Intelligent sensing interference method based on color trapping

Abstract

The invention discloses an intelligent sensing interference method based on color trapping. The method comprises the steps of image acquisition and preprocessing, color feature extraction and analysis, color-based adversarial patch design, adversarial patch generation and application, and effect evaluation and adjustment. Wherein the image acquisition and preprocessing part is used for acquiring and preprocessing data; the color feature extraction and analysis part is used for performing color multi-angle feature extraction and analysis on the acquired image; according to the color-based adversarial patch design part, adversarial patches of different sizes, shapes and colors are designed according to judgment of key positions; generating an adversarial patch and applying part, and deploying and applying the designed patch to a key part of the vehicle; and the effect evaluation and adjustment part is used for re-shooting the vehicle after the adversarial patch is deployed by using the unmanned aerial vehicle and sending the vehicle into the detection model, so that the effectiveness of the color-trapped adversarial patch is evaluated, and the design scheme is adjusted according to a detection result.

Description

Technical Field

[0001] The present application relates to the technical field of deep learning intelligent perception interference algorithms, and in particular to an intelligent perception interference method based on color trapping. Background Art

[0002] The goal of the intelligent perturbation algorithm is to create adversarial samples, which are extremely small perturbations or disturbances added to the original, uncontaminated images. These perturbations are almost imperceptible to human vision and therefore do not affect human recognition. However, these subtle changes can interfere with machine learning models and cause them to make incorrect judgments or predictions.

[0003] There are many intelligent perceptual interference methods for generating adversarial samples, each with its own characteristics. They can be mainly divided into three categories: gradient-based methods, optimization-based methods, and generative model-based methods.

[0004] First, gradient-based methods mainly rely on the gradient information of the model to construct adversarial perturbations. This method is direct and efficient because it directly uses the gradient changes within the model to locate and generate perturbations that can mislead the model. However, it may also be limited by the model structure. For some complex models or structures, it may not be able to accurately capture enough gradient information to generate effective adversarial samples.

[0005] Secondly, optimization-based methods use an iterative optimization process to find perturbations that minimize the model's prediction error. This method is more flexible because it does not depend on the specific structure of the model, but it iteratively adjusts the perturbation until the minimum perturbation that can mislead the model is found. However, this method is relatively computationally expensive because it requires multiple iterations and optimization processes.

[0006] Finally, the generative model-based method uses models such as generative adversarial networks (GANs) to generate adversarial samples. This method can generate diverse adversarial samples because models such as GANs can learn the distribution of original data and generate new data samples. However, the training process of this method is relatively complex and requires a large amount of data and computing resources to train the generative model.

[0007] These three methods have their own advantages and disadvantages and are suitable for different scenarios and needs. However, they all basically follow the idea of ​​updating based on network models and gradient backpropagation in deep learning. This idea allows the goal of generating adversarial samples to be achieved by simply designing new loss terms according to needs. However, this idea also has obvious disadvantages, because methods based on deep neural networks generally require a large amount of data, strong computing resources, and more training time. Summary of the invention

[0008] The present application provides an intelligent perception jamming method based on color trapping, which has three advantages. First, through the fine analysis and optimization of colors, the adversarial patch can more effectively mislead the detection system and improve the success rate of jamming. Second, the high degree of color fusion makes the adversarial patch visually difficult to detect, which improves the concealment of the jamming attack. Third, the method does not rely on complex gradient learning, which reduces the difficulty of implementation and computational cost.

[0009] In a first aspect, a method for intelligent perception interference based on color trapping is provided, comprising:

[0010] Image acquisition and preprocessing: Acquire target images taken from the air; preprocess the images, including color correction and denoising;

[0011] Color feature extraction and analysis: Use image processing technology to extract color features in images, including color distribution, hue, and saturation; analyze the extracted color features to find out the rules that match the target color and background color;

[0012] Complete the color-based adversarial patch design: Based on the color feature analysis results, design the color scheme of the adversarial patch so that the patch is visually integrated with the original image;

[0013] Generate and apply adversarial patches: Generate adversarial patches based on the color scheme of the adversarial patches; paste the generated adversarial patches on the target surface, and the color of the patches is coordinated with the surrounding environment.

