Adversarial robust object detection method, system, and device based on high contrast and super resolution

Abstract

This invention discloses an adversarial robust target detection method, system, and device based on high contrast and super-resolution, relating to the field of computer vision technology. The key technical points are: acquiring image data containing adversarial perturbations; binarizing the image data to obtain a high-contrast skeleton map; extracting the structural feature map from the high-contrast skeleton map and generating a candidate region set based on the structural feature map; performing super-resolution reconstruction on the image data to obtain a super-resolution reconstructed map; extracting the texture feature map from the super-resolution reconstructed map; wherein the structural feature map and the texture feature map have the same scale; weighting and fusing the structural feature map and the texture feature map according to a preset attention weight tensor to obtain a fused feature map; inputting the fused feature map and the candidate region set into a target detection head, and outputting the target detection result.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and more specifically, to adversarially robust target detection methods, systems, and devices based on high contrast and super-resolution. Background Technology

[0002] Object detection technology is widely used in scenarios such as autonomous driving, industrial inspection, video surveillance, and smart manufacturing. In these scenarios, the detection results are crucial not only for the accuracy of object category identification but also for the stable localization of the object's location. Therefore, the robustness of the model directly affects the system's security and availability. Among existing object detection models, two-stage detection models typically exhibit high detection accuracy under normal conditions because they first generate candidate regions and then perform classification and regression refinement. However, their detection results are highly dependent on the quality of the generated candidate regions. Once the front-end input is disturbed, the error tends to amplify step by step along the "candidate region generation—region feature extraction—classification and regression" chain, ultimately manifesting as missed detections, false detections, and detection box offsets.

[0003] In recent years, adversarial attacks have further amplified these shortcomings. Adversarial perturbations refer to tiny, pixel-level disturbances or local patches applied to an image. While their amplitude is small, even imperceptible to the human eye, they significantly alter how the model responds to the target region. For object detection tasks, such perturbations can not only lead to classification errors but also damage target edges and weaken local textures, thus affecting the candidate region network's judgment of the target's location. Therefore, compared to simple image classification tasks, object detection faces the problem of "simultaneous instability in localization and recognition" under adversarial attack conditions, making defense much more difficult.

[0004] Existing adversarial robustness detection schemes can be broadly categorized into three types. The first type improves the model's adaptability to known attack samples through adversarial training, but suffers from high training costs, complex sample construction, and limited generalization ability to unknown attack forms. The second type denoises, reconstructs, or cleanses the entire image at the input, which can mitigate perturbations to some extent, but often loses the original edge structure and fine-grained texture of the target, easily leading to the problem of "perturbations being suppressed while effective features are weakened." The third type focuses on enhancing or fusing feature layers, but most schemes still rely on a single information source, making it difficult to simultaneously consider the contour information required for target localization and the semantic texture information required for target recognition. Especially in two-stage detection frameworks, existing methods generally lack a dedicated robust design for the candidate region generation stage, making it difficult for subsequent detection heads to effectively compensate for any deviations in the front end. Therefore, there is an urgent need to propose an adversarial robust target detection method that can suppress the impact of adversarial perturbations while maintaining target structure, texture restoration, and detection efficiency. Summary of the Invention

[0005] The purpose of this invention is to provide an adversarial robust target detection method, system, and device based on high contrast and super-resolution, which solves the problems of candidate region shift, target structure destruction, and texture semantic loss that are easily caused by adversarial perturbation in existing target detection technologies.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0007] In a first aspect, the present invention provides an adversarial robust target detection method based on high contrast and super-resolution, the method comprising:

[0008] Acquire image data containing adversarial perturbations;

[0009] The image data is binarized to obtain a high-contrast skeleton map;

[0010] Extract the structural feature map from the high-contrast skeleton map, and generate a candidate region set based on the structural feature map;

[0011] Super-resolution reconstruction is performed on the image data to obtain a super-resolution reconstructed image;

[0012] Extract the texture feature map from the super-resolution reconstructed map; wherein the structural feature map and the texture feature map have the same scale;

[0013] The structural feature map and the texture feature map are fused together using a preset attention weight tensor to obtain a fused feature map;

[0014] The fused feature map and the candidate region set are input into the target detection head, and the target detection result is output.

