Camouflage design method and system based on adversarial samples

Abstract

The invention relates to a camouflage design method and system based on an adversarial sample, and aims to solve the problem that a current camouflage pattern is difficult to adapt to a multi-terrain background camouflage demand, and the method comprises the steps: constructing a training data set; carrying out adversarial camouflage texture initialization by adopting a multi-scale adaptive disturbance weighted Voronoi graph; constructing a camouflage texture generation model based on the confrontation camouflage texture initial value; generating a camouflage target image in the physical scene based on the target mask image in combination with the original image and the camouflage texture output by the camouflage texture generation model, calculating an adversarial loss function through the adversarial image, and obtaining the camouflage color with the optimal effect by iteratively training the camouflage texture generation model; the method can effectively interfere with the recognition capability of the AI detection model, is suitable for the stealth technology, privacy protection, security and other scenes, and remarkably improves the efficiency and adaptability of the camouflage technology.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence camouflage design, in particular to a camouflage design method and system based on adversarial samples. BACKGROUND

[0002] With the rapid development of artificial intelligence (AI) technology, intelligent vision systems based on deep learning have been widely applied in target recognition, classification, tracking and other fields. These systems, with their efficient recognition capabilities, have shown important value in monitoring, autonomous driving and military reconnaissance. However, the progress of artificial intelligence technology has also brought new challenges, namely how to cope with the efficient recognition ability of AI detection models in privacy protection, security and other scenarios that need to avoid AI detection. Traditional camouflage design methods are mostly based on fixed backgrounds, usually relying on manual collection of natural background data, using color cards to collect the colors of various ground objects in the natural background, and then calculating the size and distribution of camouflage spots according to the theory of camouflage design. This method is extremely inconvenient, time-consuming and susceptible to human factors, and it is impossible to fully and objectively represent the complexity of the background during the collection process, resulting in significant differences in the designed camouflage patterns. Especially after the equipment is damaged, the existing automatically generated camouflage patterns are difficult to accurately blend with the actual damaged background, affecting the camouflage effect. SUMMARY

[0003] The present application aims to solve the above problems and provides a camouflage design method and system based on adversarial samples, which enhances the adaptability and robustness of camouflage technology, improves the concealment and adaptability, and optimizes the efficiency and effectiveness of camouflage design.

[0004] The technical solution to achieve the purpose of the present application is as follows:

[0005] A camouflage design method based on adversarial samples, comprising:

[0006] Constructing a training data set;

[0007] Using a multi-scale self-adaptive perturbation weighted Voronoi diagram to initialize adversarial camouflage texture;

[0008] Based on the initial value of the adversarial camouflage texture, a camouflage texture generation model is constructed;

[0009] The camouflage texture generation model is trained by the training data set. During the training process, based on the target mask image, the original image and the adversarial camouflage texture image output by the camouflage texture generation model, an adversarial image in the physical scene is generated. The loss function is calculated through the adversarial image to optimize the training of the camouflage texture generation model;

[0010] A rendered image with camouflage texture is generated by the trained camouflage texture generation model.

[0011] Further, the anti-camouflage texture initialization specifically comprises:

[0012] Step 2.1, setting a patch region control point set S and a prediction color set K;

[0013] S = {s1, c1, s2, c2,..., s i , c i ,...,s l , c l};

[0014] Wherein, s i = (x i , y i ) represents the coordinates of the i-th control point, c i ∈ {1, 2, 3,..., N class} represents the color category corresponding to the control point s i , and m is the color category supported by the camouflage pattern;

[0015]

[0016] Wherein, K c is the RGB value of the c-th color;

[0017] Step 2.2, calculating the scale adaptive weight;

[0018] Step 2.3, calculating the color probability distribution according to the scale adaptive weight, coloring each pixel to obtain an initial camouflage texture map.

[0019] Further, the scale adaptive weight is:

[0020]

[0021] Wherein, is a local scale weight, is a global scale weight, and ∈ is a set coefficient.

[0022] Further, the local scale weight is:

[0023]

[0024] Wherein, x is the position of the pixel point on the texture, w c (x) represents the dynamic weight of the pixel x to the color c, λ local is a decay coefficient, is random noise, and s j is the j-th element in the patch region control point set S.

[0025] Further, the local scale weight is:

[0026]

[0027] where G(x) is the edge intensity of the pixel, p is the gradient influence coefficient, and l is the decay coefficient. global

[0028] Further, the step 2.3 specifically includes:

[0029] Calculate the color probability distribution:

[0030]

[0031] where P c (x) is the normalized color probability, w c (x) is the scale adaptive weight.

