A drone identification method and system for low-altitude security

Abstract

The present application provides a method and system for drone identification for low-altitude security. Specifically, a set of optical compensation parameters optimized for drone imaging is generated in real time through a joint optimization equation of the ambient light intensity change rate and the background motion vector field, and the parameter set is used to process the original optical sequence within the target capture window. A time-domain deconvolution kernel is constructed in combination with the drone motion characteristics to generate an enhanced optical image that is resistant to motion blur. A nonlinear weighted fusion is performed using a disturbance intensity evaluation function to form an anti-disturbance feature mask. The enhanced optical image and the anti-disturbance feature mask are spatially superimposed, and a multi-scale residual network is used to estimate the target confidence of the superimposed image to generate a drone identification result. The technical solution provided by the present application significantly improves the accuracy and anti-interference capability of drone identification in low-altitude complex environments.

Description

Technical Field

[0001] The present application relates to the field of drone identification technology, and in particular to a drone identification method and system for low-altitude security. Background Art

[0002] In low-altitude security scenarios, drone recognition faces complex environmental interference and adversarial attacks, such as dynamic lighting changes, background motion interference, and malicious adversarial sample attacks.

[0003] Existing technologies typically use deep learning-based target detection networks combined with optical flow estimation methods to track moving targets. This approach uses optical flow fields to compensate for background motion, employs deep learning models to extract target features and perform recognition, and introduces adversarial training strategies to enhance the model's defenses against adversarial examples. Existing solutions fail to effectively incorporate the rate of change in ambient light intensity for adaptive compensation when dealing with dynamic lighting changes, resulting in a significant decrease in recognition accuracy in scenes with drastic lighting changes. Furthermore, optical flow estimation methods are prone to errors when dealing with complex background motion, and adversarial training strategies have limited defenses against new adversarial examples, making them difficult to handle with diverse adversarial attacks. Summary of the Invention

[0004] The embodiments of the present application provide a method and system for identifying drones for low-altitude security, which are used to solve the problem of poor accuracy in the existing technology.

[0005] In a first aspect, an embodiment of the present application provides a method for identifying drones for low-altitude security, including:

[0006] Based on the joint optimization equation of the ambient light intensity change rate and the background motion vector field in low-altitude security scenes, a set of optical compensation parameters optimized for drone imaging is generated in real time;

[0007] The original optical sequence is acquired within the target capture window by using the optical compensation parameter set, and a time domain deconvolution kernel is constructed in combination with the motion characteristics of the UAV to generate an enhanced optical image with anti-motion blur characteristics;

[0008] Extracting adversarial sample defense features from the enhanced optical image, and performing nonlinear weighted fusion on the adversarial sample defense features by constructing a perturbation strength evaluation function to form an adversarial perturbation feature mask;

[0009] Performing a spatial superposition operation on the enhanced optical image and the adversarial perturbation feature mask, and using a multi-scale residual network based on an attention mechanism to estimate the UAV target confidence on the superposed image;

[0010] Based on the target confidence estimation results, a drone recognition result with robustness against environmental interference and adversarial attacks is generated.

[0011] Optionally, based on the joint optimization equation of the ambient light intensity change rate and the background motion vector field in the low-altitude security scene, a set of optical compensation parameters optimized for UAV imaging is generated in real time, including:

[0012] A multi-channel light intensity sensor is used to collect ambient light intensity data in low-altitude security scenarios in real time. The frequency domain characteristics of the light intensity change rate are extracted based on time series analysis, and a light intensity change rate function model is constructed.

[0013] Performing multi-scale pyramid decomposition on the light intensity change rate function model, calculating the initial motion vector of the sparse feature points on each pyramid layer to form a background motion vector field;

[0014] The light intensity change rate function model is coupled with the background motion vector field to establish a joint optimization equation with light intensity smoothing constraint and motion vector field energy minimization as objective functions, and the alternating direction multiplier method is used to iteratively solve the joint optimization equation until convergence to a stable state;

[0015] According to the output parameters of the joint optimization equation, the exposure time, gain coefficient and filtering threshold of the optical compensation module are dynamically adjusted to generate a set of optical compensation parameters optimized for UAV imaging.

[0016] Optionally, the original optical sequence is acquired within the target capture window by using the optical compensation parameter set, and a time domain deconvolution kernel is constructed in combination with the motion characteristics of the drone to generate an enhanced optical image with anti-motion blur characteristics, including:

[0017] Based on the optical compensation parameter set, N frames of optical signals continuously collected within the target capture window are subjected to channel-by-channel enhancement processing, wherein each frame of optical signal is separated into three channels, RGB, and then a dot product operation is performed with the gain coefficient matrix of the corresponding channel to generate an original optical sequence;

[0018] Performing time-domain feature modeling on the original optical sequence, extracting the brightness gradient feature and chromaticity correlation feature of each pixel in the time dimension through a multi-scale dilated convolution kernel, and outputting a feature tensor with spatiotemporal correlation;

[0019] A temporal deconvolution kernel is constructed based on the feature tensor and the UAV motion characteristics. A motion consistency constraint equation is constructed by minimizing the optical flow residual between adjacent frames. A smoothness constraint on the UAV motion trajectory is introduced, and a constrained optimization algorithm is used to solve the pixel-level temporal deconvolution kernel matrix.

[0020] The time-domain deconvolution kernel is subjected to a joint space-time deconvolution operation with the original optical sequence, wherein a back-projection calculation is performed along the time axis for each spatial position, and the phase response characteristics of the deconvolution kernel are adjusted through an iterative update strategy to generate an enhanced optical image with anti-motion blur characteristics.

[0021] Optionally, the light intensity change rate function model is coupled with the background motion vector field to establish a joint optimization equation with light intensity smoothness constraint and motion vector field energy minimization as objective functions, and the alternating direction multiplier method is used to iteratively solve the joint optimization equation until convergence to a stable state, including:

[0022] Design the coupling term parameters, fuse the intensity smoothness constraint of the intensity change rate function model with the background motion vector field through weighting factors, and form a joint optimization equation;

[0023] Decomposing the joint optimization equation into a first subproblem of constraining the light intensity gradient between adjacent pixels by introducing a light intensity smoothing regularization term of the light intensity change rate function model and a second subproblem of constraining the spatial continuity of the motion vector by introducing an energy term of the background motion vector field;

[0024] Constructing augmented Lagrangian functions for the first subproblem and the second subproblem respectively, and introducing Lagrangian multiplier terms and quadratic penalty terms of linear constraints into the objective function;

[0025] After each variable update, the augmented Lagrange multiplier is updated synchronously, and the weight coefficient of the quadratic penalty term is dynamically adjusted through an adaptive step size adjustment strategy;

[0026] Based on the quadratic penalty term, the alternating direction multiplier method is used to iteratively solve the joint optimization equation, and the relative error norm in two adjacent iterations is calculated. When the relative error norm is less than a preset threshold, it is determined that the solution has converged to a stable state.

[0027] Optionally, a temporal deconvolution kernel is constructed based on the feature tensor and the UAV motion characteristics. A motion consistency constraint equation is constructed by minimizing the optical flow residual between adjacent frames, and a smoothness constraint of the UAV motion trajectory is introduced. A constrained optimization algorithm is used to solve the pixel-level temporal deconvolution kernel matrix, including:

[0028] Based on the multi-scale spatiotemporal correlation features of the feature tensor, the spatiotemporal local contrast features are extracted to generate an initial parameter set of the dynamic convolution kernel. Combined with the high-speed translation, hovering jitter and rotor periodic motion patterns in the UAV motion characteristics, the initial parameter set is subjected to a priori motion trajectory correction, and the initial parameter set is mapped to the weight space of the deconvolution kernel to form a time domain deconvolution kernel.

[0029] The temporal deconvolution kernel is combined with the function of minimizing the residual optical flow between adjacent frames for spatiotemporal joint analysis. The smoothness constraint of the UAV motion trajectory is introduced to construct a motion consistency constraint equation with the pixel displacement vector as a variable. The smoothness constraint of the UAV motion trajectory includes the upper limit of the UAV acceleration and the heading continuity condition.

[0030] The constrained optimization algorithm is used to iteratively perform spatial and temporal domain optimization. In each iteration, the weight distribution of the optical flow residual function is adjusted according to the smoothness constraint of the UAV's motion trajectory until the mean square error of the optical flow residual converges to a preset accuracy range, and the pixel-level temporal deconvolution kernel matrix that satisfies the motion consistency constraint is output.

[0031] Optionally, time domain feature modeling is performed on the original optical sequence, and the brightness gradient feature and chromaticity correlation feature of each pixel in the time dimension are extracted by a multi-scale dilated convolution kernel, and a feature tensor with spatiotemporal correlation is output, including:

[0032] The temporal features of the original optical sequence are decomposed into a luminance gradient component, which is obtained by performing inter-frame difference operations to obtain a pixel-by-pixel luminance change matrix, and a chrominance correlation component, which uses the covariance matrix of the chrominance channel to describe the cross-frame chrominance correlation.

[0033] Performing depth-separable convolution operations on the luminance gradient component and the chrominance correlation component through each branch of the multi-scale dilated convolution kernel;

[0034] Based on the depthwise separable convolution operation, cross-scale correlation modeling is performed on the multi-branch output features, the convolution results of different expansion rates are aligned according to the time dimension, and a feature tensor with spatiotemporal correlation is output.

[0035] Optionally, extracting adversarial sample defense features from the enhanced optical image, and performing nonlinear weighted fusion on the adversarial sample defense features by constructing a perturbation strength evaluation function to form an adversarial perturbation feature mask, including:

[0036] Perform multi-scale spatial filtering on the enhanced optical image, combine the adaptive noise suppression algorithm with the local contrast enhancement operation to generate a denoised and standardized optical image;

[0037] Feature extraction is performed based on the denoised standardized optical image. The high-frequency texture features and low-frequency structural features of the drone's multi-rotor structure and metal reflective properties are screened through a cross-channel attention mechanism to construct an adversarial sample defense feature matrix. The attention mechanism focuses on key parts of the drone's rotors and fuselage.

[0038] The perturbation intensity evaluation function is used to dynamically calculate the weight coefficient of each adversarial sample defense feature channel, and the layer-by-layer reverse gradient accumulation strategy is used to adjust the weight distribution to achieve nonlinear weighted fusion of the feature space;

[0039] The adversarial sample defense feature matrix after nonlinear weighted fusion is iteratively optimized to ensure semantic consistency between the mask boundary and the original image, and the adversarial perturbation feature mask is output.

[0040] Optionally, performing a spatial superposition operation on the enhanced optical image and the anti-disturbance feature mask, and using a multi-scale residual network based on an attention mechanism to estimate the UAV target confidence on the superposed image, including:

[0041] Normalizing the anti-disturbance feature mask so that the value range of the anti-disturbance feature mask matches the pixel distribution of the enhanced optical image, and generating a superimposed image through a spatial domain superposition operation;

[0042] Constructing a multi-scale processing module to perform a multi-scale residual network operation based on an attention mechanism on the superimposed image to generate a multi-scale fusion feature tensor;

[0043] The multi-scale fusion feature tensor is input into the confidence estimation unit, local feature statistics are calculated, and a UAV target confidence estimation value is generated.