[0014] In combination with the first aspect, in some implementations of the first aspect, the method further includes:

[0015] Effect evaluation and adjustment: Use air-based ground-to-ground photography equipment to photograph the target with the adversarial patch attached, and input the captured image into the target detection model for testing; based on the test results, evaluate the interference effect of the adversarial patch; based on the evaluation results, adjust the color scheme of the adversarial patch to improve the interference efficiency.

[0016] In conjunction with the first aspect, in some implementations of the first aspect, preprocessing the image includes at least one of the following:

[0017] Convert the image from RGB color space to HSV color space to separate hue, saturation, and brightness information;

[0018] Apply Gaussian filter to smooth the image and remove noise;

[0019] Adjust the contrast and brightness of images to ensure accurate and consistent color information.

[0020] In combination with the first aspect, in some implementations of the first aspect, color feature extraction includes:

[0021] Divide the image into HSV channels H, S, and V;

[0022] A color histogram is calculated for each channel, recording the frequency of occurrence of each color component.

[0023] In combination with the first aspect, in certain implementations of the first aspect, the color feature analysis includes color feature calculation and color difference analysis;

[0024] Color feature calculation includes: calculating the average hue, average saturation and average brightness value of the image as a description of the overall color characteristics; and calculating the variance and standard deviation of the color distribution to evaluate the degree of color dispersion in the image;

[0025] Color difference analysis includes: analyzing the differences in hue and saturation between the target and the background; and determining the key color components used to distinguish the target from the background.

[0026] In conjunction with the first aspect, in certain implementations of the first aspect, determining a key color component for distinguishing a target from a background includes:

[0027] Calculate the hue difference H between the target and the background in the HSV space respectively dis and saturation difference S dis ,

[0028] Hue Difference H dis :H dis =|H objective_mean -H back_mean |,H objective_mean The target color, H back_mean For the background color,

[0029] Saturation difference S dis :S dis =|S objective_mean -S back_mean |,S objective_mean is the saturation of the target, S back_mean is the saturation of the background;

[0030] By comparing the hue difference H dis and saturation difference S dis The size of H determines the key color component for distinguishing the target from the background. dis >S dis , then hue is the key color component; if H dis dis , then saturation is the key color component.

[0031] ​In combination with the first aspect, in certain implementations of the first aspect, the color-based adversarial patch design includes:

[0032] Base color selection: Based on the color difference analysis results, a color similar to the target color is selected as the base color of the adversarial patch;

[0033] Patch shape and size design: Design the shape and size of the adversarial patch based on the key feature area of ​​the target to ensure that the adversarial patch can fully cover the key feature area without affecting the normal driving and appearance of the target;

[0034] Color fusion processing: Use color mixing technology to blend the base color of the anti-patch with the color of the surrounding environment.

[0035] In conjunction with the first aspect, in some implementations of the first aspect, determining similar colors includes:

[0036] Calculate the color mean of the H, S, and V channels of the target image sub-region to obtain the three-dimensional feature vector HSV mean =[H objective_mean ,S objective_mean ,V objective_mean ], H objective_mean is the target hue average, S objective_mean is the average saturation value of the target, V objective_mean is the average brightness of the target;

[0037] HSV mean The average value of the key color component in is replaced by the difference value of the key color component, and we get And replace the As base color.

