Anti-unmanned aerial vehicle and anti-face recognition pattern generation method and system based on bionic camouflage

Abstract

The invention belongs to the technical field of camouflage, and particularly discloses an anti-unmanned aerial vehicle and anti-face recognition pattern generation method and system based on bionic camouflage, and the method comprises the steps: selecting a target environment background picture, carrying out the color space conversion and mixed color processing, extracting background textures and bionic camouflage textures, carrying out the superposition and synthesis after visual illusion deformation, and carrying out the recognition of an anti-unmanned aerial vehicle and an anti-face recognition pattern. Performing color filling and digital camouflage design to generate a bionic camouflage pattern; meanwhile, a labeling data set is constructed, a YOLOv5l model is combined with an improved gradient CAM algorithm to carry out adversarial training, and adversarial samples enabling the accuracy of the target recognition model to be reduced by more than or equal to 40% are generated; and the system has a dynamic updating capability. Compared with the prior art, by means of the synergistic effect of bionic texture and confrontation training, efficient interference on unmanned aerial vehicle visual recognition and camera face recognition is achieved, and the method has the capacity of environment self-adaptive generation and dynamic confrontation upgrading and can be widely applied to camouflage scenes of equipment, personnel and important targets.

Description

Technical Field

[0001] This invention belongs to the field of camouflage technology, specifically relating to a method and system for generating anti-drone and anti-facial recognition patterns based on biomimetic camouflage. Background Technology

[0002] As is well known, drone technology is widely used in various fields, such as geographic surveying, environmental monitoring, logistics delivery, security patrol, and military applications. With the continuous improvement of drone performance, it possesses powerful information collection and monitoring capabilities during missions. However, this also brings new security challenges. For example, unauthorized drones may illegally spy on private areas, threatening personal privacy and public safety. Therefore, researching counter-drone technology, especially reducing drone identification efficiency through camouflage, has become an important direction for ensuring security. Currently, there are many technologies both domestically and internationally in the field of counter-AI reconnaissance. For example, the "anti-face recognition" algorithm developed by the University of Toronto significantly reduces the accuracy of face recognition; Russia's Yandex company has launched an algorithm that interferes with facial recognition through makeup; Carnegie Mellon University in the United States has developed glasses that can fool facial recognition systems; KU Leuven in Belgium has developed stickers for countering drone identification; and Wuhan University in China has designed an "invisdefense cloak" that can disable machine vision. However, these technologies are not very mature, making large-scale application difficult, and their applicable scenarios are limited, failing to meet diverse practical needs. In particular, when dealing with AI identification by drones and cameras, there is a lack of camouflage schemes that can be intelligently generated according to different environments. Traditional camouflage pattern designs, lacking biomimetic camouflage textures and effective pattern countermeasure training, have very limited interference effects on artificial intelligence recognition systems and are difficult to resist advanced intelligent recognition technologies. Current anti-facial recognition technologies are mainly divided into three categories: physical interference, digital adversarial techniques, and hybrid enhancement. Physical interference methods, such as special glasses or coatings, are simple to operate but can severely impact the user experience. They interfere with the camera's ability to capture clear facial images by reflecting infrared light or creating optical noise. Digital adversarial techniques, like the Fawkes tool, inject micro-perturbations invisible to the human eye into digital images, causing the feature vectors extracted by AI models to deviate from the true values. Although the iterative algorithm in 2024 improved the error rate of mainstream recognition models such as FaceNet to 70%, its generalization ability is insufficient, making it difficult to function stably in different scenarios and models. Hybrid enhancement strategies, such as biometric obfuscation techniques, cover real biometric features by dynamically modifying facial key points or generating "virtual faces." While they have some effect, they pose a risk of triggering deepfake detection systems and are also subject to relevant regulatory restrictions. Furthermore, existing technologies face several challenges: First, they exhibit poor robustness against attacks. Most perturbation algorithms are designed for specific recognition models, and their effectiveness diminishes significantly when the target system upgrades to a multi-model integrated architecture. Moreover, facial recognition systems can quickly adapt to new attack patterns through online updates, while anti-recognition tools update slowly and cannot respond promptly. Second, practicality and user experience are difficult to reconcile. Strong perturbations can cause significant image distortion, affecting visual perception. In real-time video stream processing, the anti-recognition processing latency for high-definition video is too high, failing to meet the real-time requirements of scenarios such as security monitoring. Third, as technological countermeasures continue to evolve, multimodal recognition technologies integrate gait, voiceprint, and other features, rendering single facial interference methods increasingly ineffective. Simultaneously, the adversarial sample detection modules deployed by the defender can efficiently identify various perturbation images. Therefore, a more efficient and adaptable anti-drone and anti-facial recognition technology is urgently needed.