[0044] Optionally, based on the drone target confidence estimation result, a drone identification result with robustness against environmental interference and counterattacks is generated, including:

[0045] Based on the UAV target confidence estimation results, a sliding window mechanism is used to calculate the environmental interference intensity index in real time. When the confidence fluctuation of consecutive frames exceeds the preset threshold, the adaptive threshold adjustment algorithm is activated;

[0046] The interference pattern recognition of the confidence abnormal area is performed through the adaptive threshold adjustment algorithm to generate the environmental interference feature mask;

[0047] The adversarial sample detection module is activated for low-confidence targets, and the gradient backpropagation method is used to analyze the sensitivity of feature space perturbations and generate robustness evaluation coefficients.

[0048] The drone target confidence estimation result, the environmental interference feature mask and the robustness evaluation coefficient are combined to generate a drone recognition result that is robust against environmental interference and counterattacks.

[0049] In a second aspect, the embodiments of the present application provide a drone identification system for low-altitude security, including:

[0050] A generation module is used to generate a set of optical compensation parameters optimized for UAV imaging in real time based on the joint optimization equation of the ambient light intensity change rate and the background motion vector field in low-altitude security scenarios;

[0051] The generation module is further configured to acquire an original optical sequence within a target capture window using the optical compensation parameter set, construct a time-domain deconvolution kernel based on the motion characteristics of the drone, and generate an enhanced optical image with motion blur resistance.

[0052] a processing module, extracting adversarial sample defense features from the enhanced optical image, and performing nonlinear weighted fusion on the adversarial sample defense features by constructing a perturbation strength evaluation function to form an adversarial perturbation feature mask;

[0053] an estimation module, performing a spatial superposition operation on the enhanced optical image and the anti-disturbance feature mask, and performing target confidence estimation on the superposed image using a multi-scale residual network based on an attention mechanism;

[0054] The generation module is also used to generate a drone identification result that is robust against environmental interference and attacks based on the target confidence estimation result.

[0055] In an embodiment of the present application, based on the joint optimization equation of the ambient light intensity change rate and the background motion vector field in the low-altitude security scene, an optical compensation parameter set optimized for drone imaging is generated in real time; the original optical sequence is obtained within the target capture window through the optical compensation parameter set, and a time domain deconvolution kernel is constructed in combination with the drone motion characteristics to generate an enhanced optical image with anti-motion blur characteristics; adversarial sample defense features are extracted from the enhanced optical image, and the adversarial sample defense features are nonlinearly weighted fused by constructing a perturbation intensity evaluation function to form an adversarial perturbation feature mask; the enhanced optical image and the adversarial perturbation feature mask are subjected to a spatial superposition operation, and a multi-scale residual network based on the attention mechanism is used to perform target confidence estimation on the superimposed image; based on the target confidence estimation result, a drone recognition result with robustness against environmental interference and adversarial attacks is generated.

[0056] The technical solution of this application has the following beneficial effects:

[0057] This application achieves adaptive dynamic adjustment of optical compensation parameters by fusing the light intensity change rate with the background motion vector field, significantly improving imaging stability and target capture quality in complex environments. The time-domain deconvolution kernel is used to eliminate image blur caused by high-speed target motion, restore detail information, and improve the clarity and recognizability of the target area. Accurately detect and identify adversarial attack areas, enhance the robustness of image processing and target recognition, and reduce the false detection rate caused by adversarial attacks. Through multi-scale feature extraction and attention mechanism, the accuracy and stability of target detection are improved, ensuring high-confidence recognition results in complex scenarios. Combining confidence-driven screening and multi-frame fusion strategies, highly reliable drone recognition results are generated, significantly improving the environmental adaptability and anti-interference capabilities of low-altitude security systems.

[0058] Furthermore, the present application uses a multi-channel light intensity sensor to collect ambient light intensity data in low-altitude security scenes in real time, extracts the frequency domain characteristics of the light intensity change rate based on time series analysis, and constructs a light intensity change rate function model; performs multi-scale pyramid decomposition on the light intensity change rate function model, calculates the initial motion vector of the sparse feature points on each layer of the pyramid, and forms a background motion vector field; couples the light intensity change rate function model with the background motion vector field, establishes a joint optimization equation with light intensity smoothing constraint and motion vector field energy minimization as objective functions, and uses the alternating direction multiplier method to iteratively solve the joint optimization equation until it converges to a stable state; according to the output parameters of the joint optimization equation, dynamically adjusts the exposure time, gain coefficient and filtering threshold of the optical compensation module to generate a set of optical compensation parameters optimized for drone imaging.

[0059] These and other aspects of the present application will become more readily apparent from the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0061] Figure 1 A flowchart of a method for identifying drones for low-altitude security provided by this application is shown;

[0062] Figure 2 The figure shows a schematic structural diagram of a drone identification system for low-altitude security provided by the present application. DETAILED DESCRIPTION

[0063] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0064] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this document or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to being different types.

[0065] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0066] Figure 1 A flowchart of a method for identifying drones for low-altitude security is provided for the embodiment of the present application. Figure 1 As shown, the method includes:

[0067] 101. Based on the joint optimization equation of the ambient light intensity change rate and the background motion vector field in low-altitude security scenes, a set of optical compensation parameters optimized for drone imaging is generated in real time;

[0068] In this step, the ambient light intensity change rate refers to the rate at which the light intensity in the low-altitude security scene changes over time, which is usually calculated using time series data collected by a light intensity sensor.

[0069] The background motion vector field is a vector field that describes the motion direction and speed of background objects in low-altitude security scenes. It is generated through optical flow estimation or feature matching methods and is used to characterize dynamic interference in the scene background (such as shaking trees and drifting clouds).

[0070] The optical compensation parameter set includes parameters such as exposure time and gain coefficient, which are used to adjust the optical system to cope with changes in ambient light intensity and background motion interference. It is used to suppress ambient light interference and compensate for imaging blur caused by high-speed movement of the drone.

[0071] A light intensity sensor collects real-time scene light intensity data, and its time-difference sequence is calculated to obtain the ambient light intensity change rate. Optical flow is estimated for the video frame sequence, and the background motion vector field is extracted. Motion consistency clustering is used to remove outliers. A joint optimization equation is constructed, coupling the light intensity change rate with the background motion vector field. The objective function includes a light intensity smoothing constraint and a motion vector field energy minimization term. The joint optimization equation is iteratively solved using the alternating direction multiplier method, outputting a set of optical compensation parameters generated in real time.

[0072] In low-altitude security scenarios, drone surveillance systems use light intensity sensors to collect ambient light intensity data and calculate its rate of change. Simultaneously, optical flow is estimated from video frame sequences to generate a background motion vector field. These two parameters are fed into a joint optimization equation to generate a set of optical compensation parameters optimized for drone imaging in real time. These parameters are used to adjust the camera's exposure time and gain factor to account for lighting variations and background motion interference.

[0073] 102. Acquire an original optical sequence within the target capture window using the optical compensation parameter set, construct a time-domain deconvolution kernel based on the UAV motion characteristics, and generate an enhanced optical image with motion blur resistance.

[0074] The original optical sequence refers to the sequence of continuous frame images acquired within the target capture window.

[0075] The temporal deconvolution kernel is a temporal filter used to eliminate motion blur in image sequences and restore clear images through deconvolution operations.

[0076] The camera parameters are adjusted using a set of optical compensation parameters to capture the original optical sequence within the target capture window. A temporal deconvolution kernel is constructed based on the drone's motion characteristics (including high-speed translation, hovering jitter, and periodic rotor motion patterns). The motion blur trajectory is estimated based on the optical flow field. Smoothness constraints on the drone's motion trajectory (such as an upper limit on acceleration and heading continuity) are introduced, and the kernel function parameters are optimized using regularization constraints. A temporal deconvolution operation is performed on the original optical sequence to generate an enhanced optical image with motion blur resistance.

[0077] The drone flew obliquely at a speed of 15 m / s. The target capture window tracked its motion and captured five blurred frames. Optical flow analysis revealed that the target formed a zigzag trajectory in the time domain (maximum displacement of 8 pixels / frame). Based on the drone's motion characteristics, a deconvolution kernel modeled the point spread function (PSF) based on this zigzag trajectory. Smoothness constraints were introduced to optimize the kernel parameters. After regularized deconvolution, the texture clarity of the target rotor blades was significantly improved (edge gradient increased from 0.3 to 0.7), and the length of the background tree afterimages was reduced by 70%.

[0078] 103. Extracting adversarial sample defense features based on the multi-rotor structure of the drone and the reflective properties of metal from the enhanced optical image, and performing nonlinear weighted fusion on the adversarial sample defense features by constructing a perturbation intensity evaluation function to form an adversarial perturbation feature mask;

[0079] Adversarial sample defense features refer to features extracted from enhanced optical images that can characterize adversarial attacks, such as abnormal gradient distribution or local texture mutations, which are used to detect potential adversarial attack areas.

[0080] The perturbation strength evaluation function is used to quantify the perturbation strength of the adversarial sample defense feature, and generates the adversarial perturbation feature mask through nonlinear weighted fusion to identify high-risk areas.

[0081] Adversarial sample defense features are extracted from enhanced optical images, including periodic texture features based on the drone's multi-rotor structure, highlight area features caused by metal reflective properties, as well as gradient amplitude, local contrast and texture complexity. Abnormal areas of the image are characterized, and a disturbance intensity evaluation function is constructed. The weight coefficient of each feature is calculated based on the drone's physical characteristics (such as rotor symmetry and metal reflective intensity) and feature sensitivity. The potential risk of adversarial attacks is quantified, and an adversarial perturbation feature mask is generated through nonlinear weighted fusion. Potential adversarial attack areas in the image are identified, providing risk warnings for subsequent processing.

[0082] For example, after generating an enhanced optical image, the system extracts the periodic texture features of the drone's multi-rotor structure and the highlights caused by the metal's reflective properties. Combined with the gradient amplitude and local contrast features, the system calculates weight coefficients using a perturbation intensity evaluation function to generate an adversarial perturbation feature mask, identifying areas in the image that may be susceptible to adversarial attacks. For example, if an area in the image shows abnormal periodic rotor texture, a sudden change in metal reflective intensity, and an abnormally high gradient amplitude, the system determines that the area may be vulnerable to adversarial attacks and marks it as a high-risk area in the mask.

[0083] 104. Performing a spatial domain superposition operation on the enhanced optical image and the anti-disturbance feature mask, and using a multi-scale residual network based on an attention mechanism to perform UAV target confidence estimation on the superimposed image;

[0084] Spatial superposition operation refers to the pixel-level fusion of the enhanced optical image and the adversarial perturbation feature mask to generate a composite image containing adversarial attack information, providing more comprehensive input data for subsequent recognition.

[0085] The multi-scale residual network is a deep learning model based on the attention mechanism, which is used to extract multi-scale features of images and perform target confidence estimation to improve the reliability of recognition results.

[0086] The enhanced optical image and the adversarial perturbation feature mask are spatially superimposed to generate a composite image. The target details and adversarial attack information are integrated. A multi-scale residual network is used to extract features from the composite image. The attention mechanism is used to focus on key areas and enhance the saliency of target features. The UAV target confidence estimation result is output through the fully connected layer to evaluate the reliability of the recognition result and provide a basis for subsequent screening.

[0087] For example, the system spatially superimposes the enhanced optical image with the adversarial perturbation feature mask, inputs the multi-scale residual network for feature extraction, outputs the target confidence estimation result, and evaluates the reliability of drone recognition. When an area in the composite image is marked as a high-risk adversarial attack area, the system lowers the confidence score of the area to avoid misidentification.

[0088] 105. Based on the target confidence estimation results, a drone identification result with robustness against environmental interference and counterattacks is generated.