[0038] In combination with the first aspect, in some implementations of the first aspect, the color fusion processing includes:

[0039] Select multiple sample points from the area around the adversarial patch, and calculate the average value of these sample points in the HSV color space as the representative color of the background. The average color value of the sample points is recorded as B = [H b ,S b ,V b ], H b is the average hue of the sample points, S b is the average saturation value of the sample points, V b is the average brightness of the sample points;

[0040] Color fusion based on transparency; introduce transparency parameter α, 0≤α≤1, and use linear difference to transform the base color Blend with the background color B; the fused color is F = [H f ,Sf ,V f ]for:

[0041] H f =(1-α)H b +αH vehicle_mean

[0042] S f =(1-α)S b +αS vehicle_mean

[0043] V f =(1-α)V b +αV vehicle_mean

[0044] H f is the hue value of the sample point after fusion, S f is the saturation value of the sample point after fusion, V f is the brightness value of the sample point after fusion.

[0045] In a second aspect, an electronic device is provided, wherein the electronic device is used to execute the method described in any one of the implementations of the first aspect.

[0046] Compared with the prior art, the solution provided by this application includes at least the following beneficial technical effects:

[0047] First, through the detailed analysis and optimization of colors, the adversarial patch can more effectively mislead the detection system and improve the success rate of deception. Second, the high degree of color integration makes the adversarial patch visually difficult to detect, which improves the concealment of the deception attack. Third, this method does not rely on complex gradient learning, which reduces the difficulty of implementation and computational cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 The figure is a flow chart of an intelligent perception jamming method based on color trapping. DETAILED DESCRIPTION

[0049] The present application is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] The present invention provides an intelligent perceptual interference method based on color trapping. This method emphasizes the role of color in adversarial sample design. By finely analyzing the color of the entire image, an adversarial patch is generated that is highly integrated with the background but can mislead the detection system (the adversarial patch is a special case of the adversarial sample). The overall block diagram is as follows Figure 1 shown.

[0051] Step 1: Image acquisition and preprocessing.

[0052] Image acquisition can refer to the use of air-based equipment such as drones or satellites to photograph ground vehicles at different times and under different lighting conditions to ensure the diversity of image samples. The acquired images should contain clear vehicle outlines, colors and background information, and contain rich color information.

[0053] Preprocessing may include at least one of the following: converting the image from RGB color space to HSV color space to better separate hue, saturation and brightness information; applying a Gaussian filter to smooth the image to remove noise and irregularities in details; adjusting the contrast and brightness of the image to ensure the accuracy and consistency of color information.

[0054] Step 2: Color feature extraction and analysis.

[0055] Color feature extraction may include color histogram extraction. Color histogram extraction is to divide the image into multiple color channels (such as HSV H, S, V channels). The color histogram is calculated for each channel, and the frequency of occurrence of each color component is recorded.

[0056] Color feature analysis can include color feature calculation and color difference analysis. Color feature calculation includes calculating the average hue, saturation, and brightness values ​​of the image as a description of the overall color features; and calculating the variance and standard deviation of the color distribution to evaluate the degree of color dispersion in the image. Color difference analysis includes analyzing the color difference between the target vehicle and the background, especially the difference in hue and saturation.

[0057] Determine which color components are most critical for distinguishing between a vehicle and a background. An implementation of determining the key color components provided by the present invention is as follows.

[0058] (1) Calculate the average values ​​of the hue (H) difference and saturation (S) difference between the target vehicle and the background in the HSV space respectively.

[0059] Hue difference: H dis =|H objective_mean -H back_mean |, H objective_mean is the average hue of the target vehicle, H back_mean is the average hue of the background.

[0060] Saturation difference: S dis =|S objective_mean -S back_mean |,S objective_mean is the average saturation value of the target vehicle, S back_mean is the average saturation of the background.

[0061] (2) Determine the key color component. By comparing the hue difference and saturation difference, determine the color component that is most critical for distinguishing the vehicle from the background. dis is larger, then hue is the key color component; if S dis Larger, saturation is the key color component.

[0062] Step 3: Color-based adversarial patch design, including base color selection, patch shape and size design, and color fusion processing.

[0063] The base color selection may refer to selecting a color similar to the color of the target vehicle as the base color of the adversarial patch according to the color difference analysis result. The similar color is determined as follows.