[0015] In one implementation, before binarizing and super-resolution reconstructing the image data, the method further includes performing smoothing filtering on the image data.

[0016] In one implementation, the image data is binarized to obtain a high-contrast skeleton map, including:

[0017] Calculate the gradient response of the image data in the horizontal and vertical directions, and calculate the gradient magnitude map based on the gradient response in the horizontal and vertical directions;

[0018] The gradient magnitude map is binarized according to a preset adaptive threshold to obtain a high-contrast skeleton map.

[0019] In one implementation, extracting the structural feature map of the high-contrast skeleton map includes:

[0020] The first feature of the high-comparison skeleton map is extracted using the backbone network;

[0021] The first feature is aligned and fused using a feature pyramid network to obtain a structural feature map.

[0022] In one implementation, generating a candidate region set based on the structural feature map includes:

[0023] The structural feature map is used to select regions through a region candidate network to obtain multiple candidate regions;

[0024] The multiple candidate regions are filtered using a nonmaximum suppression algorithm to obtain a set of candidate regions.

[0025] In one implementation, extracting the texture feature map of the super-resolution reconstructed map includes:

[0026] The second feature of the super-resolution reconstructed image is extracted using the backbone network;

[0027] The second feature is aligned and fused using a feature pyramid network to obtain a texture feature map.

[0028] In one implementation, the process of determining the attention weight tensor is as follows:

[0029] The structural feature map and texture feature map are subjected to weight normalization to obtain the weight normalization result;

[0030] The weight normalization result is multiplied by a preset weight matrix, and the multiplication result is summed with a preset bias vector. A global average pooling operation is then performed on the summed result to obtain the attention weight tensor.

[0031] A second aspect of the present invention provides an adversarial robust target detection system based on high contrast and super-resolution, the system comprising:

[0032] The image acquisition module is used to acquire image data containing adversarial perturbations;

[0033] The image processing module is used to perform binarization processing on the image data to obtain a high-contrast skeleton map;

[0034] The structural feature extraction module is used to extract the structural feature map of the high-contrast skeleton map and generate a candidate region set based on the structural feature map;

[0035] The image reconstruction module is used to perform super-resolution reconstruction on the image data to obtain a super-resolution reconstructed image;

[0036] A texture feature extraction module is used to extract the texture feature map of the super-resolution reconstructed image; wherein the structural feature map and the texture feature map have the same scale;

[0037] The feature fusion module is used to fuse the structural feature map and the texture feature map in a weighted manner according to a preset attention weight tensor to obtain a fused feature map;

[0038] The target detection module is used to input the fused feature map and the candidate region set into the target detection head and output the target detection result.

[0039] A third aspect of the present invention provides an electronic device, including a memory and a processor;

[0040] A memory for storing computer programs, the computer programs including program instructions;

[0041] A processor for executing the program instructions to cause the electronic device to perform the steps of an adversarial robust target detection method based on high contrast and super-resolution as provided in the first aspect.

[0042] A fourth aspect of the present invention provides a computer-readable storage medium comprising a computer program that, when executed by one or more processors, implements an adversarial robust target detection method based on high contrast and super-resolution as provided in the first aspect.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] This invention first constructs a high-contrast skeleton branch, making the candidate region generation process rely more on the stable contour structure of the target rather than the perturbed original pixel response, thereby reducing the interference of adversarial perturbations on the region proposal stage and mitigating the problems of candidate box offset, false target response, and missed detection of real targets. Second, by setting a super-resolution reconstruction branch, texture and deep semantic information in the perturbed image are restored, thus compensating for the lack of information in target recognition by simple structural representation. Third, through a channel attention fusion mechanism, structural features and texture features are adaptively coupled at a unified scale, enabling the model to maintain contour stability while taking into account class discrimination ability. Finally, this invention organically combines skeleton-guided candidate region generation, texture restoration, and cascaded detection refinement, ensuring detection performance while avoiding indiscriminate heavy processing of the entire image, thus achieving a good balance between robustness, accuracy, and computational efficiency. Attached Figure Description

