Method and device for generating countermeasure image, and method for processing target covering

By generating an adversarial image block suitable for infrared images, the problem of angle dependence of existing infrared pedestrian detection systems in 3D environments is solved, effective stealth effect in three-dimensional space is achieved, and the reliability and security of infrared pedestrian detection is improved. .

CN114550217BActive Publication Date: 2024-12-03TSINGHUA UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210188033.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-02-28
Publication Date
2024-12-03
Estimated Expiration
2042-02-28

AI Technical Summary

Technical Problem

When facing a 3D environment, the existing infrared pedestrian detection system has significantly weakened or failed when facing a 3D environment, and cannot effectively deal with pedestrian detection in three-dimensional space.

Method used

By generating adversarial image blocks, using probability information and object detection network to optimize images, generate binary images suitable for infrared images, cover the target object area, and form adversarial images with non-rigid deformation visual effects, suitable for pedestrian detection in three-dimensional space .

Benefits of technology

It realizes that the pedestrian angle can produce an anti-confrontation effect in infrared images, which improves the reliability and safety of infrared pedestrian detection and is suitable for pedestrian detection in three-dimensional space.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114550217B_ABST
    Figure CN114550217B_ABST
Patent Text Reader

Abstract

The present disclosure relates to an adversarial image generation method and device, and a target covering processing method, the method comprising: obtaining a first probability image according to probability information; obtaining an adversarial image block according to the first probability image; obtaining a second infrared image according to the adversarial image block and a first infrared image; detecting the second infrared image through a target detection network to obtain a target detection result; optimizing the first probability image according to the target detection result to obtain an adversarial image. According to the adversarial image generation method of an embodiment of the present disclosure, an adversarial image block can be generated based on the first probability image, and the target object in the infrared image can be covered, so that the adversarial image block is not limited by dimension and can be applied to three-dimensional pedestrians in the infrared image without making the pedestrians in the infrared image face the infrared camera at a specific angle.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to a method and device for generating a countermeasure image, and a method for processing a target covering. Background Art

[0002] Infrared imaging refers to a technology that uses photoelectric technology to detect specific infrared band signals of thermal radiation from objects, converts the signals into images that can be distinguished by the human eye, and further calculates the temperature value. Infrared imaging technology has a wide range of applications in night security, autonomous driving, and human body temperature measurement. Compared with visible light imaging, infrared imaging technology has the advantages of passive imaging, temperature sensitivity, and penetrability.

[0003] Pedestrian detection refers to the use of computer vision technology to determine whether there are pedestrians in an image or video sequence and to accurately locate them. Infrared pedestrian detection specifically refers to pedestrian detection in infrared images or video sequences. With the development of deep learning technology, pedestrian detection technology has made great progress, and the detection accuracy has gradually improved. But on the other hand, people have also found that there are some carefully designed perturbations that can make pedestrian detection algorithms prone to misjudgment or omission. These carefully designed perturbations are called adversarial samples. Interference with pedestrian detection algorithms is called adversarial attacks. According to the different fields of attack, adversarial attacks can be divided into attacks in the digital world and attacks in the physical world. The study of adversarial samples is of great significance to the safety and reliability of pedestrian detection. As infrared pedestrian detection systems are increasingly deployed in people's daily lives, the research on the safety of infrared pedestrian detection has received more and more attention and attention. The research on infrared adversarial samples has high practical significance and application value.

[0004] In the related art, for the physical confrontation method of infrared pedestrian detection, an confrontation stealth board can be designed based on heating elements (for example, light bulbs), and the position of the heating elements on the board can be optimized by AI. For example, there are N small light bulbs of the same type on the board, and each lit small light bulb will show a specific pattern in infrared imaging. The AI ​​optimization algorithm can optimize the position of these light bulbs to obtain the desired pattern in infrared imaging. This pattern can make the infrared detector unable to recognize pedestrians. As long as the pedestrian holds this confrontation board with heating elements in specific positions, he can achieve stealth to the thermal infrared pedestrian detector, thereby achieving the purpose of obtaining adversarial samples.

[0005] In the work in the visible light field, a texture pattern can be optimized through circular sampling, so that clothes covered with this texture can be invisible to the pedestrian detection model based on deep learning neural network at all angles. Pedestrians wearing such clothes can make it difficult for deep learning neural network to detect pedestrians, thereby achieving the purpose of obtaining adversarial samples in the visible light field.

[0006] However, the anti-stealth board is only effective at a certain angle. For example, when a pedestrian holds the anti-stealth board in front of the infrared camera, it can achieve the stealth effect. When facing the infrared camera at other angles, the effect will be significantly weakened, and it will be completely ineffective when facing the back. This is essentially because the board forms a 2D anti-stealth pattern, which cannot adapt to 3D stealth tasks.

[0007] Textured clothing is a method for obtaining adversarial samples in the visible light domain. The fabric patterns formed are digitally printed and colored, and these patterns cannot be displayed in thermal infrared imaging. Therefore, this method cannot obtain adversarial samples in the infrared domain either. Summary of the invention

[0008] The present disclosure proposes a method and device for generating an adversarial image, and a method for processing a target covering.

[0009] According to one aspect of the present disclosure, a method for generating an adversarial image is provided, comprising: obtaining a first probability image based on probability information, wherein the probability information includes first probability information of each pixel point, the first probability image is a binary image, and the pixel values ​​of the pixels of the first probability image are values ​​obtained based on the first probability information; obtaining an adversarial image block that covers at least a portion of an area where a target object in an infrared image is located based on the first probability image; obtaining a second infrared image that covers at least a portion of an area where the target object in the first infrared image is located based on the adversarial image block and the first infrared image; detecting the second infrared image through a target detection network to obtain a target detection result; and optimizing the first probability image based on the target detection result to obtain an adversarial image, wherein the adversarial image is a binary image.