[0003] In view of this, this invention is hereby proposed. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for generating anti-drone and anti-face recognition patterns based on biomimetic camouflage. It is mainly used to solve the problems of poor camouflage effect, weak environmental adaptability and insufficient robustness in existing anti-artificial intelligence reconnaissance technologies, and to achieve efficient interference with drone visual recognition and camera face recognition systems, thereby reducing the probability of the target being recognized by intelligent devices.

[0005] The objective of this invention is achieved through the following technical solution: In a first aspect, the present invention provides a method for generating anti-drone and anti-facial recognition patterns based on biomimetic camouflage, comprising the following steps: Step 1: Background processing: Select the background image of the target environment, perform color space conversion and noise reduction, and then use the SLIC superpixel segmentation algorithm to segment the processed image and extract the main background color; Step 2, Texture Blending: Extract the background texture outline, simultaneously extract the local and overall features of the biomimetic camouflage texture, and then combine the texture with the background texture after texture deformation based on the principle of visual illusion. Step 3, Camouflage Generation: Fill the background with the main color of the fused pattern, generate a digital camouflage base pattern by calculating the size of the camouflage unit, and add grain noise to form the final biomimetic camouflage pattern; Step 4, Adversarial Enhancement: Construct a dataset with labeled images, use the YOLOv5l model combined with the improved gradient CAM algorithm for adversarial training, and generate adversarial samples that reduce the accuracy of the target recognition model by ≥40%.

[0006] Furthermore, in step 1, the color space conversion adopts the Lab color space, and when the processed image is segmented using the SLIC superpixel segmentation algorithm, the segmented image sizes are 50 and 100.

[0007] Specifically, step 1, the noise reduction operation, includes the following processes: 1) Image preprocessing: Convert the original image to Lab color space (preserve the luminance L channel and separate the color information a and b channels); 2) Parameter initialization: Set the number of superpixel targets (e.g., 50 / 100 for a woodland scene) and the compactness factor (recommended value 15-25); 3) Pixel clustering: Execute the SLIC algorithm to calculate pixel similarity in 5-dimensional space (x,y,L,a,b) and iteratively optimize the superpixel boundaries (default 10 iterations); 4) Region merging: Merge discrete superpixel regions with an area less than 0.5% of the total image area; 5) Color smoothing: Perform median filtering (3×3 kernels) on each superpixel to eliminate noise.

[0008] Step 1, background primary color extraction, includes the following processes: 1) Superpixel sampling: Extract the RGB mean of each superpixel region after SLIC segmentation; 2) Color clustering: Use the K-means algorithm to cluster the sampled colors (K=5 is recommended for woodland scenes); 3) Primary color determination: Select the background primary color according to the following rules: the proportion of superpixels covered is >15%, the ΔE of other colors in CIELAB space is >12, and it conforms to the typical hue of the environment (e.g., yellow-green is preferred for woodland); 4) Weight optimization: Assign weights to the selected primary color according to the coverage area ratio (the sum is normalized to 100%).

[0009] Furthermore, in step 2, the biomimetic camouflage texture includes animal textures and plant textures; The animal texture is selected from zebra texture, and the plant texture is selected from poplar leaf texture.

[0010] Furthermore, the extraction of the biomimetic camouflage texture includes: extracting the local stripe contours of the zebra pattern and extracting the overall vein contours of the poplar leaf pattern.

[0011] Specifically, the implementation method and fusion algorithm of texture deformation and background texture superposition synthesis based on the principle of visual illusion in step 2 are as follows: 1) Contour feature extraction: Local and global contour extraction of biomimetic objects (such as zebras and poplar leaves) to obtain original texture features; 2) Visual illusion deformation rules: Apply geometric transformations (such as scaling, rotation, and distortion) to break the regularity of texture; introduce dynamic gradient contrast adjustment to enhance the blurring effect perceived by the human eye; 3) Dynamic adaptive adjustment: Adaptively adjust the deformation intensity according to the complexity of the background texture (such as the jagged processing of poplar leaf texture in forest scenes).

[0012] Algorithm example: Non-linear mapping of texture blocks using affine transformation matrices: The coefficients a−d control the deformation amplitude, and e−f achieves translation compensation. Background texture overlay synthesis. Fusion strategy: 1. Multi-scale texture alignment: The background texture and the deformed biomimetic texture are decomposed by the Laplacian pyramid and then fused in the frequency domain layer by layer.