[0045] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0046] Figure 1 A flowchart of an adversarial robust target detection method based on high contrast and super-resolution provided in an embodiment of the present invention;

[0047] Figure 2 This is a model architecture diagram for implementing the target detection method provided in an embodiment of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0049] It should be noted that the terms "comprising" or "may include" used in the various embodiments of this application indicate the presence of the claimed function, operation, or element, and do not limit the addition of one or more functions, operations, or elements. Furthermore, as used in the various embodiments of this application, the terms "comprising," "having," and their cognates are intended only to indicate a specific feature, number, step, operation, element, component, or combination of the foregoing, and should not be construed as primarily excluding the presence of one or more other features, numbers, steps, operations, elements, components, or combinations of the foregoing, or adding one or more combinations of the foregoing.

[0050] It should be understood that terms such as "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0051] To address the issues of candidate region shift, target structure destruction, and texture semantic loss that easily occur in existing target detection models under adversarial perturbations, this invention proposes an adversarial robust target detection method based on high contrast and super-resolution. This method does not simply stack several general modules before and after existing two-stage detection frameworks. Instead, it constructs two complementary processing branches around structurally stable localization and texture robust recognition. Through the collaborative design of candidate region generation and feature fusion mechanisms, perturbation suppression, feature recovery, and detection decision-making are coupled into an integrated detection process.

[0052] Specifically, this invention first extracts high-contrast skeleton information that characterizes the main contour of the target from the perturbed input image, and generates candidate regions using this skeleton branch. This allows the candidate region network to primarily rely on relatively stable structural features for target localization, thereby reducing the impact of adversarial perturbations on the front-end region proposal stage. Simultaneously, this invention performs super-resolution cleansing and reconstruction on the input image to restore fine-grained texture and deep semantic information, compensating for the skeleton branch's shortcomings in texture representation. Building upon this, this invention further introduces a channel attention fusion mechanism to adaptively fuse the features of the structural and texture branches, ensuring that the fusion result maintains the stability of the target contour while also considering the texture details required for category discrimination.

[0053] Through the above technical solution, the present invention does not separate perturbation suppression from target detection, but rather forms a holistic solution in which high-contrast skeleton construction, super-resolution purification and restoration, candidate region generation, and cascaded detection refinement support and cooperate with each other, thereby achieving a better balance between target detection accuracy, adversarial robustness, and computational efficiency.

[0054] like Figure 2 As shown, this invention constructs a two-branch, two-stage object detection framework for adversarial perturbation scenarios. This framework does not simply add preprocessing modules before and after existing detectors. Instead, it addresses two common problems in object detection under adversarial conditions: candidate region instability and texture semantic impairment. It establishes complementary structural and texture branches, and uses candidate region generation and feature fusion mechanisms to ensure both branches work together for the final detection decision. Specifically, the structural branch extracts relatively insensitive target contour information from the input image and generates candidate regions accordingly; the texture branch recovers fine-grained texture and deep semantic information from the perturbed image, supplementing subsequent object recognition and bounding box refinement; and the fusion module provides a unified representation of the two types of features, enabling the cascaded detection head to simultaneously utilize structural stability and texture discriminative capabilities for refined detection.

[0055] The training process of this invention is mainly divided into three stages: basic module pre-training, core module joint training, and end-to-end fine-tuning. These three stages are not isolated conventional training processes, but rather a gradual establishment of stability, synergy, and task adaptability around the dual-branch robust detection structure of this invention. It should be understood that the loss functions for each module utilize existing technologies; therefore, this embodiment does not provide a detailed description of the loss functions used in the training of each module.