[0032] For a pixel, select the color index with the maximum probability:

[0033]

[0034] According to the color index, assign the corresponding color K c (x), and obtain the initial camouflage texture map T adv .

[0035] Further, the camouflage texture generation model adopts an implicit field neural network, and the output surface texture is:

[0036] T=R(φ pos (u i , v i ), φ env (e), T adv );

[0037] where T is the generated camouflage texture target band rendering image, R is the camouflage texture generation model, φ env (e) is the environmental feature code extracted from the background image, T adv is the initial camouflage texture map, and φ pos (u i , v i ) is a high-dimensional vector after multi-scale position coding.

[0038] Further, the generated adversarial target image in the physical scene is:

[0039] X adv =m·T+(1-m)·X;

[0040] ​Wherein, m is a target target mask image, T is a target band rendering image with camouflage texture generated by the camouflage texture generation model, and X is an acquisition image.

[0041] Further, it also includes designing a loss function to optimize the training of the camouflage texture generation model; the loss function is:

[0042]

[0043] Wherein, a, β and γ are training stage weights, is an adversarial loss function, is a visual smoothing loss function, is a style transfer loss, specifically:

[0044]

[0045]

[0046] Wherein, is an IoU loss function, B p , B t are the predicted bounding box and the real bounding box respectively, is a target object confidence loss, C t is the confidence of the target detector prediction of the camouflage target, is a classification loss, K is the total number of categories, p k is the confidence of the camouflage target prediction, q k is the probability distribution of the predicted category, is the element of the i-th row and j-th column in the Gram matrix of the camouflage image, is the element of the i-th row and j-th column in the Gram matrix of the artificially designed camouflage image.

[0047] A camouflage design system based on adversarial samples, comprising:

[0048] A data set construction unit for constructing a training data set;

[0049] An adversarial camouflage texture initialization unit using a multi-scale adaptive perturbation weighted Voronoi diagram for adversarial camouflage texture initialization;

[0050] A camouflage texture generation model construction unit for constructing a camouflage texture generation model based on the initial value of the adversarial camouflage texture;

[0051] A neural renderer network unit for generating an adversarial image in a physical scene based on a target mask image, combining an original image and a camouflage texture image output by the camouflage texture generation model;

[0052] The loss function calculation unit calculates a loss function through the adversarial image, and is used for optimizing the training camouflage texture generation model;

[0053] The camouflage texture generation model training unit optimizes the training camouflage texture generation model through the training data set and the loss function;

[0054] The adversarial image generation unit generates a rendering image with camouflage texture through the trained camouflage texture generation model.

[0055] Compared with the prior art, the present application has the following advantages:

[0056] (1) Dynamic multi-terrain adaptability, the present application realizes the deep fusion of camouflage pattern and target multi-terrain background through the neural renderer combined with the adversarial sample generation technology, solves the problem that the existing camouflage pattern design is difficult to adapt to the multi-terrain environment, and ensures that the camouflage dynamic equipment has high visual concealment in complex scenes by using multi-scale self-adaptive perturbation weighted Voronoi diagram for adversarial camouflage texture initialization.

[0057] (2) Intelligent design and optimization, the present application uses a composite loss function (including adversarial loss, visual smoothing loss and style transfer loss) to guide the optimization of the camouflage pattern, realizes the automatic generation and iterative optimization of the camouflage texture through the deep learning technology, and significantly improves the intelligent level, calculation efficiency and robustness of the camouflage effect of the camouflage design.

[0058] (3) Wide applicability, the camouflage pattern designed by the present application is not only suitable for the concealment camouflage of defensive equipment, but also can be applied to privacy protection, security and other scenes that need to avoid AI detection, and the generated camouflage texture can effectively interfere with the target detection model based on deep learning, reducing the recognition success rate.

[0059] (4) Camouflage effect enhancement, the present application integrates bionics and adversarial sample characteristics into the camouflage texture, enhances the randomness and environmental adaptability of the camouflage pattern, and makes it still have high concealment ability in multi-angle, long-distance and complex scenes. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 The method flowchart of the present application.

[0061] Figure 2 The algorithm framework diagram of the present application.

[0062] Figure 3 The rendering image of the camouflage target with camouflage texture.

[0063] Figure 4 The effect diagram of the original painting target of the city detected by the yolov8 model.

[0064] Figure 5The yolov8 model is used for detecting the city original painting target heat map.

[0065] Figure 6 The yolov8 model is used for detecting the city camouflage painting target effect map.