[0013] 2. Dynamic weight allocation: Define the fusion coefficient E max Emax is the energy value of the background area, and Emax is the global maximum energy value, so that high-energy areas (such as woodland shadows) are given priority to preserve background features.

[0014] Specific steps: 1) Apply SLIC superpixel segmentation to obtain the main color distribution of the background texture; 2) Achieve seamless fusion of the deformed texture and the background through the Poisson fusion equation ∇⋅(D∇u)=0, where D is the diffusion coefficient tensor, controlling the gradient propagation direction; 3) Introduce a contrast-sensitive function: This balances the contrast competition between the biomimetic texture and the background.

[0015] The advantages include the following two aspects: First, the visual illusion enhancement mechanism: by introducing the Mach band effect through nonlinear deformation, the texture boundary perceived by the human eye is blurred; Second, the adaptive fusion algorithm: by combining frequency domain analysis and energy weight allocation, the robustness of cross-scale texture fusion is achieved.

[0016] Specifically, the noise generation rules in step 3 (based on environment adaptation and adversarial training) are as follows: 1) Noise type---Adopting a Gaussian-Poisson mixture noise model: Gaussian component (μ=0, σ²=0.01) simulates the randomness of natural texture; Poisson component enhances the high-frequency detail interference effect; 2) Density control---Dynamically adjusted according to environment type: Woodland / City: density gradient 0.2~0.4 (generated with a probability of 20%-40% per pixel); Desert / Snowfield: 0.1~0.3; 3) Intensity distribution---Based on SLIC superpixel region adaptation: Main color region: intensity coefficient α=0.3×ΔE (ΔE is the difference value with the background color); Transition region: α=0.15×ΔE; 4) Color rules---Following the diffusion of the main color. Constraints: Noise hue is assigned according to the background primary color weight in CIELAB space; brightness fluctuation range L*±5, chroma fluctuation ab±3; 5) Shape generation---apply fractal dimension control: use Mandelbrot set generator (dimension D=2.3~2.6); adjust edge blurring through Perlin noise; 6) Adversarial enhancement---based on GCAM heatmap feedback: increase noise density by 50% in YOLOv5l sensitive areas (such as bounding box prediction area); implement gradient inverse perturbation on the conv5-3 layer feature map.

[0017] Dynamic optimization mechanism: 1) Ambient light adaptation: Adjusts noise level in real time via a light sensor. (T=30s is the period, Lenv is the ambient light intensity); 2) Adversarial example iteration, updating noise parameters after each batch of training: (ΔAP50 represents the decrease in detection accuracy); 3) Multispectral compatibility, synchronous generation of thermal noise in the infrared band: (Tbg is the background temperature, N is the Gaussian distribution).

[0018] Furthermore, in step 3, the biomimetic camouflage patterns include five types: woodland, desert, mountain, snowfield, and city.

[0019] In step 4, the improved Grad-CAM algorithm is combined in adversarial training, with the core purpose of improving the effectiveness and targeting of adversarial sample generation.

[0020] I. Limitations and Improvement Directions of Gradient CAM: Gradient CAM generates heatmaps by weighted averaging of feature map gradients based on target category scores, thus locating key regions. However, it suffers from the following problems in target detection: 1) Gradient saturation and noise: Gradients in deep feature maps are prone to saturation, leading to heatmaps focusing on incorrect regions; 2) Ambiguous multi-target localization: In YOLO's multi-scale prediction head, CAM struggles to accurately correlate target instances at different scales; 3) Insufficient sensitivity to small targets: Shallow features (containing detailed information) are not effectively utilized.

[0021] II. Improvements to Gradient CAM in YOLOv5l 1) Multi-scale feature fusion mechanism---Improved method: The feature maps of the three detectors (P3, P4, P5) of YOLOv5l are fused and a comprehensive heatmap is generated through weighted aggregation: shallow features (P3) enhance the localization of small targets, and deep features (P5) improve semantic perception. Formula optimization: in αi Adaptively adjusts based on target size (smaller targets are assigned higher height). α 3); 2) Gradient-weighted strategy optimization – Improved method: Introducing gradient magnitude normalization and significance filtering to suppress gradient magnitudes below a threshold. τ The noise points are removed, and only positive gradients (regions that contribute positively to the category score) are retained, so that the heatmap focuses more on discriminative regions (such as object edges rather than background). 3) Channel Attention Guidance---Improved Method: Embed the SE (Squeeze-and-Excitation) module in the channel weighting of Grad-CAM: dynamically learn the importance weights of each channel, suppress interference from irrelevant channels, and realize the input of the output of YOLOv5's C3 layer into the SE module, and then calculate the gradient weighting; 4) Adversarial Example Generation Strategy – Improved Method: Based on Heatmap H final Generate regions to counter disturbances.