[0056] like Figure 1As shown, this embodiment of the invention provides an adversarial robust target detection method based on high contrast and super-resolution. The method is designed for target detection scenarios with adversarial perturbations, including adversarial noise and adversarial patches. This method achieves stable localization and accurate identification of targets under perturbation conditions by constructing structural and texture branches and combining a region proposal network and feature fusion mechanism. The method includes the following steps:

[0057] S101, acquire image data containing adversarial perturbations.

[0058] In this embodiment, the adversarial perturbation can be adversarial noise or adversarial patch, and the image data can be an RGB image of power equipment, power scene, or other targets to be detected. After the input image is adjusted to a preset size and input into the detection chain, in order to reduce the impact of random high-frequency perturbations on subsequent structure extraction and texture restoration, a smoothing filter is performed on the image data. This smoothing filter is a Gaussian filter, and the Gaussian convolution kernel size can be set to 5×5, and the standard deviation σ can be set to 1.0. Through the above processing, high-frequency abnormal perturbation responses can be suppressed while preserving the target's main outline and key texture information as much as possible, thereby providing a stable input for subsequent dual-branch processing.

[0059] It should be noted that the purpose of smoothing is not just to perform conventional noise reduction, but to provide a more stable initial input for subsequent dual-structure branching, making it easier for subsequent structural branches to highlight the target contour and for texture branches to recover effective image information while suppressing disturbances.

[0060] The expression for smoothing filtering is: , In the above formula, " " indicates convolution operation, This represents the grayscale value of the image at the corresponding pixel coordinates. This represents a two-dimensional Gaussian function with a standard deviation of σ. This preprocessing step can mitigate the impact of random high-frequency perturbations on subsequent feature extraction while preserving as much information as possible about the target subject.

[0061] S102, binarize the image data to obtain a high-contrast skeleton map.

[0062] In this embodiment, the image data is binarized to obtain a high-contrast skeleton map. The specific process is as follows: the gradient response of the image data in the horizontal and vertical directions is calculated, and the gradient magnitude map is calculated based on the gradient response in the horizontal and vertical directions; the gradient magnitude map is binarized according to a preset adaptive threshold to obtain a high-contrast skeleton map.

[0063] Specifically, the gradient response is calculated using the Sobel operator, as follows: ; ;in, Indicates the input image. This represents the convolution kernel in the horizontal direction. This represents the convolution kernel in the vertical direction. The Sobel operator is used to extract grayscale variation information in the horizontal and vertical directions of an image, respectively, to highlight the edge contours and structural features of potential targets, and to provide a basis for subsequent gradient magnitude map calculation and high-contrast skeleton map generation.

[0064] The gradient magnitude map is calculated as follows: Then, based on the statistical characteristics of the gradient magnitude map, an adaptive threshold is determined, and the gradient magnitude map is binarized to obtain a high-contrast skeleton map, represented as: ;in, This represents an adaptive threshold. The adaptive threshold provided in this embodiment is adaptively determined based on the statistical characteristics of the gradient magnitude map. For example, the adaptive threshold can be determined based on at least one of the mean, standard deviation, and preset percentile of the pixel values ​​in the gradient magnitude map, in order to preserve high-response structural regions and suppress weak-response background regions. Regions with a pixel value of 1 correspond to high-contrast structural regions in the image data, mainly representing the contours and subject shapes of potential targets; regions with a pixel value of 0 correspond to background regions or weak-response regions.

[0065] Unlike conventional edge extraction, this invention forms a target structure skeleton for robust detection, which can highlight the main contour structure of the target and suppress weak background responses and some abnormal responses caused by adversarial perturbations.

[0066] S103: Extract the structural feature map of the high-contrast skeleton map, and generate a candidate region set based on the structural feature map.