[0010] In one possible implementation, based on the first probability image, an adversarial image block is obtained that covers at least a portion of the area where the target object in the infrared image is located, including: periodically expanding the first probability image to obtain a second probability image whose size is larger than the first probability image; randomly cropping the second probability image to obtain a third probability image; based on the third probability image, obtaining a fourth probability image with a non-rigid deformation visual effect; and adding disturbance information to the fourth probability image to obtain the adversarial image block.

[0011] In a possible implementation, based on the third probability image, a fourth probability image with a non-rigid deformation visual effect is obtained, including: setting multiple key points in the third probability image; randomly deforming the third probability image to obtain a deformed image; and performing surface fitting processing on the deformed image based on the key points in the deformed image and the key points in the third probability image to obtain the fourth probability image.

[0012] In a possible implementation, adding disturbance information to the fourth probability image to obtain the adversarial image block includes: performing at least one of rotating, changing brightness, changing contrast, adding random noise, and changing size on the fourth probability image to obtain the adversarial image block.

[0013] In a possible implementation, the target detection result includes second probability information that a target object exists in the second infrared image, wherein the first probability image is optimized according to the target detection result to obtain an adversarial image, including: obtaining first optimization information according to the first probability information; obtaining second optimization information according to the second probability information; obtaining optimization information of the first probability image according to the first optimization information and the second optimization information; and adjusting pixel values ​​of the first probability image according to the optimization information to obtain the adversarial image.

[0014] According to one aspect of the present disclosure, a method for processing a target covering is provided, characterized in that it includes: obtaining a covering material according to the adversarial image; and obtaining a target covering according to the covering material.

[0015] In one possible implementation, obtaining the covering material includes: obtaining a covering material including the pattern of the adversarial image through a supporting material and a heat-insulating material, wherein the position of the heat-insulating material in the covering material corresponds to the position of a pixel point having a preset pixel value in the adversarial image.

[0016] According to one aspect of the present disclosure, an adversarial image generating device is provided, including: a first probability image obtaining module, used to obtain a first probability image according to probability information, wherein the probability information includes first probability information of each pixel point, the first probability image is a binary image, and the pixel value of the pixel point of the first probability image is a value obtained according to the first probability information; an adversarial image block obtaining module, used to obtain an adversarial image block covering at least a part of the area where a target object is located in an infrared image according to the first probability image; a covering module, used to obtain a second infrared image covering at least a part of the area where the target object is located in the first infrared image according to the adversarial image block and the first infrared image; a detection module, used to detect the second infrared image through a target detection network to obtain a target detection result; an optimization module, used to optimize the first probability image according to the target detection result to obtain an adversarial image, wherein the adversarial image is a binary image.

[0017] In a possible implementation, the adversarial image block acquisition module is further used to: periodically expand the first probability image to obtain a second probability image whose size is larger than the first probability image; randomly crop the second probability image to obtain a third probability image; based on the third probability image, obtain a fourth probability image with a non-rigid deformation visual effect; add disturbance information to the fourth probability image to obtain the adversarial image block.

[0018] In a possible implementation, the adversarial image block acquisition module is further used to: set multiple key points in the third probability image; randomly deform the third probability image to obtain a deformed image; and perform surface fitting processing on the deformed image based on the key points in the deformed image and the key points in the third probability image to obtain the fourth probability image.

[0019] In a possible implementation, the adversarial image block acquisition module is further used to: perform at least one of rotating, changing brightness, changing contrast, adding random noise, and changing size on the fourth probability image to obtain the adversarial image block.

[0020] In one possible implementation, the target detection result includes second probability information that a target object exists in the second infrared image, and the optimization module is further used to: obtain first optimization information based on the first probability information; obtain second optimization information based on the second probability information; obtain optimization information of the first probability image based on the first optimization information and the second optimization information; and adjust the pixel value of the first probability image based on the optimization information to obtain the adversarial image.

[0021] According to one aspect of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to call the instructions stored in the memory to execute the above method.

[0022] According to one aspect of the present disclosure, a computer-readable storage medium is provided, on which computer program instructions are stored, and the computer program instructions implement the above method when executed by a processor.

[0023] According to the adversarial image generation method of the embodiment of the present disclosure, an adversarial image block can be generated based on the first probability image, and the target object in the infrared image can be covered, so that the adversarial image block is not limited by the dimension and can be applied to three-dimensional pedestrians in the infrared image without making the pedestrians in the infrared image face the infrared camera at a specific angle. The first probability image can be optimized based on the target detection result of the target monitoring network, so that the adversarial image derived from the first probability image can achieve a good stealth effect in the infrared field, and then the adversarial image can be obtained.

[0024] It should be understood that the above general description and the following detailed description are exemplary and explanatory only and do not limit the present disclosure. Other features and aspects of the present disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and are used to illustrate the technical solutions of the present disclosure together with the specification.