[0022] To verify the effect of adversarial examples reducing the accuracy of the target recognition model by ≥40%, the test conditions must meet the following requirements: 1) Dataset requirements---Annotation quality: Use a dataset of individual soldier images containing a large number of labeled target locations to ensure annotation accuracy (such as COCO or a custom military dataset); Split ratio: The dataset is divided into a training set (70%), a validation set (15%), and a test set (15%); 2) Model configuration---Base model: Use YOLOv5l as the target detection model; Adversarial training: Combine with an improved gradient CAM algorithm (GCAM) to focus on perturbations in key feature regions; 3) Test environment---Hardware conditions: Use the same GPU configuration (such as NVIDIA RTX 3090) and software framework (such as PyTorch) as the training environment. 1.12); Test scenarios: Test scenarios including five environments such as woodland, desert, mountain, snowfield and city are generated through 3D modeling technology to simulate real combat conditions; 4) Evaluation indicators---accuracy threshold: Adversarial examples are required to reduce the recognition accuracy of the target recognition model by ≥40%; Benchmark comparison: The original image without adversarial perturbation is used as the baseline to compare the change in accuracy before and after the attack; 5) Other constraints---dynamic update: The test set must include adversarial examples that the model has not seen before to ensure that the evaluation results reflect the actual robustness; Multi-model validation: Cross-validation with other mainstream detection models (such as YOLOv8) can be selected to ensure the universality of the attack.

[0023] The adversarial example generation format and target model architecture are as follows: I. Analysis of Adversarial Example Generation Forms --- 1) The pattern itself is an adversarial example: The generated biomimetic camouflage pattern directly reduces the accuracy of the target recognition model (a decrease of ≥40%) through adversarial training, indicating that the pattern itself is adversarial; Patterns generated through texture deformation, superposition synthesis, and digital camouflage design (such as woodland camouflage) are the final form of adversarial examples and can be directly used to interfere with AI recognition systems; 2) Potential applications of adversarial patches: By generating and outputting "patches" through an improved gradient CAM algorithm, it is speculated that the patch is a local adversarial perturbation area used to enhance the specific interference effect of the pattern; The patch may exist in the form of superposition layers, dynamically adjusting the adversarial strength of key feature areas.

[0024] II. Specific Architecture of the Target Model in Adversarial Training --- 1) Unified Target Model: YOLOv5l. Adversarial training explicitly adopts the YOLOv5l model (a single-stage target detector) due to its widespread application in UAV and camera surveillance scenarios. The model structure includes the SPPF module, the C3 Neck fusion module, and multi-scale prediction heads (80×80°c, 40×40°c, 20×20°c), which is suitable for real-time detection in complex environments. 2) Improved Gradient CAM Algorithm (GCAM): By focusing on the sensitive areas of the model (such as the conv5-3 layer feature map), adversarial perturbations are generated to improve attack efficiency. Multi-scale feature fusion (P3 / P4 / P5 detection head) and channel attention mechanism (SE module) are introduced to optimize the heat map localization accuracy.

[0025] III. Supplementary Explanation---1) Environmental Adaptability of Adversarial Examples: The generated adversarial examples support five environmental modes: woodland, desert, mountain, snow, and city. Cross-scene interference is achieved by dynamically adjusting texture deformation and noise distribution; 2) Test Conditions and Evaluation Indicators: Multi-environment test scenarios are generated using 3D modeling technology, and the YOLO series models are used to verify the effect of a ≥40% decrease in recognition accuracy.

[0026] Multimodal dynamic optical jamming system (MDOS) and biomimetic camouflage patterns are solutions from different technical approaches. The former adopts a multi-layered defense mechanism that combines dynamic physical jamming with encrypted transmission, while the latter is based on static biomimetic texture design and adversarial training to generate adversarial examples.

[0027] Complementarity: The MDOS solution achieves real-time defense through dynamic optical interference, micro-motion camouflage, and blockchain encryption, while biomimetic camouflage patterns achieve static interference through environmentally adaptable camouflage and adversarial training. The two can be used in combination to deal with different scenarios.