[0089] The target confidence estimate is the reliability score of the recognition result output by the multi-scale residual network, which is used to select high-confidence recognition results. The recognition result includes information about the drone model, location, and threat level.

[0090] Robustness against environmental interference and adversarial attacks refers to the ability of the recognition system to maintain high accuracy in complex environments and adversarial attacks, ensuring the reliability of the recognition results.

[0091] Based on the target confidence estimation results, high-confidence recognition results are screened, low-confidence misidentifications are eliminated, and recognition accuracy is improved. By fusing multi-frame recognition results, the system's robustness to the environment and adversarial attacks is enhanced, the impact of single-frame misidentification is eliminated, and the final drone recognition results are output to ensure their accuracy and reliability in complex environments.

[0092] The system screens high-confidence recognition results based on the target confidence estimation results, enhances robustness through multi-frame fusion, and ultimately outputs accurate drone recognition results. In complex lighting and adversarial attack environments, the system eliminates single-frame misidentification through multi-frame fusion to ensure the reliability of the final recognition results.

[0093] Steps 101 to 105 of the present invention jointly optimize the ambient light intensity change rate and the background motion vector field to generate a set of optical compensation parameters optimized for drone imaging in real time, effectively coping with dynamic lighting and background motion interference; eliminate motion blur through a time-domain deconvolution kernel to generate a clear enhanced optical image; extract adversarial sample defense features and generate an adversarial perturbation feature mask to identify potential adversarial attack areas; use a multi-scale residual network to estimate target confidence and screen high-confidence recognition results; generate drone recognition results that are robust against environmental interference and adversarial attacks, significantly improving recognition accuracy and reliability in low-altitude security scenarios.

[0094] To address the limitations of single parameter compensation methods in dynamic and complex scenes, and to address the issue of image quality degradation caused by drastic dynamic lighting changes and complex background motion interference in low-altitude security scenes, a joint optimization equation is constructed by fusing the ambient light intensity change rate data collected in real time by the light intensity sensor with the background motion vector field data extracted from the video frame sequence. The objective function is designed using multi-source data coupling constraints, and an alternating optimization algorithm is used to solve the dynamic compensation parameters in real time to achieve dynamic adaptive adjustment of the optical system parameters. In some embodiments, based on the joint optimization equation of the ambient light intensity change rate and the background motion vector field in low-altitude security scenes, a set of optical compensation parameters optimized for drone imaging is generated in real time, including:

[0095] 201. Use a multi-channel light intensity sensor to collect ambient light intensity data in low-altitude security scenes in real time, extract the frequency domain characteristics of the light intensity change rate based on time series analysis, and construct a light intensity change rate function model;

[0096] In step 201, the multi-channel light intensity sensor refers to a sensor with an integrated multispectral photosensor unit that can simultaneously collect light intensity data in multiple bands, including visible light and infrared light. The light intensity change rate is the rate of light intensity fluctuation calculated based on time series data, reflecting the dynamic characteristics of ambient light intensity. The light intensity change rate function model is a mathematical model constructed using frequency domain characteristics (such as dominant frequency and amplitude) to quantify the regularity of light changes.

[0097] In this embodiment, a multi-channel light intensity sensor collects ambient light intensity data at a frequency of 100 Hz. This time series data is segmented using a sliding window (window length: 3 seconds, overlap ratio: 50%). A fast Fourier transform (FFT) is performed on the light intensity data within each window to extract the dominant frequency component (0.1 Hz to 10 Hz) and its amplitude, and a light intensity rate of change function model is constructed.

[0098] 202. Perform multi-scale pyramid decomposition on the light intensity change rate function model, calculate the initial motion vector of the sparse feature points on each pyramid layer, and form a background motion vector field;

[0099] In step 202, multi-scale pyramid decomposition involves decomposing the light intensity rate of change function model into components at different spatial scales (e.g., high-frequency details and low-frequency trends). Sparse feature points are key points extracted using a corner detection algorithm and are used to characterize stable structural features in the scene. The background motion vector field (BMVF) is a collection of background object motion vectors calculated using the optical flow method, reflecting the motion trends of non-target objects.

[0100] In this embodiment, the light intensity rate of change function model is decomposed into a three-layer Gaussian pyramid (with a scale factor of 2). Sparse feature points are extracted on each pyramid layer using the Harris corner detection algorithm (with a threshold set to 0.01). Subsequently, the motion vectors of the feature points are calculated using the Lucas-Kanade optical flow method. A global motion model (such as an affine transformation) is fitted using the RANSAC algorithm. Abnormal vectors with residuals exceeding 2 pixels (such as flying birds and falling leaves) are removed to generate a smooth background motion vector field.

[0101] 203. Coupling the light intensity change rate function model with the background motion vector field, establishing a joint optimization equation with light intensity smoothing constraint and motion vector field energy minimization as objective functions, and iteratively solving the joint optimization equation using an alternating direction multiplier method until convergence to a stable state;

[0102] In step 203, the intensity smoothing constraint limits sudden changes in the rate of change of light intensity to avoid overexposure or underexposure of the image due to drastic fluctuations in illumination. Minimizing the motion vector field energy suppresses abnormal jumps in background motion by constraining the spatial continuity of the motion vectors. The joint optimization equation is a multi-objective optimization model that couples intensity and motion constraints and is used to generate stable compensation parameters.

[0103] In this embodiment, a joint optimization equation is constructed, combining a light intensity smoothing constraint (constraining the gradient of the rate of change of light intensity through first-order differences) with the minimization of the motion vector field energy (ensuring motion continuity through total variation regularization). The optimization problem is decomposed into two subproblems, light intensity compensation and motion compensation, using the alternating direction multiplier method (ADMM), which are solved in alternating iterations. In each iteration, the light intensity subproblem updates its parameters using gradient descent, while the motion subproblem optimizes the vector field using the conjugate gradient method until the residual norm is less than a preset threshold (e.g., 0.01).

[0104] 204. Dynamically adjust the exposure time, gain coefficient, and filter threshold of the optical compensation module according to the output parameters of the joint optimization equation to generate an optical compensation parameter set optimized for UAV imaging.

[0105] In step 204, the exposure time is a parameter used to adjust the camera's exposure time, balancing image brightness and motion blur. The gain factor controls the image signal amplification factor, suppressing noise in low-light environments. The filter threshold dynamically adjusts image processing algorithm parameters (such as the median filter window size) to suppress noise and artifacts. The optical compensation parameter set is used to suppress ambient light interference and compensate for image blur caused by high-speed drone motion.

[0106] In this embodiment of the present application, the optical compensation module is dynamically adjusted based on the output parameters of the joint optimization equation. The exposure time is adjusted so that if the rate of change of light intensity exceeds 100 lux / s, the exposure time is proportionally extended (e.g., from 1 / 1000s to 1 / 500s). The gain coefficient is adjusted to dynamically set the gain based on the mean ambient light intensity (e.g., the gain is increased to 12dB when the light intensity is <200 lux). The filter threshold is adjusted to adjust the median filter window size based on the average speed of the motion vector field (e.g., a 5×5 window is used when the speed is >5 pixels / frame). The resulting set of optical compensation parameters can adapt to scene changes in real time.

[0107] Here's a specific example:

[0108] In a low-altitude security scenario at dusk, the light intensity drops from 1000 lux to 200 lux over 5 minutes, while the background clouds move at a speed of 2 m / s. Step 201: A multi-channel light intensity sensor detects a drop in light intensity from 1000 lux to 200 lux, with a rate of change of 160 lux / min. The system performs time series analysis on the light intensity data, extracts the dominant frequency component (0.5 Hz) through Fourier transform, and constructs a light intensity rate function model to quantify the dynamic fluctuations in light intensity. Step 202: A three-layer pyramid decomposition is performed on the light intensity rate function model to extract sparse feature points (Harris corner detection with a threshold of 0.01). The Lucas-Kanade optical flow method is used to calculate the initial motion vector, and the RANSAC algorithm is used to remove outliers such as flying birds or falling leaves. This generates a smooth background motion vector field, representing the trend of the clouds moving at a speed of 2 m / s. In step 203, a joint optimization equation is constructed, combining light intensity smoothing constraints (to suppress sudden illumination noise) and motion vector field energy minimization (to ensure motion continuity). The ADMM algorithm is then used to iteratively solve the problem. Each iteration takes less than 10 ms, and the residual norm is calculated until it falls below a preset threshold (0.01). Compensation parameters are then output: exposure time 1 / 300 s, gain 10 dB. In step 204, the optical system is dynamically adjusted based on the output parameters, adjusting the exposure time from 1 / 1000 s to 1 / 300 s and increasing the gain to 10 dB. A 3×3 median filter is also applied to suppress noise. After compensation, the image brightness standard deviation is reduced from 35 to 12, background blur is reduced by 50%, and the clarity of the target area is significantly improved.

[0109] In summary, this solution generates a set of optical compensation parameters optimized for drone imaging in real time through the joint optimization of a multi-channel light intensity sensor and the background motion vector field. This effectively solves the problem of image quality degradation caused by dynamic lighting changes and complex background motion interference in low-altitude security scenarios, significantly improves the imaging clarity and recognition stability of drone targets, and provides a high-quality input data foundation for subsequent anti-motion blur processing and anti-attack defense.

[0110] To improve the image blur problem caused by high-speed target motion in low-altitude security scenarios, this step dynamically adjusts the imaging parameters (such as exposure time and gain) within the target capture window through a set of optical compensation parameters to obtain the original optical sequence. Based on the target motion trajectory modeling and the UAV motion characteristics, a time-domain deconvolution kernel is constructed. The deconvolution operation is used to eliminate motion blur and restore details. This aims to address the defect that traditional fixed-parameter imaging cannot adaptively compensate for motion blur in dynamic scenes. A clear image is reversely solved through the time-domain degradation model, providing high-definition, anti-blur optical input data for subsequent target detection and recognition.

[0111] In some embodiments, the optical compensation parameter set is used to obtain an original optical sequence within the target capture window, and a temporal deconvolution kernel is constructed in combination with the motion characteristics of the drone to generate an enhanced optical image with anti-motion blur characteristics, including:

[0112] 301. Based on the optical compensation parameter set, perform channel-by-channel enhancement processing on N frames of optical signals continuously collected within the target capture window, wherein each frame of optical signal is separated into three channels, RGB, and then performs a dot multiplication operation with the gain coefficient matrix of the corresponding channel to generate an original optical sequence;

[0113] In step 301, the gain coefficient matrix is a parameter matrix designed for each of the three RGB channels, dynamically adjusting the brightness enhancement ratio of each channel. Channel enhancement processing decomposes each optical signal frame into three independent channels: red, green, and blue, and performs gain adjustment on each channel to optimize color balance. The original optical sequence is a collection of consecutive frame images generated after channel enhancement processing, preserving temporal correlation and spatial detail.

[0114] In this embodiment of the present application, five consecutive frames of optical signals within the target capture window are processed channel by channel based on the gain coefficients in the optical compensation parameter set (e.g., R channel gain 1.2, G channel gain 1.0, B channel gain 0.8). The specific process is to decompose each frame of the image into a grayscale image of three independent RGB channels, perform a dot product operation on the pixel value of each channel with the corresponding gain coefficient matrix (e.g., R channel pixel value × 1.2), and then re-merge the enhanced three-channel image to generate the original optical sequence. For example, in low-light scenes, increasing the R channel gain can compensate for red light attenuation to avoid an overall bluish image.

