral residuals, Laplace response, local standard deviation, and color saturation.

5. The zero-sample and texture-independent adversarial camouflage detection and cleanup method according to claim 1, characterized in that, The global semantic localization mask is fused with the global texture anomaly localization mask to obtain the fused mask. The specific expression is: ,in, This represents a global texture anomaly localization mask. This represents a global semantic location mask. This represents element-wise multiplication. This indicates the removal of small components.

6. The zero-sample and texture-independent adversarial camouflage detection and cleanup method according to claim 1, characterized in that, Use Canny to extract the edge structure of a target local image.

7. The adversarial camouflage detection and cleanup method based on zero-sample and texture-independent methods according to claim 1, characterized in that, The components of the target local image and the reconstructed target local image are smoothed separately and then fused. The specific expression is as follows: , ,in, Represents convolution. The standard deviation is Gaussian kernel, This represents the reconstructed local image of the target. Represents a local image of the target. This represents the mask corresponding to a local image of the target.

8. A zero-sample and texture-independent adversarial camouflage detection and cleanup system, characterized in that, The system includes: The target candidate region detection module is used to input the scene image into the pre-trained target detector, obtain a set of target candidate boxes through a low confidence threshold, and crop the target local image from the scene image based on the obtained target candidate boxes; The semantic-aware adversarial camouflage localization module is used to extract a dense semantic response map for each target local image using a pre-trained encoder, and obtain a binary localization mask by applying a percentile-based threshold to the semantic response map. The binary localization masks are then projected back to the original scene image coordinate system and merged to form a global semantic localization mask. The multi-domain texture anomaly localization module is used to calculate the texture anomaly score map of the target candidate box based on multiple complementary anomaly analysis methods for each target local image. By applying a percentile-based threshold to the texture anomaly score map, a local binary texture anomaly mask is obtained. Each local binary texture anomaly mask is projected back to the original scene image coordinate system and merged to form a global texture anomaly localization mask. The abnormal texture mask fusion module is used to fuse the global semantic positioning mask with the global texture abnormal positioning mask to obtain the fused mask; The adversarial texture removal and connected component decomposition module is used to decompose the fusion mask into connected components, extend the bounding box of each component by a padding margin, and extract the edge structure of the target local image. The geometric condition construction and diffusion texture reconstruction module is used to construct geometric conditions and perform diffusion texture reconstruction based on the target local image, the components decomposed by the fusion mask, edge structure, positive and negative sample prompts, to obtain the reconstructed target local image. The boundary-aware fusion module is used to smooth the components of the target local image and the reconstructed target local image separately and then fuse them to obtain the final purified image.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the zero-sample and texture-independent anti-camouflage detection and cleanup method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, wherein computer instructions are stored on the storage medium, characterized in that, When the instruction is executed by the processor, it implements the steps of the zero-sample and texture-independent anti-camouflage detection and cleanup method as described in any one of claims 1 to 7.

Citation Information

Patent Citations

  • Multi-modal large model-based remote sensing target detection sample generation method and system

    CN122391907A