[0028] The specific differences are compared below: Comparison Dimensions Multimodal Dynamic Optical Interference System (MDOS) Bionic camouflage pattern Core Principles Dynamic physical-optical interference (liquid crystal film, micromechanical motion) and biometric encryption Static texture fusion and adversarial training to generate adversarial examples Technical features Non-periodic diffraction gratings, micro servo motor drives, and zero-knowledge proofs in blockchain. Zebra / poplar leaf texture extraction, SLIC superpixel segmentation, YOLOv5l adversarial training Target jamming methods Dynamic light field perturbation, acoustic feature injection, and thermal infrared emissivity modulation Visual texture camouflage, color filling, digital camouflage design Environmental adaptability Adaptive adjustment of light spot distribution and phase cancellation of sound waves Supports five environment modes: woodland, desert, mountain, snow, and urban. Combat robustness Dynamic hardware updates (such as servo motor motion modes) Continuously update adversarial examples to adapt to model upgrades. Relationship description The technical objectives are consistent: both aim to reduce the accuracy of AI recognition systems (such as drones / cameras), and belong to passive countermeasure technologies.

[0029] Some technologies are integrated: The final biomimetic camouflage pattern may incorporate the dynamic adaptation concept from MDOS (such as environmentally adaptive camouflage generation), but its hardware components (such as liquid crystal film and servo motor) are not directly used.

[0030] Application scenario differences: Bionic camouflage patterns are suitable for fixed or slow-moving targets (such as weapons and equipment, military facilities), while MDOS is suitable for personal dynamic scenarios (such as personnel camouflage, real-time monitoring and confrontation).

[0031] Secondly, the present invention also provides a biomimetic camouflage-based anti-drone and anti-facial recognition pattern generation system, comprising: Image preprocessing module: configured to perform color space conversion and SLIC superpixel segmentation algorithm; Texture engine module: configured to extract background textures and biomimetic camouflage textures, and perform texture deformation and overlay compositing; Camouflage generation module: configured to perform color filling, digital camouflage design, and noise enhancement; Adversarial training module: Integrates YOLOv5l model with improved gradient CAM algorithm to generate adversarial examples; Environment adaptation module: Stores template libraries for five types of environments: woodland, desert, mountain, snowfield, and city.

[0032] Furthermore, the adversarial training module includes a 3D scene simulation unit, which is used to construct test scenes using Blender and connect to YOLO series detection models to verify the camouflage effect.

[0033] Compared with the prior art, the present invention has the following beneficial effects: This invention achieves significant improvements in target recognition accuracy by deeply fusing biomimetic camouflage textures such as zebra and poplar leaves with background textures and employing the YOLOv5l model and an improved gradient CAM algorithm for adversarial training. This results in a reduction of over 40% in the accuracy of the target recognition model, effectively overcoming the limited interference effect of traditional techniques. Based on Lab color space conversion and SLIC superpixel segmentation technologies, combined with five types of environmental template libraries including woodland and desert, the invention enables adaptive generation of camouflage patterns for different scenarios, overcoming the poor environmental adaptability of existing solutions. Furthermore, the system's dynamic update capability allows for continuous optimization of adversarial examples as the UAV algorithm iterates, solving the problem of insufficient adversarial robustness. Moreover, the generated patterns, while maintaining visual quality, can be widely applied to equipment, facilities, and personnel camouflage, balancing practicality and user experience, and breaking through the limitations of traditional technologies in application scenarios and real-time processing performance. Attached Figure Description

[0034] The accompanying drawings are incorporated in and form part of this specification, and together with the description serve to explain the principles of the invention.