[0066] Figure 7 The yolov8 model is used for detecting the city camouflage painting target heat map. DETAILED DESCRIPTION

[0067] This embodiment selects camouflage pattern for analysis against urban surveillance. Urban cameras can automatically detect targets driving on the road. Camouflage pattern can make surveillance cameras ineffective. Based on this, the embodiment provides a camouflage pattern design method based on adversarial samples, combining Figure 1 and Figure 2 , specifically including:

[0068] Step 1, making a data set.

[0069] Step 1.1, selecting typical background images of regions that may be involved in the target to be camouflaged, making a training data set (X, Y, θ c ). The data set contains image X, target label Y, and camera parameter θ c .

[0070] Step 1.2, making a target object mask data set. Use the large model SAM to obtain the target mask

[0071] m=U(X)

[0072] Where U(·) represents the large model SAM network function, and m is the target mask image.

[0073] Step 2, adversarial camouflage texture initialization, using multi-scale adaptive perturbation weighted Voronoi diagram to generate the initial map T adv of camouflage texture.

[0074] Step 2.1, set the patch region control point set S and the predicted color set K.

[0075] S={s1,c1,s2,c2,...,s i ,c i ,...,s l ,c l};

[0076] Where s i =(x i ,y i ) represents the coordinates of the i-th control point, c i ∈{1,2,3,...,Nclass} represents the control point corresponding point color category, a total of N class is the color class supported by the camouflage pattern;

[0077]

[0078] where K c is the RGB value of the cth color. In this fact instance, 5 colors are selected, including dark gray, light gray, dark green, dark green, brown, and black

[0079] Step 2.2, calculate the scale adaptive weight.

[0080] For a pixel point on the texture map, calculate its local scale weight to all control points:

[0081]

[0082] where x is the position of the pixel on the texture, w c (x) represents the dynamic weight of pixel x for color c, λ is the attenuation coefficient, used to control the distance to the weight of the influence speed, is random noise, so that the weight has disturbance in the local range, and increases the unpredictability.

[0083] For a pixel point on the texture map, calculate its global scale weight to all control points:

[0084]

[0085] where G(x) is the edge strength of the pixel, i.e. the gradient of the background image, and ρ controls the gradient influence coefficient.

[0086] Combine local and global scales to calculate dynamic weight:

[0087]

[0088] ∈ is a set coefficient. Through multiple experiments, by comparing the camouflage effects under different ∈ values, the optimal value of ∈ is finally determined to be 0.7

[0089] Step 2.3, patch area coloring. According to the dynamic weight, calculate the color probability distribution, and color each pixel. Calculate the color probability distribution:

[0090]

[0091] where P c (x) is the normalized color probability.

[0092] Carry out coloring processing based on probability. For a pixel, select the color index with the maximum probability:

[0093]

[0094] According to the color index, the corresponding color K is assigned c (x), to obtain the initial camouflage texture map T adv

[0095] Step 3, establish a camouflage texture generation model.

[0096] Step 3.1, target feature encoding. The three-dimensional target surface mesh is represented as a set of vertices, and the two-dimensional texture coordinates (u i , v i ) of each vertex are obtained through UV mapping, and the UV coordinates are multi-scale position encoded:

[0097] φ pos (u i , v i ) = [sin(2 0 πu i ), cos(2 0 πu i ),..., sin(2 n-1 πu i ), cos(2 n-1 πu i )]

[0098] wherein φ pos (u i , v i ) is the encoded high-dimensional vector.

[0099] Step 3.2, environment feature encoding. The environment feature extracted from the background image is encoded as φ env (e);

[0100] Step 3.3, establish a texture generation model R through an implicit field neural network, and calculate the target surface texture:

[0101] T = R(φ pos (u i , v i ), φ env (e), T adv )

[0102] wherein T is the generated camouflage target rendering image with rendering, used to simulate real-world texture mapping, and the generated adversarial camouflage target rendering image is as shown in Figure 3 .

[0103] Step 4, generate the camouflage target image in the physical scene by combining the original image and the camouflage pattern through the neural renderer network, use the conversion function Φ to generate the camouflage target image with smooth texture and realistic effect in the physical scene:

[0104] X adv = Φ(T) = m·T + (1 - M)·X

[0105] Step 5, train the texture generation model R to optimize the camouflage texture. In order to ensure that the camouflage algorithm can achieve the optimal effect in multiple dimensions, this study designs a composite loss function. The composite loss function combines adversarial loss, visual smoothness loss and style transfer loss, and reasonably balances the concealment and visual naturalness of camouflage.