[0026] Figure 1 A flowchart showing a method for generating an adversarial image according to an embodiment of the present disclosure is shown;

[0027] Figure 2 A schematic diagram showing a first probability image according to an embodiment of the present disclosure;

[0028] Figure 3 A schematic diagram showing an application of the adversarial image generation method according to an embodiment of the present disclosure;

[0029] Figure 4 A schematic diagram showing a method for processing a target covering according to an embodiment of the present disclosure;

[0030] Figure 5 A block diagram of an adversarial image generating device according to an embodiment of the present disclosure is shown;

[0031] Figure 6 A block diagram of an electronic device according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0032] Various exemplary embodiments, features and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0033] The word “exemplary” is used exclusively herein to mean “serving as an example, example, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments.

[0034] The term "and / or" herein is only a description of the association relationship of the associated objects, indicating that there may be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the term "at least one" herein represents any combination of at least two of any one or more of a plurality of. For example, including at least one of A, B, and C can represent including any one or more elements selected from the set consisting of A, B, and C.

[0035] In addition, in order to better illustrate the present disclosure, numerous specific details are given in the following specific embodiments. It should be understood by those skilled in the art that the present disclosure can also be implemented without certain specific details. In some examples, methods, means, components and circuits well known to those skilled in the art are not described in detail in order to highlight the subject matter of the present disclosure.

[0036] Figure 1 A flowchart of a method for generating an adversarial image according to an embodiment of the present disclosure is shown. Figure 1 As shown, the method includes:

[0037] In step S11, a first probability image is obtained according to the probability information, wherein the probability information includes first probability information of each pixel point, the first probability image is a binary image, and the pixel value of the pixel point of the first probability image is a value obtained according to the first probability information;

[0038] In step S12, based on the first probability image, an adversarial image block covering at least a portion of the area where the target object is located in the infrared image is obtained;

[0039] In step S13, a second infrared image covering at least a portion of the area where the target object in the first infrared image is located is obtained according to the adversarial image block and the first infrared image;

[0040] In step S14, the second infrared image is detected by a target detection network to obtain a target detection result;

[0041] In step S15, the first probability image is optimized according to the target detection result to obtain an adversarial image, where the adversarial image is a binary image.

[0042] According to the adversarial image generation method of the embodiment of the present disclosure, an adversarial image block can be generated based on the first probability image, and the target object in the infrared image can be covered, so that the adversarial image block is not limited by the dimension and can be applied to three-dimensional pedestrians in the infrared image without making the pedestrians in the infrared image face the infrared camera at a specific angle. The first probability image can be optimized based on the target detection result of the target monitoring network, so that the adversarial image derived from the first probability image can achieve a good stealth effect in the infrared field, and then the adversarial image can be obtained.

[0043] In a possible implementation, in view of the fact that the handheld countermeasure stealth plate in the related art is only applicable to the 2D field but not to the 3D field (i.e., when the angle deviates, the stealth effect is greatly reduced), and the clothing with special textures that counteracts the pedestrian detection model is only applicable to the visible light field, the present disclosure uses countermeasure image blocks to directly cover the area where the pedestrians in the infrared image are located, which can simulate the situation where the pedestrians in the infrared image are wearing clothing with the pattern of the countermeasure image, and can achieve the effect of countermeasure pedestrian detection algorithm in the detection of infrared images, that is, no matter what angle the pedestrians in the infrared image are located at the infrared camera, the countermeasure effect can be generated, which can be applicable to pedestrians in three-dimensional space, and the countermeasure image can also be applied to the infrared field. Using this countermeasure image, countermeasure samples can be generated to study the countermeasure samples and improve the reliability and safety of pedestrian detection in the infrared field.

[0044] In a possible implementation, in step S11, a first probability image can be obtained through probability information, and the probability information can represent the probability information of the specific value of the pixel value of each pixel in an image. For example, the probability information is probability information corresponding to a binary image, and the probability information is a multidimensional vector. Each element in the vector represents the probability that the pixel value of each pixel in the binary image is 1. Therefore, the multidimensional vector can also be represented as a probability set of each pixel. The probability information can be preset or randomly generated, and the present disclosure does not limit this. In the example, if the probability of an element is greater than or equal to a preset threshold (for example, 0.5), the pixel value of the pixel corresponding to the element in the first probability image is 1, otherwise, the pixel value is 0. The present disclosure does not limit the value of the pixel value and the specific numerical value of the probability threshold.

[0045] In one possible implementation, a first probability image can be obtained based on the above probability information. The first probability image is a binary image, that is, the pixel value is 1 or 0. The specific value of the pixel value of each pixel point in the first probability image is obtained based on the above probability information. For example, the pixel value corresponding to the pixel point whose probability is greater than a preset threshold can be set to 1, and the others can be set to 0.

[0046] Figure 2 A schematic diagram showing a first probability image according to an embodiment of the present disclosure is shown as follows: Figure 2 As shown, the first probability image is a binary image, in which pixels with a pixel value of 1 are black, and pixels with a pixel value of 0 are white, forming an image with a camouflage pattern similar to a QR code.

[0047] In a possible implementation, in step S12, an adversarial image block for fully or partially covering the area where the target object (e.g., a pedestrian) of the infrared image is located can be obtained based on the first probability image obtained above. The adversarial image block can be taken from the first probability image. For example, the first probability image is subjected to cropping, deformation, expansion, symmetry, rotation, and other processing to obtain the adversarial image block. Furthermore, the adversarial image block can also simulate the effect of non-rigid deformation (e.g., wrinkles of clothes) produced when the target object wears clothes with the pattern of the first probability image, so that the adversarial image block can be applicable to pedestrians in three-dimensional space. Clothes with the pattern of the first probability image or other target coverings that can cover pedestrians can also be produced based on this, so that when the pedestrian wears the target covering, the image captured is close to the effect of the above-mentioned adversarial image block (i.e., has the effect of non-rigid deformation). Therefore, when the adversarial image block can resist the pedestrian detection algorithm in the infrared field, the pedestrian wearing the target covering can also resist the pedestrian detection algorithm in the infrared field, and based on this, more adversarial samples can be captured for research.