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of the anti-drone and anti-face recognition pattern generation method based on biomimetic camouflage of the present invention; Figure 2 This is a schematic diagram of color space conversion and noise reduction in Embodiment 1 of the present invention; wherein: (a) is the original image of the forest background; (b) is the image after conversion to Lab color space and noise reduction; Figure 3 This is a schematic diagram of color extraction in Embodiment 1 of the present invention; wherein: (a) is a woodland background with reduced noise image; (b) is a segmentation size of 50; (c) is a segmentation size of 100; (d) is the extracted background main color image; Figure 4 This is a schematic diagram of forest background texture extraction in Embodiment 1 of the present invention; wherein: (a) is the original forest background image; (b) is the forest background texture image; Figure 5 This is a schematic diagram of the biomimetic object in Embodiment 1 of the present invention; wherein: (a) is a zebra; (b) is a poplar leaf; Figure 6 This is a schematic diagram of biomimetic element extraction in Embodiment 1 of the present invention; Figure 7 This is a schematic diagram of texture deformation design in Embodiment 1 of the present invention; wherein: (a) is zebra texture; (b) is optical illusion effect; (c) is poplar leaf; (d) is optical illusion effect 1; (e) is optical illusion effect 2; Figure 8 The diagram shows the texture overlay effect in Embodiment 1 of the present invention; (a) texture overlay composite effect 1; (b) texture overlay composite effect 2; (c) texture overlay composite effect 3; (d) texture overlay composite effect 4; Figure 9 This is a schematic diagram of the combination of biomimetic camouflage texture pattern and background texture pattern in Embodiment 1 of the present invention; wherein: (a) is sketch scheme 1; (b) is sketch scheme 2; Figure 10 This is a schematic diagram of the camouflage scheme in Embodiment 1 of the present invention; wherein: (a) is camouflage scheme 1; (b) is camouflage scheme 2; Figure 11 This is a schematic diagram of a digital camouflage scheme according to Embodiment 1 of the present invention; wherein: (a) is digital camouflage scheme 1; (b) is digital camouflage scheme 2; Figure 12 This is a schematic diagram of a noise enhancement scheme in Embodiment 1 of the present invention; wherein: (a) is noise enhancement scheme 1; (b) noise enhancement scheme 2; Figure 13 This is a schematic diagram of the learning model in Embodiment 1 of the present invention, showing the structure of the YOLOv5l model; Figure 14 This is a schematic diagram illustrating the generation of adversarial examples based on GAN in this invention, demonstrating the workflow of the improved gradient CAM algorithm; Figure 15This is a schematic diagram of the pattern output of the present invention, showing the generated biomimetic camouflage pattern. Detailed Implementation

[0037] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples consistent with some aspects of the invention as detailed in the appended claims.

[0038] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0039] Example 1 (Generation of biomimetic camouflage pattern in woodland) Please see Figures 1-15 Taking a woodland scenario as an example, this embodiment uses a biomimetic camouflage-based anti-drone and anti-facial recognition pattern generation method, such as... Figure 1 As shown, the specific steps include: Step 1, Background Processing: High-resolution images (4000×3000 pixels) with typical forest land features were selected from publicly available geographic image databases, such as... Figure 2 As shown in (a), the image contains elements such as tree trunks, leaves, and ground vegetation, representing a common woodland environment. The woodland background image is converted from the RGB color space to the Lab color space. The Lab color space has advantages in color processing and feature extraction, better separating color and brightness information. Image filtering algorithms (such as Gaussian filtering) are then used for preliminary noise reduction, lowering the interference of random noise in the image, resulting in... Figure 2 (b); The SLIC superpixel segmentation algorithm is used to process the... Figure 2 (b) Segmentation is performed, with segmentation sizes set to 50 and 100 respectively. When the segmentation size is 50, the image is divided into relatively small superpixel blocks, which can capture the local features of the image more precisely; when the segmentation size is 100, the superpixel blocks are larger, which helps to extract the overall structural features of the image. By statistically analyzing the color distribution within each superpixel block, the main background color is extracted, such as... Figure 3 As shown in (d), the main color characteristics of the woodland environment are obtained and used for subsequent pattern design.

[0040] Step 2, Texture Blending: The Canny edge detection algorithm is used to extract the texture contours of the woodland background image, such as... Figure 4 As shown, the background texture information, such as the texture of the tree trunk and the shape of the leaves, is highlighted, providing a foundation for subsequent integration with biomimetic textures.

[0041] Selection of biomimetic objects: Zebras and poplar leaves were selected as biomimetic objects, such as... Figure 5 As shown; the striped texture of zebras has a good visual confusion effect in natural environments, and the vein texture of poplar leaves can simulate the natural form of plants. Combining the two can enhance the biomimetic characteristics of the camouflage pattern; Biomimetic element extraction: Local stripe contours of zebra texture are extracted, and contour tracking algorithms are used to accurately extract the edges of each zebra stripe; the overall contour of poplar leaf veins is extracted, and the structure of the veins is separated through threshold segmentation and morphological processing, such as... Figure 6 As shown; Texture deformation design: Based on the principle of optical illusion, the textures of zebra stripes and poplar leaves are deformed. For example, the zebra stripes are bent and twisted to better blend them with the background; the veins of the poplar leaves are stretched and deformed to simulate changes in shape under different lighting and viewing angles, such as... Figure 7 As shown; Texture overlay synthesis: The deformed zebra texture and poplar leaf texture are overlaid and synthesized. By adjusting the transparency, position, and size of the textures, a biomimetic camouflage texture patch image with biological characteristics is formed, such as... Figure 8 As shown, the biomimetic texture is initially integrated with the background texture.