[0115] 302. Perform time domain feature modeling on the original optical sequence, extract the brightness gradient feature and chromaticity correlation feature of each pixel in the time dimension through a multi-scale dilated convolution kernel, and output a feature tensor with spatiotemporal correlation;

[0116] In step 302, temporal feature modeling analyzes the brightness and chromaticity variations of pixels across consecutive frames, capturing motion trajectories and illumination fluctuations. Multi-scale dilated convolution kernels, with varying time spans (e.g., 3, 5, and 7 frames), are used to extract both short-term mutations and long-term trends. The spatiotemporal correlation feature tensor is a tensor structure that contains pixel brightness gradients, chromaticity correlations, and temporal motion information.

[0117] In an embodiment of the present application, time domain feature modeling is performed on the original optical sequence, and convolution kernels with three expansion rates are used (expansion rate 1 corresponds to a span of 3 frames, and expansion rate 2 corresponds to a span of 5 frames). The brightness gradient (difference between adjacent frames) and chromaticity correlation (RGB channel covariance) of each pixel are calculated by sliding along the time axis, and feature maps of different scales are spliced into a three-dimensional tensor, which outputs a feature tensor with spatiotemporal correlation.

[0118] 303. Constructing a temporal deconvolution kernel based on the feature tensor and the UAV motion characteristics, constructing a motion consistency constraint equation by minimizing the optical flow residual between adjacent frames, introducing a UAV motion trajectory smoothness constraint, and using a constrained optimization algorithm to solve and obtain a pixel-level temporal deconvolution kernel matrix;

[0119] In step 303, the temporal deconvolution kernel is a filter whose parameters are dynamically adjusted based on spatiotemporal characteristics, used to eliminate motion blur. The motion consistency constraint equation ensures the continuity of motion vectors between adjacent frames by minimizing the optical flow residual. The pixel-level temporal deconvolution kernel matrix is a deconvolution kernel designed independently for each pixel, adapting to local motion differences.

[0120] Among them, the UAV motion characteristics include prior knowledge of UAV high-speed translation, hovering jitter and rotor periodic motion, which are used to guide the calculation of optical flow residuals and the optimization of deconvolution kernel parameters.

[0121] In an embodiment of the present application, dense optical flow estimation is performed on adjacent frames, the optical flow field residual (the deviation between the predicted motion and the actual motion) is calculated, and the optical flow residual is corrected for motion trajectory smoothness in combination with the UAV motion characteristics (such as high-speed translation, hovering jitter, and periodic motion of the rotor); minimization of the optical flow residual is used as a constraint condition, and an objective function is designed in combination with the spatiotemporal feature tensor and the UAV motion characteristics, wherein the UAV motion characteristics are used to constrain the weight distribution of the optical flow residual; a constrained optimization algorithm (such as the projected gradient method) is used to iteratively solve the deconvolution kernel matrix, and in each iteration, the deconvolution kernel parameters are adjusted according to the UAV motion trajectory smoothness constraint, and each pixel is optimized independently to ensure that the deconvolution kernel adapts to local motion differences.

[0122] 304. Perform a space-time joint deconvolution operation on the time-domain deconvolution kernel and the original optical sequence, wherein a back-projection calculation is performed along the time axis for each spatial position, and the phase response characteristics of the deconvolution kernel are adjusted through an iterative update strategy to generate an enhanced optical image with anti-motion blur characteristics.

[0123] In step 304, the joint spatial-temporal deconvolution operation simultaneously performs deconvolution in both the spatial and temporal dimensions to restore target details. Backprojection calculations retrace pixel motion trajectories along the temporal axis to correct blurred pixel values. Phase response adjustment dynamically optimizes the phase parameters of the deconvolution kernel to match the target motion frequency.

[0124] In an embodiment of the present application, a joint deconvolution operation in the space and time domains is performed on the time domain deconvolution kernel and the original optical sequence to generate an enhanced optical image that is resistant to motion blur. First, a back projection calculation is performed along the time axis for each spatial position, and the characteristic distribution of motion blur is extracted by analyzing the time domain information in the optical sequence. Secondly, an iterative update strategy is adopted to adjust the phase response characteristics of the deconvolution kernel, and the parameters of the deconvolution kernel are optimized by minimizing the reconstruction error function to ensure that it can effectively compensate for motion blur. In each iteration, the phase response of the deconvolution kernel is updated, and the spatial domain information is combined for joint optimization to gradually improve the image reconstruction accuracy. Finally, an enhanced optical image is generated by a joint deconvolution operation in the space and time domains, which significantly reduces the impact of motion blur, improves the image clarity and detail restoration capability, and provides high-quality data support for subsequent analysis and processing.

[0125] In order to achieve adaptive dynamic adjustment of optical compensation parameters, and to address the imaging parameter mismatch problem caused by the coupling of sudden illumination changes and background motion interference in dynamic scenes, a joint optimization equation is constructed by fusing the light intensity change rate function model (quantifying the illumination fluctuation law) with the background motion vector field (characterizing the motion trend of non-target objects). Light intensity smoothing constraints are introduced to suppress illumination mutation noise. The motion continuity is ensured by combining the energy minimization of the motion vector field. The alternating direction multiplier method (ADMM) is used to decompose the optimization variables and solve them alternately and iteratively. Dynamic parameter balance is achieved through residual convergence judgment, solving the defects of traditional single-dimensional compensation methods such as parameter conflicts or slow convergence in complex scenes. In this embodiment of the present application, the step of performing a joint deconvolution operation in the spatial and temporal domain is to trace the target motion trajectory forward and backward along the time axis with the current frame as the center to generate an initial back-projection path. The pixel values of different regions are updated in stages according to the phase response characteristics of the deconvolution kernel matrix (for example, the high-frequency kernel prioritizes correcting rotor blur). The calculation is terminated when the mean square error change rate of adjacent iteration results is less than 1%, and the enhanced optical image is output.

[0126] In some embodiments, the light intensity change rate function model is coupled with the background motion vector field to establish a joint optimization equation with light intensity smoothness constraint and motion vector field energy minimization as objective functions, and the alternating direction multiplier method is used to iteratively solve the joint optimization equation until convergence to a stable state, including:

[0127] 401. Design coupling term parameters to fuse the intensity smoothing constraint of the intensity change rate function model with the background motion vector field through weighting factors to form a joint optimization equation;

[0128] In step 401, the coupling term parameter is used to weight the intensity change rate and the background motion vector, balancing the impact of these two constraints on the optimization results. The intensity smoothing constraint limits the gradient difference in the intensity change rate between adjacent pixels, preventing noise amplification caused by sudden changes in illumination. Weighted factor fusion combines these two constraints into a unified objective function using preset weighting coefficients (e.g., 0.6 for the intensity constraint and 0.4 for the motion constraint).

[0129] In an embodiment of the present application, coupling term parameters are designed based on the physical correlation between the light intensity sensor data and the background motion vector field. The light intensity change rate data and the motion vector field data are normalized to the same dimension (such as the range of 0 to 1), and the weight factor is dynamically adjusted based on the scene characteristics (such as when the dynamic lighting is intense, the light intensity constraint weight is increased to 0.7). The light intensity smoothing term and the motion energy term are fused through the weighting factor to form a joint optimization equation to ensure the synergy of the two types of constraints. 402. Decompose the joint optimization equation into a first sub-problem of constraining the light intensity gradient between adjacent pixels by introducing the light intensity smoothing regularization term of the light intensity change rate function model and a second sub-problem of constraining the spatial continuity of the motion vector by introducing the energy term of the background motion vector field;

[0130] In step 402, the first sub-problem is an optimization problem centered around the intensity smoothing regularization term, which is used to constrain the consistency of intensity gradients between adjacent pixels. The second sub-problem is an optimization problem centered around the motion vector field energy term, which is used to ensure the spatial continuity of the motion vector.

[0131] In this embodiment, the joint optimization equation is decomposed into two independently solvable subproblems. With the motion vector field parameters fixed, the light intensity parameters are optimized using gradient descent, and the light intensity distribution is updated by calculating the grayscale gradient difference between adjacent pixels (e.g., a 3×3 neighborhood). With the light intensity parameters fixed, the motion vector field is optimized using the conjugate gradient method. The motion field is updated by constraining the directional consistency of adjacent vectors (e.g., cosine similarity > 0.9). The two subproblems are iterated alternately, and the optimization results are transferred through intermediate variables to achieve global convergence.

[0132] 403. Construct augmented Lagrangian functions for the first subproblem and the second subproblem respectively, and introduce Lagrangian multiplier terms and quadratic penalty terms of linear constraints into the objective function;

[0133] In step 403, the augmented Lagrangian function introduces Lagrangian multipliers and quadratic penalty terms into the objective function, transforming the constrained optimization problem into an unconstrained form. Linear constraints are used to enforce the physical correlation between light intensity parameters and motion parameters, such as requiring light intensity changes to match background motion trends.

[0134] In this embodiment, an augmented Lagrangian function is constructed for each subproblem. A Lagrangian multiplier term is added to the intensity smoothing term to enforce the correlation between the intensity parameter and the motion field. A quadratic penalty term is introduced into the motion energy term to balance the degree of relaxation of the constraints. The Lagrangian multiplier is initialized to 0, and the quadratic penalty term weight coefficient is initialized to 1.0.

[0135] 404. After each variable update, the augmented Lagrange multiplier is updated synchronously, and the weight coefficient of the quadratic penalty term is dynamically adjusted through an adaptive step size adjustment strategy;

[0136] In step 404, the adaptive step size adjustment strategy dynamically adjusts the multiplier update step size based on the iterative residual to accelerate convergence and avoid oscillation. The quadratic penalty weight coefficient controls the strictness of the constraint condition, with a larger weight indicating a stronger constraint.

[0137] In this embodiment, parameters are updated synchronously after each iteration. The Lagrange multiplier is updated proportionally based on the residual between the current solution and the constraints (e.g., for every 10% increase in the residual, the step size increases by 20%). If the rate of change of the residual between two adjacent iterations exceeds a threshold (e.g., 5%), the weight coefficient of the quadratic penalty term is adaptively increased (e.g., from 1.0 to 1.2). An upper limit (e.g., 2.0) is set on the weight to prevent excessive penalty from causing model rigidity.

[0138] 405. Based on the quadratic penalty term, an alternating direction multiplier method is used to iteratively solve the joint optimization equation, and a relative error norm between two adjacent iterations is calculated. When the relative error norm is less than a preset threshold, it is determined that the system has converged to a stable state.

[0139] In step 405, the alternating direction multiplier method is a decomposition-coordination optimization framework that achieves global optimization by alternately solving subproblems. The relative error norm is a measure of the difference between two consecutive iteration results and is used to determine convergence.

[0140] In an embodiment of the present application, the main optimization objective (such as minimizing the bandwidth allocation cost) and the constraints (such as energy consumption not exceeding a threshold) are determined. A quadratic penalty term is added to the objective function to balance the accuracy and stability of the solution. The initial values of the main variables, auxiliary variables and Lagrange multipliers are set, usually taking zero vectors or reasonable values based on prior knowledge. The penalty coefficient (used to control the strength of the penalty term) and the convergence threshold (used to determine whether the algorithm has converged) are set. The auxiliary variables and Lagrange multipliers are fixed to solve the subproblem of the main variables. The main variables and Lagrange multipliers are fixed to solve the subproblem of the auxiliary variables. The Lagrange multiplier is updated based on the residuals of the main variables and the auxiliary variables to adjust the strength of the penalty term. The difference between the main variables of two adjacent iterative solutions is calculated and divided by the norm of the main variables of the current iterative solution to obtain the relative error norm. If the relative error norm is less than the preset threshold (such as one millionth), it is determined that the algorithm has converged to a stable state.