[0048] In a possible implementation, step S12 may include: periodically expanding the first probability image to obtain a second probability image whose size is larger than that of the first probability image; randomly cropping the second probability image to obtain a third probability image; obtaining a fourth probability image with a non-rigid deformation visual effect based on the third probability image; and adding disturbance information to the fourth probability image to obtain the adversarial image block.

[0049] In a possible implementation, the first probability image may be periodically extended. In an example, the first probability image may be copied multiple times, and the multiple first probability images may be spliced ​​into a second probability image. In addition, during the splicing process, the edges of each first probability image are smooth, that is, the pixel values ​​of the pixels at the edges of the two spliced ​​first probability images are kept consistent. In an example, if the pixel value of a pixel at the edge of a first probability image is 1, the pixel values ​​of the pixels adjacent to the above pixel at the edge of the first probability image spliced ​​therewith are also set to 1, so that the connection at the edge is smoother.

[0050] In the example, the first probability image can be periodically expanded 9 times, that is, with the first probability image as the center, its left side, right side, upper side, lower side, upper left side, upper right side, lower left side and lower right side are all spliced ​​with the first probability image, and the spliced ​​edges are smoothed to obtain a second probability image that is 9 times the size of the first probability image. The present disclosure does not limit the number of periodic expansions of the first probability image, for example, it can be expanded four times, twice, six times, etc. This process can simulate the process of making the first probability image into a target covering (for example, clothes), the same pattern is periodically expanded multiple times to obtain a larger pattern, and the carrier material (for example, cloth) for making the target covering has the larger pattern, for example, so that the pattern including multiple first probability images covers the entire carrier material.

[0051] In a possible implementation, the size of the second probability image is larger than the first probability image, and an image block can be randomly selected in the second probability image for cropping, that is, a position is randomly selected in the second probability image with a larger size, and a size is randomly set for cropping, and the cropped image block is the third probability image. This process can simulate that after the pattern of the first probability image is made into a target cover (for example, clothes), pedestrians and other targets may face the infrared camera at any angle of the clothes. Therefore, a local part of the larger pattern can be obtained through random cropping, and the local pattern (third probability image) is the pattern of the target cover (for example, clothes) facing the infrared camera.

[0052] In a possible implementation, as described above, the image block that fully or partially covers the area where the pedestrian is located can also have a non-rigid deformation (e.g., wrinkles) visual effect, that is, simulating the effect of the target covering (e.g., clothes) worn on the pedestrian. The third probability image can be deformed to obtain a fourth probability image. This step may include: setting a plurality of key points in the third probability image; randomly deforming the third probability image to obtain a deformed image; performing surface fitting processing on the deformed image according to the key points in the deformed image and the key points in the third probability image to obtain the fourth probability image.

[0053] In a possible implementation, the above deformation process can be performed by thin plate spline interpolation processing. For example, multiple key points for interpolation or fitting can be set, and the third probability image can be randomly deformed to obtain a deformed image. The deformation effect of the deformed image may not be smooth, that is, it is impossible to simulate the smooth non-rigid deformation (for example, wrinkles) of the target cover. Therefore, the position of the key points in the deformed image and the position of the key points in the third probability image can be used for interpolation or fitting processing. For example, the position transformation relationship between the position of the key points in the third probability image and the position in the deformed image can be used to determine the position transformation relationship between each multiple pixel point in the third probability image and the multiple pixel points in the deformed image, and then the position transformation relationship (for example, the position transformation relationship that can be expressed by an equation) is fitted to make the deformation in the deformed image smoother, and obtain a fourth probability image. In the example, a deformed image with multiple edges and corners (that is, the edges and corners generated during deformation) can be surface fitted to obtain a fourth probability image with smoother deformation.

[0054] In a possible implementation, further, more disturbance information may be added to the fourth probability image to increase more visual effects of simulating pedestrians wearing the target covering. After adding the disturbance information, the adversarial image block may be obtained. This step may include: performing at least one of the following processings on the fourth probability image: rotating, changing brightness, changing contrast, adding random noise, and changing size to obtain the adversarial image block. Through at least one of the above processings, adversarial image blocks with more visual effects may be obtained.

[0055] In the example, the above processing can generate multiple adversarial image blocks. For example, in the process of random cropping, multiple third probability images can be cropped, and each third probability image can be respectively subjected to surface fitting and disturbance information addition to obtain multiple adversarial image blocks. In the process of adding disturbance information, multiple adversarial image blocks can also be obtained. For example, the fourth probability image can be subjected to the above-mentioned multiple processes of adding disturbance information to obtain multiple adversarial image blocks. For example, the fourth probability image can be rotated to obtain an adversarial image block, random noise can be added to the fourth probability image to obtain an adversarial image block, the fourth probability image can be resized and the contrast can be changed to obtain an adversarial image block... The present disclosure does not limit the process of adding random disturbances.

[0056] In the example, the above multiple processes for obtaining the adversarial image block from the first probability image do not need to be performed in full, and only a part of them may be performed, for example, only the period expansion, random clipping, and surface fitting are performed, and the process of adding disturbance information is not performed, and the adversarial image block can also be obtained. The present disclosure does not limit which specific processes are performed.