[0042] By combining biomimetic camouflage texture design patterns with woodland background outline textures, two sketch schemes were designed, such as... Figure 9 As shown.

[0043] Step 3: Camouflage Generation: Based on the background primary color extracted in Step 1, fill and design the sketch scheme to generate a camouflage scheme, such as... Figure 10 As shown, during the filling process, the saturation and brightness of the colors are adjusted to make the pattern more harmonious with the background color; the mosaic size is determined by calculating the camouflage unit size. Considering the recognition resolution of drones and cameras, the camouflage unit size is set to 10×10 pixels, generating a digital camouflage scheme, as shown below. Figure 11 As shown; to enhance the camouflage effect, a fine granular texture was designed on the digital camouflage pattern to increase noise effect, such as... Figure 12 As shown, a random noise generation algorithm is used to add Gaussian noise to the pattern to simulate subtle texture changes in the natural environment, thereby improving the realism and camouflage performance of the pattern. Step 4, Adversarial Enhancement: Collect 5000 individual soldier images with labeled target locations, covering different poses, clothing, and woodland environmental conditions. Divide the dataset into a 7:2:1 ratio for training (3500 images), validation (1000 images), and test (500 images) sets for subsequent model training and performance evaluation; generate adversarial examples using the attack identification and detection model: employing methods such as... Figure 13The YOLOv5l model shown converts the training and validation sets to a format suitable for the model (such as Darknet format), and then trains it for 100 epochs. Each epoch uses a batch size of 16, an initial learning rate of 0.001, and a cosine annealing learning rate adjustment strategy, gradually decreasing the learning rate as training progresses to improve convergence. During training, the model's accuracy and recall on the validation set are recorded, and the model with the best training performance is selected. The trained model and test set images are then loaded, and target recognition is tested. An improved Gradient CAM (GCAM) algorithm is used to select and modify target images to generate adversarial examples. Specifically, by calculating the model's gradient response to different regions in the image, regions with significant impact on recognition results are identified. Designed perturbations are then added to these regions to generate adversarial examples that can mislead the recognition model. The completed patched biomimetic camouflage pattern is then output, such as... Figure 15 As shown.

[0044] To verify the effectiveness of the biomimetic camouflage pattern generated in this embodiment, mainstream detection models YOLOv5s, YOLOv6, and YOLOv7 were used for testing. In a simulated forest scene, using uncamouflaged target images, these models achieved an average recognition accuracy of 92%; after using the biomimetic camouflage pattern generated by this invention, the average recognition accuracy dropped to 48%, achieving the expected interference effect.

[0045] Example 2 (Generation of Desert Biomimetic Camouflage Pattern) The specific process of generating the desert biomimetic camouflage pattern in this embodiment is as follows: Step 1: Background Processing: Select a desert background image with a resolution of 3840×2160, containing sand dunes, rocks, and a small number of drought-resistant plants; convert the image to the Lab color space, perform median filtering to reduce noise, and then use the SLIC superpixel segmentation algorithm to segment the image with segmentation sizes of 40 and 80 respectively to extract the main background colors, which are mainly different shades of yellowish-brown and gray.

[0046] Step 2, Texture Blending: The background texture extraction uses the LBP (Local Binary Pattern) algorithm to highlight the ripples of sand dunes and the surface texture of rocks. The biomimetic objects selected are the fur texture of a desert fox and the leaf texture of a camel thorn. Local hair direction extraction is performed on the desert fox fur texture, and the overall outline and spike features of the camel thorn leaf texture are extracted. Based on the principle of visual illusion, the fur texture is distorted in direction, and the leaf texture is scaled and deformed. These are then superimposed to create a biomimetic camouflage texture, which is combined with the background texture to design two sketch schemes.

[0047] Step 3, Camouflage Generation: Fill the sketch with the main background color, calculate the camouflage unit size as 8×8 pixels to generate digital camouflage, and add salt and pepper noise to enhance the camouflage effect.