[0067] Specifically, the structural feature map of the high-contrast skeleton map is extracted as follows: the first feature of the high-contrast skeleton map is extracted through the backbone network; the first feature is aligned and fused through the feature pyramid network to obtain the structural feature map.

[0068] The focus of this embodiment is not on which conventional backbone network to use, but on constructing a structurally robust branch using the contour structure information carried by the skeleton map, so that the generation of subsequent candidate regions depends more on the target contour than on the disturbed original texture. ,in, , , These represent the height, width, and number of channels of the corresponding layer feature map, respectively. In this embodiment, the backbone network can be ResNet-50. After the backbone network extracts features, a feature pyramid network is used to fuse the multi-scale feature maps to form a unified structural representation suitable for candidate region generation. In the above formula, Indicates an upsampling operation. This indicates a convolution operation, after which the number of channels in each layer's feature map is unified to the same dimension. Through this process, the structural branches can maintain a relatively stable representation of the target contour at different scales.

[0069] Furthermore, the structural feature map is input into the region proposal network, which outputs multiple candidate regions. Subsequently, a non-maximum suppression algorithm is used to filter and deduplicate the multiple candidate regions, resulting in a candidate region set.

[0070] Right now In the formula, M represents the number of proposed regions, and NMS is the non-maximum suppression algorithm. Each candidate region is characterized by its top-left corner coordinates and width and height parameters. Since the candidate region generation stage is based on structural features rather than directly on the features of the perturbed original image, the direct impact of adversarial perturbations on the candidate region localization stage can be reduced, improving the consistency between the candidate regions and the real target locations.

[0071] S104, perform super-resolution reconstruction on the image data to obtain a super-resolution reconstructed image.

[0072] Specifically, the mathematical expression for super-resolution reconstruction is: ,in, This represents the super-resolution reconstruction function. Super-resolution reconstruction can be implemented using a lightweight super-resolution model based on convolutional neural networks. The purpose of this embodiment is to restore the local texture, edge transitions, and fine-grained semantic information in image data that has been damaged by perturbations, thereby forming a texture restoration branch that complements the structural branch. This not only improves the visual resolution of the image but also enhances the effective texture information required for subsequent recognition while suppressing the effects of perturbations.

[0073] S105, extract the texture feature map from the super-resolution reconstructed map; wherein, the structural feature map and the texture feature map have the same scale.

[0074] In this embodiment, the texture feature map of the super-resolution reconstructed image is extracted as follows: the second feature of the super-resolution reconstructed image is extracted through the backbone network; the second feature is aligned and fused through the feature pyramid network to obtain the texture feature map.

[0075] It is important to understand that the backbone network in this step is the same as that in step S103. First, the super-resolution reconstructed image is input into the backbone network and the feature pyramid network to obtain a texture feature map that corresponds in scale to the structural feature map in step S103. The feature maps output by the texture branch at each level maintain consistency with the corresponding layers of the structural branch in terms of spatial size and channel dimension, facilitating subsequent layer-by-layer fusion. Since the input to the texture branch is the cleaned and reconstructed image, its output features can better preserve the fine-grained texture information and semantic discrimination information of the target. The structural branch corresponding to step S103 and the texture branch corresponding to this step are executed in parallel to improve processing efficiency and avoid unnecessary prior coupling between the two branches.

[0076] The same scale for structural feature maps and texture feature maps means that the two feature maps have the same height, width and number of channels. This is common knowledge in the technical field and will not be described in detail in this embodiment.

[0077] S106, the structural feature map and texture feature map are fused according to the preset attention weight tensor to obtain the fused feature map.

[0078] In this embodiment, the structural feature map and the texture feature map are first fused one-to-one at the corresponding scale level. Then, a channel attention-based weighted fusion method is adopted, that is, the channel weights are first calculated based on the channel statistics of the structural features and texture features, and then the two features are weighted and combined according to the attention weight tensor to obtain the fused feature map.