[0057] In this way, through the processing of periodic expansion, random cropping, surface fitting and adding disturbance information, when the adversarial image block covers the area where the pedestrian is located in the infrared image, a visual effect of the pedestrian wearing a target covering (for example, clothing) with a pattern of the first probability image can be obtained. The scene of the pedestrian wearing the target covering can be simulated, so that the above method is applicable to the three-dimensional field where the pedestrian wears the target covering, and the subsequent target detection results are more accurate, thereby improving the adversarial effect of the adversarial image and providing a basis for obtaining adversarial samples with better adversarial effects.

[0058] In a possible implementation, in step S13, the adversarial image block obtained above may be overlaid on the target object (e.g., pedestrian) region in the first infrared image, and the adversarial image block may partially or completely cover the target object region to obtain a second infrared image. In the example, the obtained multiple adversarial image blocks may be respectively overlaid on multiple target object regions in multiple first infrared images.

[0059] In a possible implementation, although the second infrared image includes the target object, the target object is covered by the adversarial image block. In a physical sense, the black pixel points (e.g., pixel points with a pixel value of 1) on the adversarial image block in the second infrared image represent points where no infrared thermal radiation is detected in the infrared imaging. Therefore, the probability that the target object covered by the adversarial image block is detected by the target detection algorithm in the infrared field (e.g., the target detection network based on deep learning in step S14) will decrease.

[0060] In a possible implementation, in step S14, the second infrared image may be detected by the target detection network to obtain a target detection result. In the example, the target detection result f(x,θ) of the target detection network (wherein f represents the operation process of the target detection network, x represents the input of the target monitoring network, such as the second infrared image, and θ represents the parameter of the target detection network) may include three aspects, namely, the second probability information f of the existence of the target object in the second infrared image, obj (x,θ), the location information of the detection box of the target object f pos (x,θ), and the category information f of the target object cls (x,θ). The purpose of the adversarial sample is to make it difficult for the target detection network to detect the existence of the target object, that is, to make the second probability information f of the target object obj (x,θ) decreases.

[0061] In one possible implementation, when the target detection network detects the second infrared image, the higher the proportion of black pixels in the adversarial image block, the more points that cannot detect infrared thermal radiation in a physical sense. If all pixels in the adversarial image block are black pixels, infrared thermal radiation cannot be detected at all, that is, the target object cannot be detected at all. However, the black pixels correspond to the thermal insulation materials on the target covering, that is, the thermal insulation materials can be used to achieve the effect of the black pixels in the adversarial image block to isolate infrared thermal radiation. Therefore, in order to reduce the amount of thermal insulation materials used, reduce the cost of the target covering, and make the target object more breathable and comfortable to wear, it can be considered to minimize the use of thermal insulation materials on the basis of reducing the probability of the target detection network detecting the existence of the target object, that is, to minimize the proportion of black pixels in the adversarial image block.

[0062] In a possible implementation, in step S15, the first probability image can be optimized based on the above two aspects so that the adversarial image derived therefrom can reduce the proportion of black pixels as much as possible while being able to reduce the probability of the target detection network detecting the existence of the target object.

[0063] In a possible implementation, step S15 may include: obtaining first optimization information based on the first probability information; obtaining second optimization information based on the second probability information; obtaining optimization information of the first probability image based on the first optimization information and the second optimization information; and adjusting the pixel values ​​of the first probability image based on the optimization information to obtain the adversarial image.

[0064] In a possible implementation, the first probability information is the probability information that the pixel value of each pixel in the first probability image is 1. As analyzed above, the proportion of black pixels can be reduced as much as possible on the basis of reducing the probability that the target object exists detected by the target detection network. Therefore, the probability information that the pixel value of each pixel is 1 can be reduced as much as possible. Therefore, the first optimization information can be obtained based on the first probability information. During the optimization process, the first optimization information can be minimized to reduce the probability information that the pixel value of each pixel is 1.

[0065] In an example, the first optimization information may be determined by the following formula (1):

[0066]

[0067] Among them, L black is the first optimization information, π0(i, j) represents the probability that the pixel value with coordinates (i, j) in the first probability image is 1 (ie, a black pixel).

[0068] In a possible implementation, as described above, the optimization goal is to reduce the proportion of black pixels as much as possible on the basis of being able to reduce the probability of the target object being detected by the target detection network, so that the first optimization information is minimized to meet the optimization goal of "reducing the proportion of black pixels as much as possible", and the second probability information f of the existence of the target object in the detection result of the target detection network is minimized. obj Minimizing (x,θ) meets the goal of "reducing the probability that the target detection network detects the existence of the target object".

[0069] In a possible implementation, the second optimization information may be obtained based on the second probability information. In an example, the second optimization information may be obtained according to the following formula (2):

[0070]

[0071] Among them, L obj is the second optimization information, M is the number of second infrared images, is the second probability information of the target object existing in the i-th second infrared image, E teT In the optimization process, minimizing the second optimization information meets the goal of "reducing the probability that the target detection network detects the existence of the target object".

[0072] In one possible implementation, the above two objectives can be combined to obtain optimization information of the first probability image. For example, the optimization goal is to minimize both the first optimization information and the second optimization information. Therefore, both can be minimized at the same time, which can achieve the optimization goal of reducing the proportion of black pixels as much as possible on the basis of reducing the probability of the target object detected by the target detection network.

[0073] In the example, the two can be weighted summed to obtain the optimization information of the first probability image, and the optimization information can be minimized to achieve the above goal. In the example, the above optimization information can be obtained by the following formula (3):

[0074] L=L obj +λL black (3)

[0075] Among them, L is the optimization information, λ is the weight, and during the optimization process, the optimization information can be minimized.