[0064] (1) Calculate the color mean of the H, S, and V channels of the image sub-region where the target vehicle is located, and concatenate the three means into a three-dimensional feature vector, denoted as HSV mean =[H vehicle_mean ,S vehicle_mean ,V vehicle_mean ], H objective_mean is the average hue of the target vehicle, S objective_mean is the average saturation value of the target vehicle, V objective_mean is the average brightness of the target vehicle.

[0065] (2) HSV mean The average value of the key color component in is replaced by the difference value of the key color component, and we get If the key color component is hue, then H objective_mean Replace with H dis ; If the key color component is saturation, then S back_mean Replace with S dis ;

[0066] (3) Replace the As base color.

[0067] Patch shape and size design can refer to designing the shape and size of the countermeasure patch according to the key feature areas of the target vehicle (such as the roof, windows, etc.), ensuring that the countermeasure patch can completely cover the key feature areas without affecting the normal driving and appearance of the vehicle.

[0068] Color fusion processing may refer to using a color mixing technique to blend the base color of the confrontation patch with the color of the surrounding environment. A transparency adjustment method may be used to make the patch more coordinated with the background. The fusion method is as follows.

[0069] (1) Select multiple sample points from the area around the adversarial patch and calculate the average value of these sample points in the HSV color space as the representative color of the background. The average color value of the sample points is recorded as B = [H b ,S b ,V b ], H b is the average hue of the sample points, S b is the average saturation value of the sample points, V b is the average brightness of the sample points.

[0070] (2) Color fusion based on transparency. Introduce the transparency parameter α, 0≤α≤1, and use linear difference to transform the base color Blend with the background color B; the fused color is F = [H f ,S f ,V f ]for:

[0071] H f =(1-α)H b +αH vehicle_mean

[0072] S f =(1-α)S b +αS vehicle_mean

[0073] V f =(1-α)V b +αV vehicle_mean

[0074] H f is the hue value of the sample point after fusion, S f is the saturation value of the sample point after fusion, V f is the brightness value of the sample point after fusion.

[0075] Step 4: Adversarial patch generation and application.

[0076] Patch generation. Use image processing software (such as Photoshop, GIMP, etc.) or professional printing equipment to generate physical or digital patterns of adversarial patches according to the designed color scheme and shape size. Ensure that the generated adversarial patches have sufficient accuracy and resolution to ensure their effectiveness in practical applications.

[0077] Patch application. Select appropriate adhesive materials (such as tape, spray paint, etc.) to ensure that the anti-patch can be firmly attached to the vehicle surface. According to the designed shape and size, select the appropriate position on the vehicle surface and stick the anti-patch on.

[0078] Step 5: Effect evaluation and adjustment.

[0079] Test and Evaluation. Use airborne ground-based photography equipment to retake vehicle images with adversarial patches pasted on them. Input these images into the vehicle detection model for testing and record the performance changes of the detection model.

[0080] Performance analysis. Analyze the changes in detection accuracy, false positive rate and other indicators of the detection model before and after the adversarial patch is applied. Evaluate the interference effect of the adversarial patch and determine its applicability in different scenarios and conditions.

[0081] Adjustment and optimization. According to the evaluation results, adjust the color, shape, size, and position of the adversarial patch. Repeat steps 4 and 5 until a satisfactory interference effect is achieved.

[0082] The intelligent perception interference method based on color trapping provided by the present invention has good physical meaning, strong interpretability, low calculation complexity, simple implementation method, and good interference effect.

[0083] Although the present invention is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be based on the scope defined by the claims of the present invention.

Claims

1. An intelligent perception interference method based on color trapping, characterized in that: include: Image acquisition and preprocessing: Acquire target images taken from the air; Preprocess the image, including color correction and denoising; Color feature extraction and analysis: Use image processing technology to extract color features in images, including color distribution, hue, and saturation; Analyze the extracted color features to find the rules that match the target color and background color; Complete the color-based adversarial patch design: Based on the color feature analysis results, design the color scheme of the adversarial patch so that the patch is visually integrated with the original image; Generate and apply adversarial patches: Generate adversarial patches based on the color scheme of the adversarial patches; paste the generated adversarial patches on the target surface, and the color of the patches is coordinated with the surrounding environment.