[0079] Through the weighted fusion step in this embodiment, contour information that is conducive to stable positioning in the structural branch can be preserved, while semantic details that are conducive to category discrimination in the texture branch can be enhanced, thereby forming a unified and robust detection feature representation.

[0080] Weighted fusion does not simply superimpose the two feature maps, but dynamically allocates them according to the contribution of different channels to the current detection task. The weighted fusion process can be represented by the following mathematical expression: ;in, express, express, express.

[0081] Secondly, the process of determining the attention weight tensor is as follows: perform weight normalization operation on the structural feature map and texture feature map to obtain the weight normalization result; multiply the weight normalization result with the preset weight matrix, then sum the multiplication result with the preset bias vector, and perform global average pooling operation on the summation result to obtain the attention weight tensor.

[0082] In this embodiment, the mathematical expression for the attention weight tensor is:

[0083] In the formula, This indicates a global average pooling operation. This indicates the weight normalization operation. and Here, represents the weight matrix and bias vector of the channel attention, respectively, and represents the learnable parameters that need to be trained.

[0084] Through this fusion process, features beneficial for stable localization in the structural branch are preserved, while features beneficial for category discrimination in the texture branch are enhanced, thus forming a unified robust detection feature representation.

[0085] S107 inputs the fused feature map and candidate region set into the target detection head and outputs the target detection result.

[0086] In this embodiment, the candidate region set obtained in step S103 and the fused feature map obtained in step S106 are jointly input into the cascaded ROI target detection head, and the final detection result is output through multi-level classification and bounding box regression. First, the candidate regions are mapped onto the fused feature map, and the region features corresponding to each candidate region are obtained through ROI feature extraction. Then, the region features are input into the cascaded ROI detection head, and multi-level classification and bounding box regression are performed to obtain the final target category, target location, and corresponding confidence score. Since the front-end candidate regions are mainly provided by structural branches, and the back-end classification and localization refinement are completed based on the fusion representation of structural features and texture features, this step can improve the target category discrimination accuracy and bounding box localization accuracy while reducing the impact of perturbations.

[0087] In this embodiment, to obtain the model parameters for performing steps S101 to S107, a dataset containing a mixture of clean and adversarial samples can be used to train the backbone network, region proposal network, super-resolution cleanup module, fusion module, and detection head. Adversarial samples can be constructed using FGSM, PGD, or adversarial patching. The loss function during training may include super-resolution reconstruction loss, region proposal loss, classification loss, bounding box regression loss, and feature alignment loss to constrain the collaborative representation of structural and texture branches. After training, the model can perform adversarial robust object detection according to steps S101 to S107.

[0088] In summary, firstly, by constructing a high-contrast skeleton branch, the candidate region generation process relies more on the stable contour structure of the target rather than the perturbed original pixel response, thereby reducing the interference of adversarial perturbations on the region proposal stage and mitigating the problems of candidate box offset, false target response, and missed detection of real targets. Secondly, by setting a super-resolution cleanup branch, texture and deep semantic information in the perturbed image are restored, thus compensating for the lack of information in target recognition by simple structural representation. Thirdly, through a channel attention fusion mechanism, structural features and texture features are adaptively coupled at a unified scale, enabling the model to maintain contour stability while taking into account class discrimination ability. Finally, this invention organically combines skeleton-guided candidate region generation, texture restoration, and cascaded detection refinement, ensuring detection performance while avoiding indiscriminate heavy processing of the entire image, thus achieving a good balance between robustness, accuracy, and computational efficiency. This embodiment achieves robust target detection for adversarial perturbation scenarios by constructing a high-contrast skeleton, super-resolution purification and restoration, structure-guided candidate region generation, and adaptive fusion of structural and texture features. It improves the stability of candidate region localization and overall detection robustness while ensuring detection accuracy.