[0141] A drone was flying in dusty weather. Light intensity fluctuated by 300 lux per minute due to dust concentration fluctuations, and background dust particles moved in irregular trajectories. The light intensity change rate was detected to have a dominant frequency of 0.6 Hz (dust concentration periodically varied). The motion vector field indicated an average dust particle velocity of 3 m / s. Dynamic weighting factors (0.65 for intensity constraint and 0.35 for motion constraint) were assigned to construct a joint optimization equation. After decomposing the problem into subproblems, the first subproblem optimized the light intensity parameters using gradient descent, reducing the gradient difference between adjacent pixels by 40%. The second subproblem optimized the motion field using the conjugate gradient method, improving the vector direction consistency to 0.85. An augmented Lagrangian function was constructed with an initial penalty weight of 1.0 and a multiplier initialized to 0. The penalty weight was dynamically adjusted to 1.3 based on the residual change rate, and the multiplier step size was increased by 15%. After 12 ADMM iterations, the relative error norm dropped to 0.008, indicating convergence. The final output light intensity parameters (exposure time 1 / 400s) and motion compensation parameters (filter threshold 5×5) reduced image noise caused by sand and dust interference by 60% and improved target outline clarity by 55%.

[0142] In summary, this scheme solves the problems of convergence speed and stability of coupled optimization of light intensity and motion parameters in dynamic scenes through coupling term parameter design, subproblem decomposition, augmented Lagrangian function construction and adaptive ADMM optimization.

[0143] In order to eliminate image blur caused by high-speed target motion, and to address the problem of insufficient adaptability of traditional deconvolution kernels caused by the complex and changeable target motion trajectories in dynamic scenes, a time-domain deconvolution kernel is constructed based on the spatiotemporal correlation information (such as brightness gradient and chromaticity correlation) extracted from the feature tensor; a motion consistency constraint equation is designed by minimizing the optical flow residual between adjacent frames to ensure that the deconvolution kernel matches the actual motion trajectory of the target; a constrained optimization algorithm (such as the projected gradient method) is used to iteratively solve the pixel-level deconvolution kernel matrix to address the performance bottleneck of the global deconvolution kernel in scenarios with local motion differences.

[0144] In some embodiments, a temporal deconvolution kernel is constructed based on the feature tensor and the UAV motion characteristics. A motion consistency constraint equation is constructed by minimizing the optical flow residual between adjacent frames, and a UAV motion trajectory smoothness constraint is introduced. A constrained optimization algorithm is used to solve the pixel-level temporal deconvolution kernel matrix, including:

[0145] 501. Based on the multi-scale spatiotemporal correlation features of the feature tensor, the spatiotemporal local contrast features are extracted to generate an initial parameter set of the dynamic convolution kernel. In combination with the high-speed translation, hovering jitter, and rotor periodic motion patterns in the UAV motion characteristics, the initial parameter set is subjected to a priori motion trajectory correction. The initial parameter set is mapped to the weight space of the deconvolution kernel to form a time domain deconvolution kernel.

[0146] In step 501, the spatiotemporal local contrast feature is used to characterize the detail changes caused by target motion by analyzing the brightness differences of local regions of the image in the temporal and spatial dimensions (such as the rate of change of pixel values in adjacent frames).

[0147] Dynamic convolution kernel initial parameter set: initial filter parameters generated based on spatiotemporal features, used to describe blur trajectories under different motion modes.

[0148] Weight space mapping: Map the initial parameters to the weight matrix of the deconvolution kernel through nonlinear transformation to adapt to different motion speeds and directions.

[0149] UAV motion characteristics: including high-speed translation, hovering jitter, and rotor periodic motion patterns of the UAV, which are used to perform a priori correction of the motion trajectory of the initial parameter set to ensure that the deconvolution kernel matches the actual motion pattern of the UAV.

[0150] In an embodiment of the present application, first, the spatiotemporal local contrast features are extracted from the feature tensor, the spatiotemporal contrast (the average brightness difference between the current frame and the previous and next two frames) is calculated for the 3×3 neighborhood of each pixel, and high-contrast areas are screened (the threshold is set to 0.2).

[0151] Secondly, the initial parameter set of the dynamic convolution kernel is generated based on the contrast feature, and the motion trajectory is corrected a priori for the initial parameter set in combination with the high-speed translation, hovering jitter and rotor periodic motion patterns in the UAV motion characteristics. For example:

[0152] High-speed translation area: corresponds to the short-term high-frequency kernel, adapting to the blur caused by fast motion;

[0153] Hover jitter area: corresponds to the medium-time and medium-frequency core, adapting to the blur caused by slight jitter;

[0154] Rotor periodic motion region: corresponds to the periodic kernel and adapts to the periodic blur caused by rotor rotation.

[0155] Finally, the corrected initial parameters are mapped to the weight space of the deconvolution kernel through a fully connected neural network to generate a time domain deconvolution kernel.

[0156] 502. Perform a spatiotemporal joint analysis on the temporal deconvolution kernel and the function for minimizing the optical flow residual between adjacent frames, introduce a smoothness constraint on the UAV motion trajectory, and construct a motion consistency constraint equation with the pixel displacement vector as a variable, wherein the UAV motion trajectory smoothness constraint includes an upper limit on the UAV acceleration and a heading continuity condition.

[0157] In step 502, the optical flow residual function measures the difference between the predicted motion vector and the actual pixel displacement between adjacent frames and is used to evaluate the accuracy of motion estimation. The motion consistency constraint equation is an optimization equation constructed with the goal of minimizing the optical flow residual, forcing the deconvolution kernel to match the target's actual motion trajectory. The drone's trajectory smoothness constraints, including the upper limit of the drone's acceleration and the heading continuity condition, are used to constrain the optical flow residual function, ensuring that the deconvolution kernel matches the drone's actual motion pattern.

[0158] In the embodiment of the present application, a dense optical flow algorithm (such as the Farneback algorithm) is first used to calculate the pixel displacement vectors of adjacent frames. For each pixel, the residual (such as the Euclidean distance) between the predicted displacement (derived from the deconvolution kernel parameters) and the actual optical flow vector is calculated.

[0159] Secondly, the sum of squares of the optical flow residual is used as the objective function, and the constraint equation is constructed by combining the spatiotemporal local contrast characteristics and the smoothness constraints of the UAV trajectory (such as the acceleration limit and heading continuity conditions). For example:

[0160] Acceleration upper limit constraint: limits the rate of change of optical flow residuals to avoid abnormal residuals caused by high-speed movement of drones;

[0161] Heading continuity constraint: ensures smooth transition of optical flow residuals when the drone's heading changes, avoiding residual jumps caused by sudden changes in heading.

[0162] Finally, the motion consistency constraint equation is iteratively solved through an optimization algorithm (such as gradient descent method) until the optical flow residual converges to a preset accuracy range.

[0163] 503. Iteratively perform spatial and temporal domain optimization through a constrained optimization algorithm. In each iteration, the weight distribution of the optical flow residual function is adjusted according to the smoothness constraint of the UAV motion trajectory until the mean square error of the optical flow residual converges to a preset accuracy range, and the pixel-level temporal deconvolution kernel matrix that satisfies the motion consistency constraint is output.

[0164] In step 503, spatial domain optimization involves adjusting the weight distribution of the deconvolution kernel within the image plane to enhance the ability to recover local motion features. The weight decay coefficient of high-motion areas is dynamically adjusted based on the smoothness constraints of the drone's trajectory (such as the upper limit on acceleration). Temporal domain optimization involves optimizing the response characteristics of the deconvolution kernel along the time axis to match the target's periodic motion pattern. The time span of the deconvolution kernel is adjusted based on periodic motion features such as the drone's rotor speed.

[0165] In an embodiment of the present application, first, the initial deconvolution kernel parameters and the weight distribution of the optical flow residual function are set, and the weight attenuation coefficient is initialized according to the smoothness constraint of the drone motion trajectory (for example, the weight attenuation coefficient of the high-speed translation area is set to 0.8).

[0166] Secondly, the projected gradient method is used to update the spatial weight of the deconvolution kernel, giving priority to optimizing the weight distribution of high-contrast areas (such as the edge of the drone rotor), and dynamically adjusting the regional weight of the residual function according to the upper limit constraint of the drone acceleration. For example:

[0167] Acceleration out-of-limit region: Reduce the weight attenuation coefficient to suppress residual interference caused by abnormal motion;

[0168] Heading mutation area: Add weight penalty term to avoid residual jump caused by heading discontinuity.

[0169] Next, the time span of the deconvolution kernel is adjusted based on the target motion frequency (e.g., rotor speed 20 Hz), and the time window length is set according to the periodic motion characteristics of the rotor (e.g., 5 frames correspond to half a rotor rotation cycle) to suppress low-frequency background interference.

[0170] Finally, when the mean square error of the optical flow residual changes less than 1% for three consecutive iterations, the optimization is terminated and the pixel-level temporal deconvolution kernel matrix is output.

[0171] A drone flew at 20 m / s in heavy rain. Falling raindrops caused high-frequency background noise, and the rotor motion caused periodic blur. Spatiotemporal local contrast features were extracted from the feature tensor. The rotor region had a mean contrast of 0.5 (threshold 0.2) and an initial short-term high-frequency kernel (spanning three frames). The fuselage region had a mean contrast of 0.1, and a long-term low-frequency kernel (spanning seven frames) was generated. Optical flow residual calculations showed a rotor region residual of 3.2 pixels (threshold 2.5 pixels). A motion consistency constraint was constructed, assigning a 1.5x weight to the rotor region residual. Spatial domain optimization increased the rotor region kernel weight by 30%, and temporal domain optimization matched the rotor's 20 Hz motion frequency. After 15 iterations, the mean squared error of the optical flow residual decreased from an initial 5.6 to 0.8, meeting the convergence criteria. The rotor texture clarity (SSIM) improved from 0.4 to 0.75, and rain noise interference was reduced by 70%. The target tracking and positioning error decreased from 6 pixels to 1.5 pixels, meeting real-time detection requirements.

[0172] In summary, this scheme solves the problem of insufficient adaptability of deconvolution kernel in complex dynamic scenes through spatiotemporal local contrast feature extraction, motion consistency constraint equation construction and spatiotemporal joint optimization.

[0173] In order to accurately capture the rapid motion of the target, and to address the problem that the target motion speed in dynamic scenes is variable and the traditional single time-domain convolution kernel is difficult to effectively capture the spatiotemporal correlation features, a multi-scale dilated convolution kernel is designed (e.g., a short span of 3 frames and a long span of 7 frames). The brightness gradient features (characterizing the brightness and darkness changes caused by motion) and chromaticity correlation features (reflecting the color consistency under illumination fluctuations) of the pixel points are layered extracted in the time dimension. The local and global features of different time spans are integrated to construct a three-dimensional feature tensor with spatiotemporal correlation. This solves the defect of the traditional method's insufficient feature representation ability in complex motion patterns. In some embodiments, time-domain feature modeling is performed on the original optical sequence, and the brightness gradient features and chromaticity correlation features of each pixel point in the time dimension are extracted using a multi-scale dilated convolution kernel. The output is a feature tensor with spatiotemporal correlation, including:

[0174] 601. Decompose the temporal features of the original optical sequence into a luminance gradient component, which is a pixel-by-pixel luminance change matrix obtained by performing a difference operation between adjacent frames, and a chrominance correlation component, which is a chrominance channel covariance matrix that describes cross-frame chrominance correlation.

