Flapping wing target detection and identification method based on laser micro-motion and polarization-infrared feature combination

Abstract

The flapping wing target detection and recognition method based on laser micro-motion and polarization-infrared feature combination belongs to the field of laser detection and recognition.The implementation steps of the method are as follows: step one, selecting the identification features of laser micro-motion, polarization and infrared;step two, proposing a multi-modal feature fusion criterion, and distributing weights to the features;step three, constructing a data set, and combining a machine learning algorithm to perform target recognition.The present application utilizes the differences in micro-motion, polarization and infrared modal features between flying birds and flapping wing unmanned aerial vehicles, and combines a machine learning algorithm to effectively detect and recognize flying birds and flapping wing unmanned aerial vehicles.Compared with the existing flapping wing target recognition method, the flapping wing target detection and recognition method of the present application significantly improves the detection and recognition accuracy of flying birds and flapping wing unmanned aerial vehicles, and provides a direction for the detection and recognition of flying birds and flapping wing unmanned aerial vehicles.

Description

Technical Field

[0001] This invention relates to a flapping-wing target detection and recognition method based on multimodal feature fusion, specifically a flapping-wing target detection and recognition method based on laser micro-motion and polarization-infrared feature combination, belonging to the field of laser detection and recognition. Background Technology

[0002] With the development of drone detection and identification capabilities, and inspired by the excellent flight skills of insects and birds in nature, the biomimetic performance of flapping-wing drones has been greatly improved. Flapping-wing drones possess advantages such as strong stealth, high maneuverability, strong adaptability, and low noise. At the same time, this also brings greater challenges to public safety and military defense.

[0003] Teams at Shanghai Jiao Tong University have developed "feather-PVDF bio-hybrid mechanoreceptors," and Northwestern Polytechnical University has developed "Carrier Pigeon," both flapping-wing drones capable of blending into flocks of birds to deceive microwave radar systems. By using certain techniques to fuse deceptive feathers or coatings with flapping-wing drones, they can become virtually indistinguishable from real birds, truly becoming "bird spies" undetectable by optical systems and radar. Therefore, distinguishing between birds and flapping-wing drones faces significant challenges, primarily due to the high similarity in their flight patterns and motion characteristics. The flapping-wing drones, mimicking the wings of birds to generate lift and propulsion, make their micro-motion characteristics extremely similar to those of birds in terms of frequency and amplitude. Furthermore, the diverse signal variations of flapping-wing drones and birds under different flight modes further increase the difficulty of differentiation.

[0004] Therefore, how to accurately detect and identify birds and flapping-wing drones is a problem that needs to be explored in the field of detection and identification.

[0005] Currently, birds and biomimetic flapping-wing drones exhibit a high degree of similarity in their motion characteristics. Traditional millimeter-wave radar detection methods, limited by physical characteristics such as long wavelengths and insufficient resolution, struggle to effectively capture the fine micro-motion characteristics of flapping-wing targets and distinguish between birds and biomimetic flapping-wing drones. In contrast, laser micro-Doppler technology, with its high-resolution advantage, can accurately resolve kinematic parameters such as flapping frequency and amplitude, thus achieving higher spatial and temporal resolution for target detection. However, most existing methods are limited to single-mode time-frequency feature analysis. This single-mode analytical framework has certain limitations: firstly, it ignores the essential differences in polarization scattering characteristics between mechanical structures and biological tissues; secondly, it cannot integrate the optical properties of the target material; and thirdly, it is susceptible to interference from complex environments and background heat sources, leading to distortion in the feature space.

[0006] To address the aforementioned problems of traditional single-mode detection methods, this invention proposes a flapping-wing target detection and identification method based on the combination of laser micro-motion and polarization-infrared features, which can improve the accuracy of bird and flapping-wing UAV identification. Summary of the Invention

[0007] The purpose of this invention is to solve the above-mentioned problems in the background technology by proposing a flapping-wing target detection and identification method based on the combination of laser micro-motion and polarization-infrared features.

[0008] This invention utilizes the differences in micro-motion, polarization, and infrared modal characteristics between birds and flapping-wing UAVs, and combines them with machine learning algorithms to effectively detect and identify birds and flapping-wing UAVs, providing a direction for the detection and identification of birds and flapping-wing UAVs.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A flapping-wing target detection and identification method based on laser micro-motion and polarization-infrared feature combination is described below:

[0011] Step 1: Select the recognition features of three modes: laser micro-motion, polarization, and infrared.

[0012] Step 2: Propose a multimodal feature fusion criterion and assign weights to the features;

[0013] Step 3: Construct a dataset and combine it with machine learning algorithms for target recognition.

[0014] Furthermore, in step one, the selection of recognition features for three modes—laser micro-motion, polarization, and infrared—is specifically as follows:

[0015] (1) Regarding the identification characteristics of the laser micro-motion mode;

[0016] For a single-stage flapping-wing UAV, the selection of bi-peak consistency without wing sweeping frequency shift is used as the discrimination index for a single-stage flapping-wing UAV.