[0089] This invention also provides an adversarial robust target detection system based on high contrast and super-resolution, the system comprising:

[0090] The image acquisition module is used to acquire image data containing adversarial perturbations;

[0091] The image processing module is used to binarize image data to obtain a high-contrast skeleton map;

[0092] The structural feature extraction module is used to extract the structural feature map of the high-contrast skeleton map and generate a candidate region set based on the structural feature map;

[0093] The image reconstruction module is used to perform super-resolution reconstruction of image data to obtain a super-resolution reconstructed image.

[0094] The texture feature extraction module is used to extract texture feature maps from the super-resolution reconstructed image; the structure feature map and the texture feature map have the same scale.

[0095] The feature fusion module is used to fuse the structural feature map and the texture feature map by weighting them according to the preset attention weight tensor to obtain the fused feature map;

[0096] The target detection module is used to input the fused feature map and the candidate region set into the target detection head and output the target detection result.

[0097] This application provides an adversarial robust target detection system based on high contrast and super-resolution, which is similar to the one described above. Figure 1 The adversarial robust target detection method based on high contrast and super-resolution shown is a technical solution under the same inventive concept. Through the detailed description of the adversarial robust target detection method based on high contrast and super-resolution provided in the above embodiments, those skilled in the art can clearly understand the implementation process of an adversarial robust target detection system based on high contrast and super-resolution in this embodiment. Therefore, for the sake of brevity, it will not be described again here.

[0098] Accordingly, the adversarial robust target detection system based on high contrast and super-resolution provided in this embodiment of the invention firstly constructs a high-contrast skeleton branch, making the candidate region generation process rely more on the stable contour structure of the target rather than the perturbed original pixel response, thereby reducing the interference of adversarial perturbations on the region proposal stage and mitigating the problems of candidate box offset, false target response, and missed detection of real targets. Secondly, by setting a super-resolution cleanup branch, the texture and deep semantic information in the perturbed image are restored, thereby compensating for the lack of information in target recognition by simple structural representation. Thirdly, through a channel attention fusion mechanism, structural features and texture features are adaptively coupled at a unified scale, enabling the model to maintain contour stability while taking into account class discrimination ability. Finally, this invention organically combines skeleton-guided candidate region generation, texture restoration, and cascaded detection refinement, ensuring detection performance while avoiding indiscriminate heavy processing of the entire image, thus achieving a good balance between robustness, accuracy, and computational efficiency.

[0099] This invention also provides an electronic device. The electronic device includes a processor, a memory, a communication interface, and at least one communication bus for connecting the processor, the memory, and the communication interface. The memory includes, but is not limited to, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (PROM), or portable read-only memory (CD-ROM), and is used for related instructions and data.

[0100] The communication interface is used to receive and send data. The processor can be one or more CPUs; if the processor is a single CPU, it can be a single-core CPU or a multi-core CPU. The processor in the electronic device reads one or more programs stored in memory and performs the following operations: acquiring image data containing adversarial perturbations; binarizing the image data to obtain a high-contrast skeleton map; extracting the structural feature map from the high-contrast skeleton map and generating a candidate region set based on the structural feature map; performing super-resolution reconstruction on the image data to obtain a super-resolution reconstructed map; extracting the texture feature map from the super-resolution reconstructed map; wherein the structural feature map and the texture feature map have the same scale; fusing the structural feature map and the texture feature map according to a preset attention weight tensor to obtain a fused feature map; inputting the fused feature map and the candidate region set into the target detection head and outputting the target detection result.

[0101] It should be noted that the specific implementation of each operation can be described above. Figure 1 The corresponding description of the method embodiments shown indicates that the electronic device can be used to execute an adversarial robust target detection method based on high contrast and super-resolution of the above method embodiments of this application, which will not be described in detail here.