[0175] In step 601, a pixel-by-pixel luminance change matrix is generated by calculating the luminance difference between corresponding pixels in adjacent frames. This matrix represents the brightness changes caused by target motion. The chromaticity correlation component, based on the covariance matrix of the RGB channels, describes the correlation between colors across frames (e.g., whether changes in the red channel are synchronized with those in the blue channel). Temporal eigendecomposition (TED) splits the original optical sequence into two independent dimensions: luminance and chromaticity, analyzing the effects of motion and illumination separately.

[0176] In an embodiment of the present application, time domain feature decomposition is performed on the original optical sequence, and the brightness difference between adjacent frames (such as the difference between the t-th frame and the t+1 frame) is calculated pixel by pixel for 5 consecutive frames of images to generate a 4-layer brightness change matrix. The RGB channel data of each frame is extracted, and the cross-frame covariance matrix is calculated, such as the covariance value of the red channel and the blue channel in consecutive frames. The chromaticity consistency is quantified, and the brightness gradient component (4-layer matrix) and the chromaticity correlation component (3×3 covariance matrix) are output for subsequent multi-scale convolution processing.

[0177] 602. Perform depthwise separable convolution operations on the luminance gradient component and the chrominance correlation component respectively through each branch of the multi-scale dilated convolution kernel;

[0178] In step 602, multi-scale dilated convolution kernels with different time spans (e.g., a dilation rate of 1 corresponds to a span of 3 frames, and a dilation rate of 2 corresponds to a span of 5 frames) are used to capture both short-term mutations and long-term trends. Depthwise separable convolution decomposes the standard convolution into depthwise convolution (channel-by-channel processing) and pointwise convolution (channel fusion), reducing computational complexity and improving feature differentiation.

[0179] In an embodiment of the present application, depthwise separable convolution is performed on the luminance and chrominance components, and three convolution branches with different expansion rates are set (expansion rates 1, 2, and 3), which correspond to short-, medium-, and long-time feature extraction, respectively. Layer-by-layer convolution is performed on the luminance gradient component (single channel) to extract motion features of different time spans; channel-independent convolution is performed on the chrominance correlation components (3 channels) to retain color correlation, and the outputs of each branch are channel-fused through 1×1 convolution, for example, short-time luminance features and long-time chrominance features are associated across channels.

[0180] 603. Perform cross-scale correlation modeling on the multi-branch output features based on the depthwise separable convolution operation, align the convolution results of different expansion rates according to the time dimension, and output a feature tensor with spatiotemporal correlation;

[0181] In step 603, cross-scale correlation modeling aligns and concatenates feature maps of different time spans along the time axis to construct a spatiotemporal correlation tensor. Temporal alignment uses interpolation or cropping to ensure that feature maps of different scales have the same temporal length, facilitating subsequent fusion.

[0182] In the embodiment of the present application, cross-scale correlation modeling is performed on the multi-branch output, and linear interpolation is performed on the long-term feature map with an expansion rate of 3 to make its time dimension consistent with the short-term feature map. Figure 1 The feature maps with different expansion rates are concatenated according to the channel dimension (such as interpolating from 5 frames to 7 frames) to form a three-dimensional tensor (time × space × channel). The spatiotemporal features are further fused through 3D convolution to output a feature tensor containing global motion trends and local details.

[0183] A drone flew at 15 m / s in a dust storm. The background dust caused the light intensity to fluctuate by 500 lux per minute, resulting in irregular motion blur for both the target and the background. Dust occlusion caused the brightness difference between adjacent frames to reach up to 80 (in the 0-255 range), generating a highly dynamic brightness variation matrix. The yellow hue of the dust increased the covariance between the red and green channels to 0.8 (compared to approximately 0.3 in normal scenes). A short-term dilated convolution (3-frame span) captured the rapid vibration of the drone's rotor (25 Hz frequency). A long-term dilated convolution (7-frame span) extracted the slow diffusion of the dust cloud (0.5 m / s). The long-term feature map was interpolated from 7 frames to 9 frames and concatenated with the short-term feature map. A 3D convolution enhanced the correlation between the high-frequency vibration of the rotor and the low-frequency motion of the dust, outputting a feature tensor. This improved the discriminability of the spatiotemporal features in the rotor region by 50%, and reduced the false detection rate due to dust interference by 65%. The target motion trajectory prediction error is reduced from 5 pixels to 1.2 pixels, meeting the real-time tracking requirements in complex environments.

[0184] In order to enhance the defense capability of low-altitude security systems against adversarial sample attacks, and to address the problem of decreased image recognition accuracy caused by adversarial sample attacks in low-altitude security scenarios, adversarial sample defense features (such as local gradient direction consistency and frequency domain energy concentration) are extracted from enhanced optical images, a perturbation strength evaluation function is constructed to quantify the potential risk of adversarial attacks, and a nonlinear weighted fusion strategy (such as a random forest model or an attention mechanism) is used to integrate multi-dimensional features to generate an adversarial perturbation feature mask, identify high-risk areas, and address the shortcomings of traditional methods in their insufficient ability to detect new adversarial attacks. In some embodiments, adversarial sample defense features are extracted from the enhanced optical image, and the adversarial sample defense features are nonlinearly weighted fused by constructing a perturbation strength evaluation function to form an adversarial perturbation feature mask, including:

[0185] 701. Perform multi-scale spatial filtering on the enhanced optical image, combining an adaptive noise suppression algorithm with a local contrast enhancement operation to generate a denoised standardized optical image;

[0186] In step 701, multiscale spatial filtering uses filters of different sizes (e.g., 3×3, 5×5, and 7×7) to layer the image, preserving detail while suppressing noise. An adaptive noise suppression algorithm dynamically adjusts the filter strength based on the local noise level to avoid detail loss caused by oversmoothing. Local contrast enhancement enhances the visibility of details in local areas of the image through histogram equalization or contrast stretching.

[0187] In an embodiment of the present application, multi-scale spatial filtering is performed on the enhanced optical image, and 3×3, 5×5, and 7×7 Gaussian filters are used to smooth the image, respectively, retaining detail information at different scales, and the filtering intensity is dynamically adjusted based on the local noise variance (such as the standard deviation of pixel values in a 3×3 neighborhood). Strong filtering is used when the noise level is high, and weak filtering is used when the noise level is low. Local histogram equalization is performed on the filtered image to enhance the contrast of the target area (such as the rotor of a drone) and generate a denoised standardized optical image.

[0188] 702. Feature extraction is performed based on the denoised standardized optical image. The high-frequency texture features and low-frequency structural features of the drone's multi-rotor structure and metal reflective properties are screened through a cross-channel attention mechanism to construct an adversarial sample defense feature matrix.

[0189] In step 702, a cross-channel attention mechanism dynamically selects important features in the RGB channels using attention weights to enhance the distinction between high-frequency texture and low-frequency structure. The attention mechanism focuses on key areas of the drone's rotors and fuselage. High-frequency texture features represent image details such as edges and corners and are used to detect anomalous textures introduced by adversarial attacks. Low-frequency structural features describe the overall contours and regional distribution of the image and are used to assess the impact of adversarial attacks on the global structure.

[0190] In an embodiment of the present application, feature extraction is performed based on the denoised standardized optical image, the image is decomposed into three independent RGB channels, the gradient amplitude and direction of each channel are calculated respectively, the weight coefficient of each channel is calculated through the cross-channel attention mechanism (such as the red channel weight 0.6, the green channel 0.3, and the blue channel 0.1), the high-frequency texture and low-frequency structural features are dynamically fused, the filtered features are spliced according to the channel dimension, and the adversarial sample defense feature matrix is generated for subsequent perturbation intensity evaluation.

[0191] 703. Dynamically calculate the weight coefficients of each adversarial sample defense feature channel using the perturbation intensity evaluation function, and use a layer-by-layer reverse gradient accumulation strategy to adjust the weight distribution to achieve nonlinear weighted fusion of the feature space;

[0192] In step 703, the perturbation intensity assessment function calculates the potential risk level of the adversarial attack based on the feature matrix and assigns weights to different feature channels. The layer-by-layer reverse gradient accumulation strategy optimizes the weight distribution through backpropagation, enhancing the detection accuracy of high-perturbation areas.

[0193] In an embodiment of the present application, a disturbance intensity evaluation function is used to perform feature weighted fusion, an initial weight is assigned to each feature channel (such as a high-frequency texture weight of 0.7 and a low-frequency structure weight of 0.3), the gradient value of the feature channel is calculated by back propagation, the weight distribution is dynamically adjusted (such as the high-frequency texture weight is increased to 0.8), and the weighted feature channels are fused according to a nonlinear function (such as Sigmoid) to generate a disturbance intensity distribution map.

[0194] 704. The adversarial sample defense feature matrix after nonlinear weighted fusion is iteratively optimized to ensure semantic consistency between the mask boundary and the original image, and an adversarial perturbation feature mask is output.

[0195] In step 704, the mask boundaries are iteratively optimized: the mask edges are adjusted through multiple iterations to align with the semantic information of the original image (e.g., the target outline). Semantic consistency ensures that the mask area is precisely aligned with the target area in the image (e.g., the drone body or rotor).

[0196] In an embodiment of the present application, multi-dimensional information such as gradient amplitude, pixel intensity distribution, and texture features are extracted from the adversarial sample. The features are weighted by a nonlinear function (such as Sigmoid) to generate a fused feature matrix. A high-dimensional matrix containing the fused features is generated as input for subsequent mask optimization. Based on the fused feature matrix, an initial mask is generated to cover the main area of the adversarial perturbation. The semantic similarity between the mask boundary area and the original image is calculated to evaluate the semantic consistency of the mask. If the semantic similarity is lower than a threshold (such as 0.9), the mask boundary is adjusted to expand or reduce the mask coverage to ensure that the boundary area is aligned with the semantic information of the original image. Repeat the semantic consistency evaluation and boundary adjustment until the semantic similarity between the mask boundary and the original image reaches a preset threshold. Test the mask's defense performance (such as the adversarial sample recognition accuracy) and semantic consistency (such as the similarity between the boundary area and the original image) on the validation set. Output the optimized adversarial perturbation feature mask.

[0197] In summary, this scheme solves the problem of decreased image recognition accuracy caused by adversarial sample attacks through multi-scale filtering, cross-channel attention feature extraction, perturbation strength evaluation and mask optimization.

[0198] In order to improve the accuracy and robustness of target detection in adversarial disturbance scenarios, and to address the problem of image semantic information destruction and decreased target detection accuracy caused by adversarial disturbances in low-altitude security scenarios, the enhanced optical image and the adversarial disturbance feature mask are spatially superimposed to fuse the original image details with the disturbance area identification information; a multi-scale residual network based on an attention mechanism is used to extract multi-scale features (such as local texture and global structure) from the superimposed image, and dynamically focus on key areas (such as drone rotors) through attention weights to output target confidence estimation results, thereby solving the problem of insufficient feature extraction and inaccurate target positioning in traditional methods in adversarial disturbance scenarios. In some embodiments, the enhanced optical image and the adversarial disturbance feature mask are spatially superimposed, and a multi-scale residual network based on an attention mechanism is used to estimate the target confidence of the superimposed image, including:

[0199] 801. Normalize the anti-disturbance feature mask so that the value range of the anti-disturbance feature mask matches the pixel distribution of the enhanced optical image, and generate a superimposed image through a spatial domain superposition operation;

[0200] In step 801, normalization is performed to scale the pixel values of the anti-disturbance feature mask to the same numerical range as the enhanced optical image (e.g., 0–255) to facilitate subsequent overlay operations. Spatial overlay operations combine the normalized mask with the enhanced image on a pixel-wise weighted basis to generate an overlaid image that includes the perturbed regions.