[0076] In a possible implementation, during the optimization process, back propagation can be performed based on the above optimization information. For example, the gradient descent method of the neural network can be used to adjust the pixel value of each pixel of the first probability image so that the above optimization information obtained based on the adjusted first probability image is minimized. Alternatively, if the binary first probability image is difficult to optimize by the gradient descent method, the above probability information can be adjusted, that is, the first probability information of each pixel represented by the probability information can be adjusted to achieve the purpose of adjusting the first probability image. For example, the first probability information of a certain pixel is adjusted from 0.6 (above the threshold) to 0.4 (below the threshold), then the pixel in the first probability image is adjusted from a pixel with a pixel value of 1 to a pixel with a pixel value of 0.

[0077] In a possible implementation, the above optimization process can be iteratively executed multiple times, and when the optimization information converges, the optimization process is completed to obtain the optimized first probability image, that is, the adversarial image. After the image block derived from the adversarial image is covered on the area where the target object is located in the infrared image, it can make it difficult for the target detection network to detect the target object in the infrared image, and at the same time, there are fewer black pixels in the adversarial image. Correspondingly, while enabling the target covering to achieve the adversarial target detection network, the amount of thermal insulation material used is minimized to save materials and improve comfort.

[0078] According to the adversarial image generation method of the embodiment of the present disclosure, the first probability image can be processed by periodic expansion, random clipping, surface fitting and adding disturbance information to generate an adversarial image block, so that when the adversarial image block covers the area where the pedestrian is located in the infrared image, the visual effect of the pedestrian wearing the target covering with the pattern of the first probability image is obtained, and the scene of the pedestrian wearing the target covering can be simulated, so that the above method is applicable to the three-dimensional field of the pedestrian wearing the target covering. And the target object in the infrared image is covered, so that the adversarial image block is not limited by the dimension, and can be applied to the three-dimensional pedestrian in the infrared image, without making the pedestrian in the infrared image face the infrared camera at a specific angle. And the first probability image can be optimized based on the target detection result of the target monitoring network, so that the adversarial image derived from the first probability image can achieve a good stealth effect in the infrared field, and through the above optimization process, the number of black pixels in the adversarial image can be minimized, correspondingly, the amount of thermal insulation material used for the target covering can be saved and the comfort can be improved. More adversarial images can also be obtained based on the target covering for research.

[0079] Figure 3 A schematic diagram showing an application of the adversarial image generation method according to an embodiment of the present disclosure is shown. Figure 3 As shown, the first probability image P can be obtained based on the probability information z basicIn the example, the Gumbel-Softmax function can be used to obtain the probability that the pixel value of each pixel in the first probability image is 1 based on the probability information, and then determine the pixel value of each pixel in the first probability image to obtain the first probability image.

[0080] In a possible implementation, the first probability image can be periodically extended to obtain a second probability image P tiled For example, the first probability image period can be expanded 9 times to obtain Figure 3 The second probability image P is shown tiled Furthermore, the second probability image can be randomly cropped, that is, the image block at any position in the second probability image is cropped to obtain a third probability image P of any size. crop .

[0081] In a possible implementation, the third probability image P crop The random deformation and surface fitting processing, that is, thin plate spline interpolation processing (TPS), is performed to obtain a fourth probability image P of a non-rigid deformation (eg, wrinkles) visual effect. crop-T Furthermore, the fourth probability image P crop-T Add disturbance information, for example, rotate the fourth probability image, change the brightness, change the contrast, add random noise, change the size of at least one of the above, to obtain multiple adversarial image blocks.

[0082] In a possible implementation, multiple adversarial image blocks can be respectively covered on multiple pedestrian areas of multiple infrared images to obtain covered infrared images and form a data set. Further, a detector (i.e., a target detection network) can be used to perform detection to obtain a target detection result. Further, the optimization information can be obtained by minimizing the proportion of black pixels based on the optimization goal of reducing the probability of the target object being detected by the target detection network, as shown in formula (3).

[0083] In a possible implementation, the pixel values ​​of the pixels of the first probability image can be optimized based on formula (3), so that the proportion of black pixels is reduced as much as possible, and the probability of the target detection network detecting the existence of the target object is reduced as much as possible. Correspondingly, the target cover can resist the target detection network, save the amount of thermal insulation material, and improve comfort.

[0084] According to another aspect of the present disclosure, a method for processing a target covering is provided, comprising: obtaining a covering material according to the counter image, and obtaining a target covering according to the covering material.

[0085] In a possible implementation, the above-mentioned confrontation image may be periodically extended to obtain a pattern of a covering material. For example, a covering material including the pattern of the confrontation image may be obtained through a carrier material and a heat-insulating material, wherein the position of the heat-insulating material in the covering material corresponds to the position of a pixel point having a preset pixel value in the confrontation image.

[0086] In a possible implementation, the carrier material may include fabric for processing clothes. According to the pattern of the covering material, a heat-insulating material may be added to the fabric so that the position of the heat-insulating material corresponds to the position of a pixel point (i.e., a black pixel point) with a preset pixel value in the pattern, and the covering material can be obtained. Furthermore, the covering material can be processed into a target covering such as clothes. When the target object wears the target covering and is photographed by an infrared camera, the generated infrared image is detected by a target detection network, and the probability of detecting the target object is low. Moreover, the amount of heat-insulating material used in the target covering obtained in this way is small, which can save materials and improve wearing comfort and breathability.