[0106] Step 5.1, input the camouflage target image X adv to the target detector F to calculate the adversarial loss function.

[0107]

[0108] Where, IoU loss is

[0109]

[0110] Where, B t is the predicted bounding box, B k is the real bounding box;

[0111] The target object confidence loss is:

[0112]

[0113] Where, C k is the confidence of the target detector predicting the camouflage target;

[0114] The classification loss is:

[0115]

[0116] Where, K is the total number of categories, p is the confidence of the target detector predicting the camouflage target, q

[0001] is the probability distribution of the target detector predicting the category.

[0117] Step 5.2, calculate the visual smoothness loss function.

[0118]

[0119] Step 5.3, calculate the style transfer loss.

[0120]

[0121] wherein, is the i,j element of the Gram matrix of the camouflage pattern image. is the i,j element of the Gram matrix of the artificially designed camouflage pattern.

[0122] Step 5.4, calculating the total loss function

[0123]

[0124] wherein, a, b and g are training stage weights, respectively controlling the contribution of the adversarial loss, the visual smoothing loss and the style transfer loss to the total loss. In this embodiment, a = 0.6, b = 0.1 and g = 0.3. Through the optimization of this composite loss function, the camouflage algorithm can maintain concealment in multi-angle, long-distance and complex environments, while ensuring the naturalness and environmental adaptability of the camouflage texture.

[0125] Step 5.5: generating a new camouflage texture by gradient optimization

[0126]

[0127] Step 5.6: iterative training until the maximum training step N is reached, and outputting the camouflage texture T. To verify the effectiveness and robustness of the camouflage pattern design method based on adversarial samples of the present application, the following experimental verification is carried out. The experimental method is to generate camouflage textures on urban and grassland backgrounds respectively, and to design and generate corresponding camouflage patterns using the method of the present application; target detection model testing is carried out on the generated camouflage pattern target, a classic target detection model YOLOv8 is selected as the benchmark, and the detection accuracy and false detection rate are recorded. At the same time, a heat map is selected to interpret the effectiveness of the camouflage.

[0128] Table 1, experimental comparison data under different scenarios

[0129]

[0130] From the data in Table 1 and Figures 4-7 It can be seen that the camouflage pattern designed by the method of the present application significantly reduces the recognition accuracy of the target detection model, and significantly improves the false detection rate, confirming the effective interference ability of the camouflage pattern design method of the present application on the target detection model under different terrain backgrounds.

[0131] The present embodiment also provides a camouflage pattern design system, which aims to solve the problem that current camouflage patterns are difficult to adapt to multi-terrain background camouflage requirements, comprising:

[0132] A data set construction unit for constructing a training data set;

[0133] The adversarial camouflage texture initialization unit uses a multi-scale adaptive perturbation weighted Voronoi diagram to initialize the adversarial camouflage texture.

[0134] The camouflage texture generation model construction unit constructs a camouflage texture generation model based on the initial value of the adversarial camouflage texture.

[0135] The neural renderer network unit generates an adversarial image in a physical scene based on a target mask image, combined with an original image and a camouflage texture image output by the camouflage texture generation model.

[0136] The loss function calculation unit calculates a loss function through the adversarial image, which is used to optimize and train the camouflage texture generation model.

[0137] The camouflage texture generation model training unit optimizes and trains the camouflage texture generation model through a training data set and a loss function.

[0138] The adversarial target image generation unit generates a target rendering image with camouflage texture through the trained camouflage texture generation model.

[0139] The method uses a neural renderer to map 3D equipment models and their textures to 2D images, and uses deep learning to generate camouflage patterns that blend with the background depth. Through adversarial sample technology, the recognition ability of AI detection models is disturbed, and the adaptability and robustness of camouflage technology are enhanced. Thus, while improving concealment and adaptability, the efficiency and effectiveness of camouflage design are optimized. This innovative design method is widely used in privacy protection, security, and other scenarios that need to evade AI detection, providing a new idea for the future development of camouflage technology. The invention provides a new technical path for camouflage pattern design based on adversarial samples.

[0140] The invention can effectively interfere with the recognition ability of AI detection models, and is suitable for stealth technology, privacy protection, security, and other scenarios, significantly improving the efficiency and adaptability of camouflage technology.

[0141] The embodiments of the invention are described in detail above in conjunction with the drawings, but the invention is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the spirit of the invention.