[0017] For two-stage flapping-wing UAVs, the absence of bi-peak consistency and no sweeping wing frequency shift is selected as the discrimination index for two-stage flapping-wing UAVs.

[0018] For birds, the non-double-peak consistency with swept-wing frequency shift is selected as the bird discrimination index;

[0019] (2) Regarding the identification features of the polarization mode;

[0020] Natural biological materials with a debiasing value higher than 0.5 were selected as the discrimination index for birds.

[0021] Artificial composite materials with a deflection value of less than 0.5 were selected as the discrimination index for flapping-wing UAVs;

[0022] (3) Regarding the identification features of the infrared mode;

[0023] The texture structure features of the target surface are selected as the discrimination index between birds and flapping-wing UAVs. The infrared image is preprocessed and the target is extracted using a target locking and image cropping method based on maximizing brightness.

[0024] Furthermore, the selection of the target surface texture structure features as the discrimination index between birds and flapping-wing UAVs, and the use of a target locking and image cropping method based on maximizing brightness to preprocess the infrared image and extract the target, specifically:

[0025] (3.1) First, read the experimental image data and set the target size, size tolerance, brightness threshold and center cropping expansion according to the user-defined target size, size tolerance, brightness threshold and center cropping expansion amount; then convert the image to grayscale, find all the brightest points, calculate the average of the row coordinates of these brightest points, and round it to the nearest integer as the row coordinates of the center point. Similarly, calculate the average of the column coordinates of all the brightest points and round it to the nearest integer as the column coordinates of the center point. Thus, the average center of the brightest point is obtained.

[0026] (3.2) Locate the center of the average brightest point, obtain the original image size, determine the cropping range and adjust the boundary based on the above center cropping expansion amount, and complete the image cropping centered on the center of the average brightest point;

[0027] (3.3) Target localization: First, adaptive histogram equalization is used to enhance contrast; then dynamic thresholding is performed: the top 10% of pixels with the highest image brightness are selected, and the threshold is set to the minimum value of these 10% of pixels, and the threshold is not less than 200, generating a binary mask; then morphological processing is performed, and noise is removed and the boundaries are smoothed by combining opening and closing operations; finally, region analysis and screening are performed on images whose size is within ±50% of the target size and whose aspect ratio is between 0.5 and 2; the target size is 120×120 pixels;

[0028] (3.4) Finally, target extraction is performed, the candidate region closest to the center of the image is selected, the bounding box is expanded, boundary protection is performed to prevent cross-boundary, and the final cropping is performed.

[0029] Furthermore, the texture structure features of the target surface include contrast, inverse difference moment, energy, and entropy;

[0030] For birds, the mean contrast ratio ranges from 0.269 to 0.312, the mean inverse moment ranges from 0.931 to 0.946, the mean energy ranges from 0.398 to 0.538, and the mean entropy ranges from 1.683 to 2.225.

[0031] For flapping-wing UAVs, the average contrast ratio ranges from 0.027 to 0.068, the average inverse moment ranges from 0.969 to 0.989, the average energy ranges from 0.599 to 0.916, and the average entropy ranges from 0.391 to 1.122.

[0032] Furthermore, the formula for calculating the deflection degree is as follows:

[0033] (1)

[0034] In equation (1), Indicates the degree of deflection. Indicates degree of polarization; polarization degree (2)

[0035] In equation (2), S0 represents the total incident light intensity, S1 represents the linearly polarized light intensity in the 0° and 90° directions, S2 represents the linearly polarized light intensity in the 45° and 135° directions, and S3 represents the circularly polarized light intensity in both left and right directions.

[0036] Furthermore, , , , The calculation formulas are as follows:

[0037] (3)

[0038] (4)

[0039] ; (5)

[0040] (6)

[0041] In equations (3)-(6), i is the gray value of the reference pixel or the starting pixel, and j is the gray value of the neighboring pixel paired with the reference pixel i under a specific spatial relationship. Represents pixel pairs in the grayscale matrix The probability of.

[0042] Furthermore, in step two, the multimodal feature fusion criterion is proposed, and weights are assigned to the features; specifically:

[0043] By combining the features of the above modalities with machine learning, and adopting a multi-dimensional information fusion criterion with target micro-motion modal features as the main factor and polarization and infrared features as auxiliary factors, efficient and accurate target classification is achieved; and weight allocation is performed based on the physical salience of the features.

[0044] Furthermore, the weight allocation is based on the physical saliency of the features, combined with machine learning algorithm training. The weight allocation principle is as follows:

[0045] The total weight of the micro-motion feature is 0.70, of which the weight of the bimodal consistency is 0.55 and the weight of the swept-wing frequency shift is 0.15;

[0046] The polarization feature weight is 0.10;

[0047] The total weight of image texture features is 0.20, of which the weight of contrast is 0.15, the weight of inverse moment is 0.02, the weight of energy is 0.01, and the weight of entropy is 0.02.