[0087] Figure 4 A schematic diagram showing a method for processing a target covering according to an embodiment of the present disclosure is shown. Figure 4 As shown, after periodic expansion of the counter image, the pattern P of the covering material can be obtained. tiled , a covering material can be printed based on the pattern, wherein the heat-insulating material corresponds to the black pixels in the pattern. Furthermore, the covering material can be processed into target coverings such as clothes. When the target object wearing the clothes is photographed by an infrared camera, the infrared image generated is difficult to be detected by the target detection network. Therefore, more adversarial samples can be taken for research.

[0088] Figure 5 A block diagram of a confrontation image generating device according to an embodiment of the present disclosure is shown as follows: Figure 5As shown, the device includes: a first probability image acquisition module 11, used to obtain a first probability image according to probability information, wherein the probability information includes first probability information of each pixel point, the first probability image is a binary image, and the pixel value of the pixel point of the first probability image is a value obtained according to the first probability information; an adversarial image block acquisition module 12, used to obtain an adversarial image block covering at least a part of the area where the target object in the infrared image is located according to the first probability image; a covering module 13, used to obtain a second infrared image covering at least a part of the area where the target object in the first infrared image is located according to the adversarial image block and the first infrared image; a detection module 14, used to detect the second infrared image through a target detection network to obtain a target detection result; an optimization module 15, used to optimize the first probability image according to the target detection result to obtain an adversarial image, wherein the adversarial image is a binary image.

[0089] In a possible implementation, the adversarial image block acquisition module is further used to: periodically expand the first probability image to obtain a second probability image whose size is larger than the first probability image; randomly crop the second probability image to obtain a third probability image; based on the third probability image, obtain a fourth probability image with a non-rigid deformation visual effect; add disturbance information to the fourth probability image to obtain the adversarial image block.

[0090] In a possible implementation, the adversarial image block acquisition module is further used to: set multiple key points in the third probability image; randomly deform the third probability image to obtain a deformed image; and perform surface fitting processing on the deformed image based on the key points in the deformed image and the key points in the third probability image to obtain the fourth probability image.

[0091] In a possible implementation, the adversarial image block acquisition module is further used to: perform at least one of rotating, changing brightness, changing contrast, adding random noise, and changing size on the fourth probability image to obtain the adversarial image block.

[0092] In one possible implementation, the target detection result includes second probability information that a target object exists in the second infrared image, and the optimization module is further used to: obtain first optimization information based on the first probability information; obtain second optimization information based on the second probability information; obtain optimization information of the first probability image based on the first optimization information and the second optimization information; and adjust the pixel value of the first probability image based on the optimization information to obtain the adversarial image.

[0093] It can be understood that the above-mentioned various method embodiments mentioned in the present disclosure can be combined with each other to form a combined embodiment without violating the principle logic. Due to space limitations, the present disclosure will not repeat them. It can be understood by those skilled in the art that in the above-mentioned method of the specific implementation method, the specific execution order of each step should be determined according to its function and possible internal logic.

[0094] In addition, the present disclosure also provides an adversarial image generation device, an electronic device, a computer-readable storage medium, and a program, all of which can be used to implement any adversarial image generation method provided by the present disclosure. The corresponding technical solutions and descriptions are referred to in the corresponding records of the method part and will not be repeated here.

[0095] In some embodiments, the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the method described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0096] The embodiment of the present disclosure also provides a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the above method is implemented. The computer-readable storage medium may be a non-volatile computer-readable storage medium.

[0097] An embodiment of the present disclosure further proposes an electronic device, comprising: a processor; and a memory for storing instructions executable by the processor; wherein the processor is configured to call the instructions stored in the memory to execute the above method.

[0098] The embodiments of the present disclosure also provide a computer program product, including a computer-readable code. When the computer-readable code is run on a device, a processor in the device executes instructions for implementing the adversarial image generation method provided in any of the above embodiments.

[0099] The embodiments of the present disclosure also provide another computer program product for storing computer-readable instructions, which, when executed, enable a computer to perform the operations of the adversarial image generation method provided in any of the above embodiments.

[0100] The electronic device may be provided as a terminal, a server, or a device in other forms.

[0101] Figure 6 1 is a block diagram of an electronic device 1900 according to an embodiment of the present disclosure. For example, the electronic device 1900 may be provided as a server or a terminal. Figure 6, the electronic device 1900 includes a processing component 1922, which further includes one or more processors, and a memory resource represented by a memory 1932 for storing instructions executable by the processing component 1922, such as an application. The application stored in the memory 1932 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 1922 is configured to execute instructions to perform the above method.

[0102] The electronic device 1900 may also include a power supply component 1926 configured to perform power management of the electronic device 1900, a wired or wireless network interface 1950 configured to connect the electronic device 1900 to a network, and an input / output (I / O) interface 1958. The electronic device 1900 may operate based on an operating system stored in the memory 1932, such as Windows Server 2000. TM , Mac OS X TM , Unix TM ,Linux TM , FreeBSD TM or similar.

[0103] In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, such as a memory 1932 including computer program instructions, which can be executed by the processing component 1922 of the electronic device 1900 to perform the above method.

[0104] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0105] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0106] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0107] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed completely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be customized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present disclosure.

[0108] Various aspects of the present disclosure are described herein with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer-readable program instructions.