[0201] In this embodiment, the anti-disturbance feature mask is normalized, linearly mapping the mask pixel values from the range 0–1 to 0–255 to match the pixel distribution of the enhanced optical image. The normalized mask is then fused with the enhanced image using a pixel-weighted fusion method (e.g., a mask weight of 0.3 and an image weight of 0.7) to generate a superimposed image. The perturbed areas are marked in semi-transparent red, while the unperturbed areas retain the original image details.

[0202] 802. Construct a multi-scale processing module to perform a multi-scale residual network operation based on an attention mechanism on the superimposed image to generate a multi-scale fusion feature tensor;

[0203] In step 802, the multi-scale processing module includes convolutional layers with different receptive fields (e.g., 3×3, 5×5, and 7×7) to extract local details and global structural features. The attention mechanism uses dynamic weight allocation to focus on key areas (e.g., drone rotors), enhancing the targeted nature of feature extraction. The multi-scale fusion feature tensor concatenates feature maps of different scales into a three-dimensional tensor, preserving both local and global information.

[0204] In this embodiment, a multi-scale processing module is constructed and a residual network is implemented. 3×3, 5×5, and 7×7 convolution kernels are used to extract local texture and global contour features. An attention mechanism is used to calculate weight coefficients for features at each scale (e.g., a weight of 0.8 for the rotor region and 0.2 for the background region), dynamically fusing features from key regions. Feature maps at different scales are concatenated along the channel dimension to generate a multi-scale fused feature tensor for subsequent confidence estimation.

[0205] 803. Input the multi-scale fusion feature tensor into a confidence estimation unit, perform local feature statistics calculation, and generate a target confidence estimation value.

[0206] In step 803, the multi-scale processing module includes convolutional layers with different receptive fields (e.g., 3×3, 5×5, and 7×7) to extract local details and global structural features. The attention mechanism uses dynamic weight allocation to focus on key areas (e.g., the drone's rotor blades), enhancing the targeted nature of feature extraction. The multi-scale fusion feature tensor concatenates feature maps of different scales into a three-dimensional tensor, preserving both local and global information.

[0207] In an embodiment of the present application, a multi-scale processing module is constructed and a residual network operation is performed. 3×3, 5×5, and 7×7 convolution kernels are used to extract local texture and global contour features, respectively. The weight coefficients of the features at each scale are calculated through the attention mechanism (such as a weight of 0.8 for the rotor area and a weight of 0.2 for the background area). The features of key areas are dynamically fused, and the feature maps of different scales are spliced according to the channel dimension to generate a multi-scale fused feature tensor for subsequent confidence estimation.

[0208] A drone was flying in strong winds. An adversarial attack injected high-frequency noise (PSNR = 28dB) into the image, causing target detection failure. The pixel values of the adversarial perturbation feature mask were mapped from 0–1 to 0–255. The mask and the enhanced image were superimposed with a weighting of 0.3:0.7, and the perturbed areas were marked in semi-transparent red. 3×3, 5×5, and 7×7 convolution kernels were used to extract rotor texture and fuselage contour features. The weight of the rotor region was increased to 0.8, while the weight of the background region was reduced to 0.2. The mean and variance of the features were calculated for a 3×3 neighborhood of the rotor region, and the mean was significantly higher than that of the background region. The confidence level of the rotor region was increased to 0.9, while the confidence level of the background noise region was reduced to 0.2. The detection accuracy of the rotor region was improved to 95%, and the false detection rate caused by the adversarial attack was reduced by 80%. Target detection accuracy was restored from 60% to 90%, meeting low-altitude security requirements.

[0209] In summary, this scheme solves the problem of decreased target detection accuracy in adversarial disturbance scenarios through normalized superposition, multi-scale residual network and confidence estimation.

[0210] To improve the target recognition accuracy and stability of low-altitude security systems in complex dynamic scenarios, and to address the problem of decreased drone recognition accuracy caused by environmental interference (such as sudden changes in illumination and background motion) and adversarial attacks (such as high-frequency noise and image tampering) in low-altitude security scenarios, a dynamic threshold screening and multi-frame fusion strategy are used based on target confidence estimation results (such as local area confidence scores) to eliminate low-confidence false detection areas and enhance the recognition weight of high-confidence target areas. The recognition results are optimized by combining spatiotemporal context information to address the defects of traditional methods such as high false detection rates and poor robustness in complex scenarios. In some embodiments, drone recognition results with robustness against environmental interference and adversarial attacks are generated based on the target confidence estimation results, including:

[0211] 901. Based on the target confidence estimation results, a sliding window mechanism is used to calculate the environmental interference intensity index in real time. When the fluctuation of the confidence of consecutive frames exceeds the preset threshold, the adaptive threshold adjustment algorithm is activated;

[0212] In step 901, the target confidence change trend is analyzed in real time through a sliding window (such as a window length of 5 frames and an overlap rate of 50%). The environmental interference intensity index quantifies the intensity of environmental interference based on the confidence fluctuation amplitude (such as the confidence difference between adjacent frames). The adaptive threshold adjustment algorithm dynamically adjusts the confidence threshold according to the interference intensity to screen high-confidence target areas. In an embodiment of the present application, the environmental interference intensity index is calculated based on the target confidence estimation result, and a sliding window analysis is performed on the confidence data of 5 consecutive frames to calculate the window built-in confidence mean and variance. If the confidence difference between adjacent frames exceeds a preset threshold (such as 0.2), it is determined that environmental interference exists, and the interference intensity index is set to the absolute value of the difference. When the interference intensity index is >0.3, the adaptive threshold adjustment algorithm is started to increase the confidence threshold from 0.6 to 0.7 to screen high-confidence areas.

[0213] 902. Perform interference pattern recognition on the confidence abnormal area through an adaptive threshold adjustment algorithm to generate an environmental interference feature mask;

[0214] In step 902, image regions with confidence outlier fluctuations exceeding a preset threshold are identified as potential interference areas. Interference pattern recognition identifies interference types (e.g., sudden changes in illumination, background motion) through feature analysis (e.g., texture and motion trajectory). The environmental interference feature mask is used to create a binary mask marking the interference areas, assisting in subsequent identification optimization.

[0215] In this embodiment, interference pattern recognition is performed on areas with abnormal confidence levels. Texture features (such as gradient direction histograms) and motion features (such as optical flow vectors) are extracted from these areas. A random forest model (consisting of 100 decision trees) is then used to identify the type of interference. For example, areas with sudden changes in illumination have concentrated gradient directions, while areas with background motion have consistent optical flow vectors. Based on the recognition results, an environmental interference feature mask is generated, marking the interference areas (e.g., areas with sudden changes in illumination are marked in yellow, and areas with background motion are marked in green).

[0216] 903. Start the adversarial sample detection module for low-confidence targets, use the gradient backpropagation method to analyze the sensitivity of feature space perturbations, and generate a robustness evaluation coefficient.

[0217] In step 903, the adversarial example detection module detects potential adversarial attack regions by analyzing the sensitivity of feature space perturbations. The gradient backpropagation method calculates the sensitivity of features to perturbations through backpropagation, quantifying the impact of adversarial attacks. The robustness evaluation coefficient represents an indicator of the target region's resistance to adversarial attacks; a higher coefficient indicates greater robustness.

[0218] In this embodiment, adversarial sample detection is initiated for low-confidence targets. Backpropagation is used to calculate the gradient of the feature to the perturbation, quantifying the perturbation sensitivity (e.g., a gradient magnitude > 0.5 indicates high sensitivity). A robustness evaluation coefficient is generated based on the gradient value. For example, the coefficient is set to 1.0 for regions with a gradient magnitude < 0.2, and to 0.3 for regions with a gradient magnitude > 0.5. The robustness evaluation coefficient is output for each target region for subsequent recognition optimization.

[0219] 904. Combining the target confidence estimation result, the environmental interference feature mask and the robustness evaluation coefficient, a drone recognition result with robustness against environmental interference and counterattack is generated.

[0220] In step 904, the confidence estimate, interference feature mask, and robustness coefficient are combined through joint optimization to generate a robust recognition result. Robustness against environmental interference and adversarial attacks is enhanced through dynamic screening and fusion strategies, improving the stability of recognition results in complex scenarios.

[0221] In this embodiment, multi-dimensional information is combined to generate robust recognition results. High-confidence target areas (e.g., confidence > 0.7 and robustness > 0.8) are selected based on a confidence threshold and robustness coefficient. These results are then fused with interference feature masks. For example, low-confidence targets in areas with sudden changes in illumination are ignored, while high-confidence targets in areas with background motion are retained. This generates drone recognition results that are robust against environmental interference and attacks, which are used for subsequent tracking and decision-making.

[0222] A drone was flying in heavy rain. Raindrops caused high-frequency image noise (PSNR = 28dB), and light intensity fluctuated by 300 lux per minute due to cloud cover. The average confidence score for five consecutive frames fluctuated by 0.25 (with a threshold of 0.2), and the interference intensity index was set to 0.25. The confidence threshold was increased from 0.6 to 0.7 to screen for high-confidence areas. The gradient direction in areas with sudden illumination changes was concentrated, while the optical flow vectors in background raindrop areas were dispersed. The interference of sudden illumination changes and raindrops was identified to generate an environmental interference feature mask. The gradient amplitude in the raindrop area was 0.6, and the robustness coefficient was set to 0.3. Coefficient generation: The gradient amplitude in the drone body area was 0.1, and the robustness coefficient was set to 1.0. Target areas with a confidence score greater than 0.7 and a robustness coefficient greater than 0.8 were retained. Low-confidence targets in the raindrop area were ignored to generate robust recognition results. This improved drone recognition accuracy from 65% to 92%, and the false detection rate due to raindrop interference was reduced by 85%. The false detection rate in areas with sudden changes in illumination is reduced by 70%, meeting the real-time detection needs in complex environments.

[0223] In summary, this solution solves the problem of decreased drone recognition accuracy caused by environmental interference and adversarial attacks in low-altitude security scenarios through sliding window interference detection, adaptive threshold adjustment, adversarial sample detection, and multi-dimensional joint optimization.

[0224] Figure 2 The present invention provides a schematic diagram of a low-altitude security drone identification device. Figure 2 As shown, the device includes:

[0225] A generation module 21 is configured to generate, in real time, a set of optical compensation parameters optimized for drone imaging based on a joint optimization equation for the rate of change of ambient light intensity and the background motion vector field in a low-altitude security scenario. The generation module 21 is further configured to acquire an original optical sequence within a target capture window using the optical compensation parameter set, construct a temporal deconvolution kernel based on the drone's motion characteristics, and generate an enhanced optical image with motion blur resistance.

[0226] A processing module 22 extracts adversarial sample defense features from the enhanced optical image, and performs nonlinear weighted fusion on the adversarial sample defense features by constructing a perturbation strength evaluation function to form an adversarial perturbation feature mask;

[0227] An estimation module 23 performs a spatial superposition operation on the enhanced optical image and the anti-disturbance feature mask, and uses a multi-scale residual network based on an attention mechanism to estimate the target confidence of the superposed image;

[0228] The generation module 21 is further configured to generate a drone identification result that is robust against environmental interference and attacks based on the target confidence estimation result.