[0102] This invention also provides a computer-readable storage medium, which is a memory device in a computer device for storing programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the operating system of the terminal. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the above embodiments regarding a method for adversarial robust target detection based on high contrast and super-resolution. Those skilled in the art should understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0103] This invention also provides a computer program product containing program instructions. The computer program product may be software or program products containing program instructions, capable of running on a computing device or stored on any usable medium. When the computer program product runs on at least one electronic device, it causes the at least one electronic device to perform an adversarial robust target detection method based on high contrast and super-resolution.

[0104] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An adversarial robust target detection method based on high contrast and super-resolution, characterized in that, The methods include: Acquire image data containing adversarial perturbations; The image data is binarized to obtain a high-contrast skeleton map; Extract the structural feature map from the high-contrast skeleton map, and generate a candidate region set based on the structural feature map; Super-resolution reconstruction is performed on the image data to obtain a super-resolution reconstructed image; Extract the texture feature map from the super-resolution reconstructed map; wherein the structural feature map and the texture feature map have the same scale; The structural feature map and the texture feature map are fused together using a preset attention weight tensor to obtain a fused feature map; The fused feature map and the candidate region set are input into the target detection head, and the target detection result is output.

2. The method according to claim 1, characterized in that, Before performing binarization and super-resolution reconstruction on the image data, the method further includes performing smoothing filtering on the image data.

3. The method according to claim 1, characterized in that, The image data is binarized to obtain a high-contrast skeleton map, including: Calculate the gradient response of the image data in the horizontal and vertical directions, and calculate the gradient magnitude map based on the gradient response in the horizontal and vertical directions; The gradient magnitude map is binarized according to a preset adaptive threshold to obtain a high-contrast skeleton map.

4. The method according to claim 1, characterized in that, Extracting the structural feature map of the high-contrast skeleton map includes: The first feature of the high-comparison skeleton map is extracted using the backbone network; The first feature is aligned and fused using a feature pyramid network to obtain a structural feature map.

5. The basic method according to claim 1, characterized in that, Generate a candidate region set based on the structural feature map, including: The structural feature map is used to select regions through a region candidate network to obtain multiple candidate regions; The multiple candidate regions are filtered using a nonmaximum suppression algorithm to obtain a set of candidate regions.

6. The method according to claim 1, characterized in that, Extracting the texture feature map from the super-resolution reconstructed map includes: The second feature of the super-resolution reconstructed image is extracted using the backbone network; The second feature is aligned and fused using a feature pyramid network to obtain a texture feature map.

7. The method according to claim 1, characterized in that, The process for determining the attention weight tensor is as follows: The structural feature map and texture feature map are subjected to weight normalization to obtain the weight normalization result; The weight normalization result is multiplied by a preset weight matrix, and the multiplication result is summed with a preset bias vector. A global average pooling operation is then performed on the summed result to obtain the attention weight tensor.

8. An adversarial robust target detection system based on high contrast and super-resolution, characterized in that, The system includes: The image acquisition module is used to acquire image data containing adversarial perturbations; The image processing module is used to perform binarization processing on the image data to obtain a high-contrast skeleton map; The structural feature extraction module is used to extract the structural feature map of the high-contrast skeleton map and generate a candidate region set based on the structural feature map; The image reconstruction module is used to perform super-resolution reconstruction on the image data to obtain a super-resolution reconstructed image; A texture feature extraction module is used to extract the texture feature map of the super-resolution reconstructed image; wherein the structural feature map and the texture feature map have the same scale; The feature fusion module is used to fuse the structural feature map and the texture feature map in a weighted manner according to a preset attention weight tensor to obtain a fused feature map; The target detection module is used to input the fused feature map and the candidate region set into the target detection head and output the target detection result.

9. An electronic device, characterized in that, Including memory and processor; A memory for storing computer programs, the computer programs including program instructions; A processor is configured to execute the program instructions to cause the electronic device to perform the steps of an adversarial robust target detection method based on high contrast and super-resolution as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a computer program that, when executed by one or more processors, implements an adversarial robust target detection method based on high contrast and super-resolution as described in any one of claims 1 to 7.

Images