[0048] Step 4, Adversarial Enhancement: Construct a dataset of 4000 individual soldier images in a desert scene with labeled target locations, and divide it into training, validation, and test sets in a 7:2:1 ratio. Train the YOLOv5l model for 80 epochs, and generate adversarial examples using an improved gradient CAM algorithm. Testing showed that in a desert scene, the patterns generated by this invention reduced the recognition accuracy of mainstream detection models from 89% to 45%.

[0049] Furthermore, this invention also provides a biomimetic camouflage-based anti-drone and anti-face recognition pattern generation system. This system includes an image preprocessing module: using Python's OpenCV library to implement color space conversion and SLIC superpixel segmentation algorithm, supporting input and processing of various image formats; a texture engine module: based on the deep learning framework PyTorch, utilizing convolutional neural networks for texture extraction and deformation operations, with GPU acceleration to improve processing efficiency; a camouflage generation module: using MATLAB to write color filling, digital camouflage design, and noise enhancement algorithms, providing an adjustable user interface to facilitate users adjusting camouflage effects according to different needs; an adversarial training module: integrating the YOLOv5l model and an improved gradient CAM algorithm, built using the TensorFlow framework, including a 3D scene simulation unit, constructing realistic test scenarios using Blender, and connecting to YOLO series detection models to verify camouflage effects; and an environment adaptation module: storing template libraries for five types of environments: woodland, desert, mountain, snowfield, and city. Each template library contains typical background images, color features, and texture parameters, supporting users to quickly switch and customize camouflage patterns for different environments.

[0050] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention.

[0051] It should be understood that the present invention is not limited to the content already described above, and various modifications and changes can be made without departing from its scope. The scope of the present invention is limited only by the appended claims.

Claims

1. A method for generating anti-drone and anti-facial recognition patterns based on biomimetic camouflage, characterized in that, Includes the following steps: Step 1: Background processing: Select the background image of the target environment, perform color space conversion and noise reduction, and then use the SLIC superpixel segmentation algorithm to segment the processed image and extract the main background color; Step 2, Texture Blending: Extract the background texture outline, simultaneously extract the local and overall features of the biomimetic camouflage texture, and then combine the texture with the background texture after texture deformation based on the principle of visual illusion. Step 3, Camouflage Generation: Fill the background with the main color of the fused pattern, generate a digital camouflage base pattern by calculating the size of the camouflage unit, and add grain noise to form the final biomimetic camouflage pattern; Step 4, Adversarial Enhancement: Construct a dataset with labeled images, use the YOLOv5l model combined with the improved gradient CAM algorithm for adversarial training, and generate adversarial samples that reduce the accuracy of the target recognition model by ≥40%.

2. The method for generating anti-drone and anti-face recognition patterns based on biomimetic camouflage according to claim 1, characterized in that, In step 1, the color space conversion adopts the Lab color space, and when the SLIC superpixel segmentation algorithm is used to segment the processed image, the size of the segmented image is 50 to 100.

3. The method for generating anti-drone and anti-face recognition patterns based on biomimetic camouflage according to claim 1, characterized in that, In step 2, the biomimetic camouflage texture includes animal textures and plant textures; The animal texture is selected from zebra texture, and the plant texture is selected from poplar leaf texture.

4. The method for generating anti-drone and anti-face recognition patterns based on biomimetic camouflage according to claim 3, characterized in that, The extraction of the biomimetic camouflage texture includes: extracting the local stripe contours of zebra patterns and extracting the overall vein contours of poplar leaf patterns.

5. The method for generating anti-drone and anti-face recognition patterns based on biomimetic camouflage according to claim 1, characterized in that, In step 3, the biomimetic camouflage patterns include five types: woodland, desert, mountain, snowfield, and city.

6. A biomimetic camouflage-based anti-drone and anti-facial recognition pattern generation system, characterized in that, include: Image preprocessing module: configured to perform color space conversion and SLIC superpixel segmentation algorithm; Texture engine module: configured to extract background textures and biomimetic camouflage textures, and perform texture deformation and overlay compositing; Camouflage generation module: configured to perform color filling, digital camouflage design, and noise enhancement; Adversarial training module: Integrates YOLOv5l model with improved gradient CAM algorithm to generate adversarial examples; Environment adaptation module: Stores template libraries for five types of environments: woodland, desert, mountain, snowfield, and city.

7. The anti-drone and anti-facial recognition pattern generation system based on biomimetic camouflage according to claim 6, characterized in that, The adversarial training module includes a 3D scene simulation unit, which is used to construct test scenes using Blender and connect to YOLO series detection models to verify the camouflage effect.

Images