[0109] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processor of the computer or other programmable data processing device, a device that implements the functions / actions specified in one or more boxes in the flowchart and / or block diagram is generated. These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause the computer, programmable data processing device, and / or other equipment to work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0110] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operating steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0111] The flow chart and block diagram in the accompanying drawings show the possible architecture, function and operation of the system, method and computer program product according to multiple embodiments of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a part of a module, program segment or instruction, and a part of the module, program segment or instruction includes one or more executable instructions for realizing the specified logical function. In some alternative implementations, the function marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two continuous square boxes can actually be executed substantially in parallel, and they can sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs the specified function or action, or can be implemented with a combination of special hardware and computer instructions.

[0112] The computer program product may be implemented in hardware, software or a combination thereof. In one optional embodiment, the computer program product is embodied as a computer storage medium, and in another optional embodiment, the computer program product is embodied as a software product, such as a software development kit (SDK) and the like.

[0113] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A method for generating an adversarial image, characterized in that: include: Obtaining a first probability image according to the probability information, wherein the probability information includes first probability information of each pixel point, the first probability image is a binary image, and the pixel values ​​of the pixels of the first probability image are values ​​obtained according to the first probability information; Obtaining, according to the first probability image, an adversarial image block covering at least a portion of an area where the target object in the infrared image is located; Obtaining, according to the adversarial image block and the first infrared image, a second infrared image covering at least a portion of an area where the target object in the first infrared image is located; Detecting the second infrared image through a target detection network to obtain a target detection result; Optimizing the first probability image according to the target detection result to obtain an adversarial image, where the adversarial image is a binary image; The target detection result includes second probability information that a target object exists in the second infrared image. The step of optimizing the first probability image according to the target detection result to obtain the adversarial image includes: Obtaining first optimization information according to the first probability information; Obtaining second optimization information according to the second probability information; Obtaining optimization information of the first probability image according to the first optimization information and the second optimization information; Adjusting the pixel values ​​of the first probability image according to the optimization information to obtain the adversarial image; The first optimization information is calculated by the following formula: Among them, L black is the first optimization information, π0(i, j) represents the probability that the pixel value at the coordinate (i, j) in the first probability image is 1; The second optimization information is calculated by the following formula: Among them, L obj is the second optimization information, M is the number of second infrared images, is the second probability information of the target object existing in the i-th second infrared image, E t∈T represents mathematical expectation; The optimization information L of the first probability image is calculated by the following formula: L=L obj +λL black Among them, λ is the weight.

2. The method according to claim 1, characterized in that: Obtaining, according to the first probability image, an adversarial image block covering at least a portion of an area where the target object is located in the infrared image, comprising: Periodically expanding the first probability image to obtain a second probability image that is larger in size than the first probability image; Randomly cropping the second probability image to obtain a third probability image; Obtaining a fourth probability image having a non-rigid deformation visual effect according to the third probability image; Add disturbance information to the fourth probability image to obtain the adversarial image block.

3. The method according to claim 2, characterized in that Obtaining a fourth probability image with a non-rigid deformation visual effect according to the third probability image, comprising: Setting a plurality of key points in the third probability image; Randomly deforming the third probability image to obtain a deformed image; According to the key points in the deformation image and the key points in the third probability image, a surface fitting process is performed on the deformation image to obtain the fourth probability image.

4. The method according to claim 2, characterized in that: Adding disturbance information to the fourth probability image to obtain the adversarial image block includes: The fourth probability image is subjected to at least one of rotation, brightness change, contrast change, random noise addition, and size change to obtain the adversarial image block.

5. A method for processing a target covering, characterized in that: include: Obtaining a covering material according to the adversarial image generation method according to any one of claims 1 to 4; According to the covering material, a target covering is obtained.

6. The method according to claim 5, characterized in that The obtaining of the covering material comprises: A covering material including a pattern of the confrontation image is obtained by using a supporting material and a heat-insulating material, wherein the position of the heat-insulating material in the covering material corresponds to the position of a pixel point having a preset pixel value in the confrontation image.

7. A device for generating an adversarial image, characterized in that: include: A first probability image acquisition module, configured to acquire a first probability image according to the probability information, wherein the probability information includes first probability information of each pixel point, the first probability image is a binary image, and the pixel values ​​of the pixels of the first probability image are values ​​obtained according to the first probability information; An adversarial image block obtaining module, configured to obtain, based on the first probability image, an adversarial image block covering at least a portion of an area where a target object is located in the infrared image; A covering module, configured to obtain, based on the adversarial image block and the first infrared image, a second infrared image covering at least a portion of an area where the target object in the first infrared image is located; A detection module, used to detect the second infrared image through a target detection network to obtain a target detection result; An optimization module, configured to optimize the first probability image according to the target detection result to obtain an adversarial image, wherein the adversarial image is a binary image; The target detection result includes second probability information that a target object exists in the second infrared image. The step of optimizing the first probability image according to the target detection result to obtain the adversarial image includes: Obtaining first optimization information according to the first probability information; Obtaining second optimization information according to the second probability information; Obtaining optimization information of the first probability image according to the first optimization information and the second optimization information; Adjusting the pixel values ​​of the first probability image according to the optimization information to obtain the adversarial image; The first optimization information is calculated by the following formula: Among them, L black is the first optimization information, π0(i, j) represents the probability that the pixel value at the coordinate (i, j) in the first probability image is 1; The second optimization information is calculated by the following formula: Among them, L obj is the second optimization information, M is the number of second infrared images, is the second probability information of the target object existing in the i-th second infrared image, E t∈T represents mathematical expectation; The optimization information L of the first probability image is calculated by the following formula: L=L obj +λL black Among them, λ is the weight.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 4.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 4 is implemented.