Claims

1. A camouflage design method based on adversarial samples, characterized in that: include: Build a training dataset; Adopting multi-scale adaptive perturbation weighted Voronoi diagram for adversarial camouflage texture initialization; Based on the initial value of the adversarial camouflage texture, a camouflage texture generation model is constructed; The camouflage texture generation model is trained using a training dataset. During the training process, adversarial images in the physical scene are generated based on the target mask image, the original image, and the camouflaged texture image output by the camouflage texture generation model. The loss function is calculated using the adversarial images to optimize the training of the camouflage texture generation model. Generate rendered images with camouflage textures through the trained camouflage texture generation model.

2. The camouflage design method based on adversarial samples according to claim 1, characterized in that: Initializing the anti-camouflage texture specifically includes: Step 2.1, set the patch area control point set S and the predicted color set K; S={s1,c1,s2,c2,...,s i ,c i ,...,s l ,c l }; Among them, s i =(x i ,y i ) represents the coordinates of the i-th control point, c i ∈{1, 2, 3, ..., N class } represents the control point s i The corresponding point color category, m is the color category supported by the camouflage pattern; Among them, K c is the RGB value of the cth color; Step 2.2, calculate scale adaptive weight; In step 2.3, the color probability distribution is calculated according to the scale-adaptive weight, and each pixel is colored to obtain the initial camouflage texture map.

3. The camouflage design method based on adversarial samples according to claim 2, characterized in that: The scale adaptation weight is: in, is the local scale weight, is the global scale weight, ∈ is the setting coefficient.

4. The camouflage design method based on adversarial samples according to claim 3 is characterized in that: The local scale weight is: Among them, x is the pixel position on the texture, w c (x) represents the dynamic weight of pixel x to color c, λ local is the attenuation coefficient, is random noise, s j It is the jth element in the patch area control point set S.

5. The camouflage design method based on adversarial samples according to claim 3 is characterized in that: The local scale weight is: Among them, G(x) is the edge strength of the pixel, ρ controls the gradient influence coefficient, and λ global is the attenuation coefficient.

6. The camouflage design method based on adversarial samples according to claim 2, characterized in that: The step 2.3 specifically includes: Compute the color probability distribution: Among them, P c (x) is the normalized color probability, w c (x) is the scale-adaptive weight; For a pixel, choose the color index with the maximum probability: Then assign the corresponding color K according to the color index c (x), get the initial camouflage texture map T adv .

7. The camouflage design method based on adversarial samples according to claim 1, characterized in that: The camouflage texture generation model adopts an implicit field neural network, and the output surface texture is: T=R(φ pos (u i ,v i ),φ env (e),T adv ); Among them, T is the generated camouflage texture and rendering image, R is the camouflage texture generation model, φ env (e) Environmental feature encoding extracted from the background image, T adv is the initial camouflage texture map, φ pos (u i , v i ) is a high-dimensional vector after multi-scale position encoding.

8. The camouflage design method based on adversarial samples according to claim 1, characterized in that: Generate adversarial images in physical scenes as: X adv =m·T+(1-m)·X; Where m is the target mask image, T is the rendered image with camouflage texture generated by the camouflage texture generation model, and X is the captured image.

9. The camouflage design method based on adversarial samples according to claim 1, characterized in that: The loss function is: Among them, α, β and γ are weights that are dynamically adjusted during the training phase. To counter the loss function, is the visual smoothness loss function, is the style transfer loss, specifically: Among them, α0, β0 and γ0 are set initial values, is the IoU loss function, B p 、B t are the predicted bounding box and the true bounding box, respectively. is the target object confidence loss, C t is the confidence of the target detector in predicting the disguised target, is the classification loss, K is the total number of categories, and p k is the confidence of the camouflaged target prediction, q k is the probability distribution of the predicted category, is the element in row i and column j of the Gram matrix of the camouflage image, is the element in row i and column j of the Gram matrix of the artificially designed camouflage image.

10. A camouflage design system for implementing the method according to any one of claims 1 to 9, characterized in that: include: A data set construction unit, used to construct a training data set; The adversarial camouflage texture initialization unit uses a multi-scale adaptive perturbation weighted Voronoi diagram to initialize the adversarial camouflage texture; A camouflage texture generation model construction unit, which constructs a camouflage texture generation model based on an initial value of the adversarial camouflage texture; The neural renderer network unit generates an adversarial image in the physical scene based on the target mask image, the original image, and the camouflaged texture image output by the camouflaged texture generation model; The loss function calculation unit calculates the loss function through the adversarial image, which is used to optimize the training of the camouflage texture generation model; A camouflage texture generation model training unit, which trains the camouflage texture generation model through training data sets and loss function optimization; The adversarial image generation unit generates a rendered image with camouflage texture through the trained camouflage texture generation model.

Images