2. The method according to claim 1, characterized in that The method further comprises: Effect evaluation and adjustment: Use air-based ground-to-ground photography equipment to photograph the target with the adversarial patch attached, and input the captured image into the target detection model for testing; based on the test results, evaluate the interference effect of the adversarial patch; based on the evaluation results, adjust the color scheme of the adversarial patch to improve the interference efficiency.

3. The method according to claim 1, characterized in that Preprocessing the image includes at least one of the following: Convert the image from RGB color space to HSV color space to separate hue, saturation, and brightness information; Apply Gaussian filter to smooth the image and remove noise; Adjust the contrast and brightness of images to ensure accurate and consistent color information.

4. The method according to claim 1, characterized in that: Color feature extraction includes: Divide the image into HSV channels H, S, and V; A color histogram is calculated for each channel, recording the frequency of occurrence of each color component.

5. The method according to claim 1, characterized in that Color feature analysis includes color feature calculation and color difference analysis; Color feature calculation includes: calculating the average hue, average saturation and average brightness value of the image as a description of the overall color characteristics; and calculating the variance and standard deviation of the color distribution to evaluate the degree of color dispersion in the image; Color difference analysis includes: analyzing the differences in hue and saturation between the target and the background; and determining the key color components used to distinguish the target from the background.

6. The method according to claim 5, characterized in that Identify key color components used to distinguish target from background, including: Calculate the hue difference H between the target and the background in the HSV space respectively dis and saturation difference S dis , Hue Difference H dis :H dis =|H objective_mean -H back_mean |, H objective_mean The target color, H back_mean For the background color, Saturation difference S dis :S dis =|S objective_mean -S back_mean |,S objective_mean is the saturation of the target, S back_mean is the saturation of the background; By comparing the hue difference H dis and saturation difference S dis The size of H determines the key color component for distinguishing the target from the background. dis >S dis , then hue is the key color component; if H dis dis , then saturation is the key color component.​ 7. The method according to claim 1, characterized in that Color-based adversarial patch design, including: Base color selection: Based on the color difference analysis results, a color similar to the target color is selected as the base color of the adversarial patch; Patch shape and size design: Design the shape and size of the adversarial patch based on the key feature area of ​​the target to ensure that the adversarial patch can fully cover the key feature area without affecting the normal driving and appearance of the target; Color fusion processing: Use color mixing technology to blend the base color of the anti-patch with the color of the surrounding environment.

8. The method according to claim 7, characterized in that The determination of similar colors includes: Calculate the color mean of the H, S, and V channels of the target image sub-region to obtain the three-dimensional feature vector HSV mean =[H objective_mean ,S objective_mean ,V objective_mean ], H objective_mean is the target hue average, S objective_mean is the average saturation value of the target, V objective_mean is the average brightness of the target; HSV mean The average value of the key color component in is replaced by the difference value of the key color component, and we get And replace the As base color.

9. The method according to claim 8, characterized in that Color fusion processing includes: Select multiple sample points from the area around the adversarial patch, and calculate the average value of these sample points in the HSV color space as the representative color of the background. The average color value of the sample points is recorded as B = [H b ,S b ,V b ], H b is the average hue of the sample points, S b is the average saturation value of the sample points, V b is the average brightness of the sample points; Color fusion based on transparency; introduce transparency parameter α, 0≤α≤1, and use linear difference to transform the base color Blend with the background color B; the fused color is F = [H f ,S f ,V f ]for: H f =(1-α)H b +αH vehicle_mean S f =(1-α)S b +αS vehicle_mean V f =(1-α)V b +αV vehicle_mean H f is the hue value of the sample point after fusion, S f is the saturation value of the sample point after fusion, V f is the brightness value of the sample point after fusion.

10. An electronic device, characterized in that: The electronic device is configured to execute the method according to any one of claims 1 to 9.

Images