[0229] Figure 2 The drone identification device for low-altitude security can perform Figure 1 The implementation principles and technical effects of the method for identifying drones for low-altitude security described in the illustrated embodiment are not further elaborated. The specific manner in which each module and unit performs operations in the aforementioned embodiment of a drone identification device for low-altitude security has been described in detail in the relevant embodiments of the method and will not be further elaborated here.

[0230] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A drone identification method for low-altitude security, characterized in that: include: Based on a joint optimization equation for the ambient light intensity change rate and the background motion vector field in low-altitude security scenarios, a set of optical compensation parameters optimized for drone imaging is generated in real time. The joint optimization equation is established by coupling the light intensity change rate function model with the background motion vector field, with light intensity smoothing constraints and motion vector field energy minimization as the objective functions. The original optical sequence is acquired within the target capture window by using the optical compensation parameter set, and a time domain deconvolution kernel is constructed in combination with the motion characteristics of the UAV to generate an enhanced optical image with anti-motion blur characteristics; Extracting adversarial sample defense features from the enhanced optical image, and performing nonlinear weighted fusion on the adversarial sample defense features by constructing a perturbation strength evaluation function to form an adversarial perturbation feature mask; Performing a spatial superposition operation on the enhanced optical image and the adversarial perturbation feature mask, and using a multi-scale residual network based on an attention mechanism to estimate the UAV target confidence on the superposed image; Based on the target confidence estimation results, a drone recognition result with robustness against environmental interference and adversarial attacks is generated.

2. The method according to claim 1, characterized in that Based on the joint optimization equation of the ambient light intensity change rate and the background motion vector field in low-altitude security scenes, a set of optical compensation parameters optimized for drone imaging is generated in real time, including: A multi-channel light intensity sensor is used to collect ambient light intensity data in low-altitude security scenarios in real time. The frequency domain characteristics of the light intensity change rate are extracted based on time series analysis, and a light intensity change rate function model is constructed. Performing multi-scale pyramid decomposition on the light intensity change rate function model, calculating the initial motion vector of the sparse feature points on each pyramid layer to form a background motion vector field; The light intensity change rate function model is coupled with the background motion vector field to establish a joint optimization equation with light intensity smoothing constraint and motion vector field energy minimization as objective functions, and the alternating direction multiplier method is used to iteratively solve the joint optimization equation until convergence to a stable state; According to the output parameters of the joint optimization equation, the exposure time, gain coefficient and filtering threshold of the optical compensation module are dynamically adjusted to generate a set of optical compensation parameters optimized for UAV imaging.

3. The method according to claim 1, characterized in that The original optical sequence is acquired within the target capture window by using the optical compensation parameter set, and a time domain deconvolution kernel is constructed in combination with the UAV motion characteristics to generate an enhanced optical image with anti-motion blur characteristics, including: Based on the optical compensation parameter set, N frames of optical signals continuously collected within the target capture window are subjected to channel-by-channel enhancement processing, wherein each frame of optical signal is separated into three channels, RGB, and then a dot product operation is performed with the gain coefficient matrix of the corresponding channel to generate an original optical sequence; Performing time-domain feature modeling on the original optical sequence, extracting the brightness gradient feature and chromaticity correlation feature of each pixel in the time dimension through a multi-scale dilated convolution kernel, and outputting a feature tensor with spatiotemporal correlation; A temporal deconvolution kernel is constructed based on the feature tensor and the UAV motion characteristics. A motion consistency constraint equation is constructed by minimizing the optical flow residual between adjacent frames. A smoothness constraint on the UAV motion trajectory is introduced, and a constrained optimization algorithm is used to solve the pixel-level temporal deconvolution kernel matrix. The time-domain deconvolution kernel is subjected to a joint space-time deconvolution operation with the original optical sequence, wherein a back-projection calculation is performed along the time axis for each spatial position, and the phase response characteristics of the deconvolution kernel are adjusted through an iterative update strategy to generate an enhanced optical image with anti-motion blur characteristics.

4. The method according to claim 2, characterized in that The light intensity change rate function model is coupled with the background motion vector field, and a joint optimization equation with light intensity smoothness constraint and motion vector field energy minimization as objective functions is established. The alternating direction multiplier method is used to iteratively solve the joint optimization equation until convergence to a stable state, including: Design the coupling term parameters, fuse the intensity smoothness constraint of the intensity change rate function model with the background motion vector field through weighting factors, and form a joint optimization equation; Decomposing the joint optimization equation into a first subproblem of constraining the light intensity gradient between adjacent pixels by introducing a light intensity smoothing regularization term of the light intensity change rate function model and a second subproblem of constraining the spatial continuity of the motion vector by introducing an energy term of the background motion vector field; Constructing augmented Lagrangian functions for the first subproblem and the second subproblem respectively, and introducing Lagrangian multiplier terms and quadratic penalty terms of linear constraints into the objective function; After each variable update, the augmented Lagrange multiplier is updated synchronously, and the weight coefficient of the quadratic penalty term is dynamically adjusted through an adaptive step size adjustment strategy; Based on the quadratic penalty term, the alternating direction multiplier method is used to iteratively solve the joint optimization equation, and the relative error norm in two adjacent iterations is calculated. When the relative error norm is less than a preset threshold, it is determined that the solution has converged to a stable state.

5. The method according to claim 3, characterized in that Based on the feature tensor and the UAV motion characteristics, a temporal deconvolution kernel is constructed. The motion consistency constraint equation is constructed by minimizing the optical flow residual between adjacent frames. The smoothness constraint of the UAV motion trajectory is introduced, and the pixel-level temporal deconvolution kernel matrix is obtained by using a constrained optimization algorithm, including: Based on the multi-scale spatiotemporal correlation features of the feature tensor, the spatiotemporal local contrast features are extracted to generate an initial parameter set of the dynamic convolution kernel. Combined with the high-speed translation, hovering jitter and rotor periodic motion patterns in the UAV motion characteristics, the initial parameter set is subjected to a priori motion trajectory correction, and the initial parameter set is mapped to the weight space of the deconvolution kernel to form a time domain deconvolution kernel. The temporal deconvolution kernel is combined with the function of minimizing the residual optical flow between adjacent frames for spatiotemporal joint analysis. The smoothness constraint of the UAV motion trajectory is introduced to construct a motion consistency constraint equation with the pixel displacement vector as a variable. The smoothness constraint of the UAV motion trajectory includes the upper limit of the UAV acceleration and the heading continuity condition. The constrained optimization algorithm is used to iteratively perform spatial and temporal domain optimization. In each iteration, the weight distribution of the optical flow residual function is adjusted according to the smoothness constraint of the UAV's motion trajectory until the mean square error of the optical flow residual converges to a preset accuracy range, and the pixel-level temporal deconvolution kernel matrix that satisfies the motion consistency constraint is output.

6. The method according to claim 3, characterized in that Perform time domain feature modeling on the original optical sequence, extract the brightness gradient features and chromaticity correlation features of each pixel in the time dimension through a multi-scale dilated convolution kernel, and output a feature tensor with spatiotemporal correlation, including: The temporal features of the original optical sequence are decomposed into a luminance gradient component, which is obtained by performing inter-frame difference operations to obtain a pixel-by-pixel luminance change matrix, and a chrominance correlation component, which uses the covariance matrix of the chrominance channel to describe the cross-frame chrominance correlation. Performing depth-separable convolution operations on the luminance gradient component and the chrominance correlation component through each branch of the multi-scale dilated convolution kernel; Based on the depthwise separable convolution operation, cross-scale correlation modeling is performed on the multi-branch output features, the convolution results of different expansion rates are aligned according to the time dimension, and a feature tensor with spatiotemporal correlation is output.

7. The method according to claim 1, wherein Extracting adversarial sample defense features from the enhanced optical image, and performing nonlinear weighted fusion on the adversarial sample defense features by constructing a perturbation strength evaluation function to form an adversarial perturbation feature mask, including: Perform multi-scale spatial filtering on the enhanced optical image, combine the adaptive noise suppression algorithm with the local contrast enhancement operation to generate a denoised and standardized optical image; Feature extraction is performed based on the denoised standardized optical image. The high-frequency texture features and low-frequency structural features of the drone's multi-rotor structure and metal reflective properties are screened through a cross-channel attention mechanism to construct an adversarial sample defense feature matrix. The attention mechanism focuses on key parts of the drone's rotors and fuselage. The perturbation intensity evaluation function is used to dynamically calculate the weight coefficient of each adversarial sample defense feature channel, and the layer-by-layer reverse gradient accumulation strategy is used to adjust the weight distribution to achieve nonlinear weighted fusion of the feature space; The adversarial sample defense feature matrix after nonlinear weighted fusion is iteratively optimized to ensure semantic consistency between the mask boundary and the original image, and the adversarial perturbation feature mask is output.

8. The method according to claim 1, wherein The enhanced optical image and the adversarial perturbation feature mask are spatially superimposed, and a multi-scale residual network based on an attention mechanism is used to estimate the UAV target confidence of the superimposed image, including: Normalizing the anti-disturbance feature mask so that the value range of the anti-disturbance feature mask matches the pixel distribution of the enhanced optical image, and generating a superimposed image through a spatial domain superposition operation; Constructing a multi-scale processing module to perform a multi-scale residual network operation based on an attention mechanism on the superimposed image to generate a multi-scale fusion feature tensor; The multi-scale fusion feature tensor is input into the confidence estimation unit, local feature statistics are calculated, and a UAV target confidence estimation value is generated.

9. The method according to claim 1, wherein Based on the drone target confidence estimation results, a drone recognition result with robustness against environmental interference and adversarial attacks is generated, including: Based on the UAV target confidence estimation results, a sliding window mechanism is used to calculate the environmental interference intensity index in real time. When the confidence fluctuation of consecutive frames exceeds the preset threshold, the adaptive threshold adjustment algorithm is activated; The interference pattern recognition of the confidence abnormal area is performed through the adaptive threshold adjustment algorithm to generate the environmental interference feature mask; The adversarial sample detection module is activated for low-confidence targets, and the gradient backpropagation method is used to analyze the sensitivity of feature space perturbations and generate robustness evaluation coefficients. The drone target confidence estimation result, the environmental interference feature mask and the robustness evaluation coefficient are combined to generate a drone recognition result that is robust against environmental interference and counterattacks.

10. A drone identification system for low-altitude security, characterized by: include: A generation module for generating, in real time, a set of optical compensation parameters optimized for UAV imaging based on a joint optimization equation for the ambient light intensity change rate and the background motion vector field in a low-altitude security scenario. The joint optimization equation is established by coupling a light intensity change rate function model with the background motion vector field to establish an optimization equation with light intensity smoothing constraints and motion vector field energy minimization as objective functions. The generation module is further configured to acquire an original optical sequence within a target capture window using the optical compensation parameter set, construct a time-domain deconvolution kernel based on the motion characteristics of the drone, and generate an enhanced optical image with motion blur resistance. a processing module, extracting adversarial sample defense features from the enhanced optical image, and performing nonlinear weighted fusion on the adversarial sample defense features by constructing a perturbation strength evaluation function to form an adversarial perturbation feature mask; an estimation module, performing a spatial superposition operation on the enhanced optical image and the anti-disturbance feature mask, and performing target confidence estimation on the superposed image using a multi-scale residual network based on an attention mechanism; The generation module is also used to generate a drone identification result that is robust against environmental interference and attacks based on the target confidence estimation result.

Images