[0048] Furthermore, in step three, the construction of the dataset and the combination of machine learning algorithms for target recognition specifically involves:

[0049] First, import the training dataset containing features of birds, single-stage drones, and two-stage drones. Randomize the data by shuffling the order, normalize the data, and set feature weights of 0.55, 0.15, 0.10, 0.15, 0.02, 0.01, and 0.02. Apply the weights to the normalized features, import the test set, process the test set using the same normalization parameters as the training set, and apply the same feature weights for preprocessing. Next, train three algorithm models: MLP, RF, and SVM. Finally, use the trained models to predict the test set and calculate the accuracy using the following formula.

[0050] Correct predictions / total number of samples = classification accuracy;

[0051] Both the number of correct predictions and the total number of samples can be reflected in the program. The total number of samples is the total number of samples in the test set, and the number of correct predictions is obtained in the training process of the trained model predicting the test set.

[0052] The advantages of this invention compared to existing technologies are as follows: The flapping-wing target detection and recognition method of this invention uses the micro-motion characteristics inherent in complex flapping-wing motion targets as its core, combining polarization and infrared features, and integrating machine learning algorithms to fuse the features of the three modalities for target recognition. The flapping-wing target detection and recognition method of this invention, based on the combination of laser micro-motion and polarization-infrared features, significantly improves the accuracy compared to feature recognition methods using only a single micro-motion modality. In the absence of noise, the accuracy of single-modal features using RF, SVM, and MLP algorithms is 58.83%, 59.83%, and 57.5%, respectively; while the accuracy of multimodal features after fusing RF, SVM, and MLP algorithms is 91.83%, 93.17%, and 93.50%, respectively. (These figures are calculated by training the algorithm using the training set and weights, and then applying the trained algorithm to the test set.) The specific calculation process of the multimodal feature algorithm is as follows: First, import the training dataset, shuffle the data order to randomize it, normalize the data, and set feature weights of 0.55, 0.15, 0.10, 0.15, 0.02, 0.01, and 0.02. Apply the weights to the normalized features, import the test set, process the test set using the normalization parameters of the training set, and apply the same feature weights for preprocessing. Next, train three models: Random Forest (RF), Support Vector Machine (SVM), and others.

[0053] The three models—Multilayer Perceptron (MLP), Multilayer Perceptron (MLP), and Multilayer Perceptron (MLP)—are existing machine learning models. Finally, the trained models are used to predict the test set, and the accuracy rate mentioned above is calculated.

[0054] Compared to existing flapping-wing target recognition methods, the flapping-wing target detection and recognition method of this invention, based on the combination of laser micro-motion and polarization-infrared features, significantly improves the detection and recognition accuracy of birds and flapping-wing UAVs. Attached Figure Description

[0055] Figure 1 Time-frequency graphs showing the characteristics of birds and drones; where: Figure 1 (a) is a time-frequency diagram of the characteristics of a single-wing flapping-wing UAV. Figure 1 (b) is a time-frequency diagram of bird characteristics. Figure 1 (c) is a time-frequency diagram of the characteristics of a two-stage flapping-wing UAV.

[0056] Figure 2 This is a depolarization diagram of natural biomaterials and artificial composite materials; where: Figure 2 (a) is a debiasing diagram of downy biological materials. Figure 2(b) is a debiasing diagram of hard-feathered biological materials. Figure 2 (c) is a diagram of the de-skewing degree of artificial composite materials.

[0057] Figure 3 The image shows the texture feature parameters of a real bird and a bionic bird at different distances; where: Figure 3 (a) shows the texture feature parameters of a real bird and a bionic bird at a distance of 5m. Figure 3 (b) shows the texture feature parameters of a real bird and a bionic bird at a distance of 10 m. Figure 3 (c) shows the texture feature parameters of a real bird and a bionic bird at a distance of 15 m.

[0058] Figure 4 Here is a confusion matrix diagram for the three algorithms; where: Figure 4 (a) is the confusion matrix diagram using the RF algorithm. Figure 4 (b) is the confusion matrix diagram using the SVM algorithm. Figure 4 (c) is the confusion matrix diagram using the MLP algorithm.

[0059] Figure 5 This is a technical roadmap of the flapping-wing target detection and identification method based on the combination of laser micro-motion and polarization-infrared features of the present invention. Detailed Implementation

[0060] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention. Specific implementation method one:

[0062] This embodiment provides a method for detecting and identifying flapping-wing targets based on the combination of laser micro-motion and polarization-infrared features. The implementation steps of the method are as follows:

[0063] Step 1: Select the recognition features of three modes: laser micro-motion, polarization, and infrared.

[0064] Step 2: Propose a multimodal feature fusion criterion and assign weights to the features;

[0065] Step 3: Construct a dataset and combine it with machine learning algorithms for target recognition.

[0066] Furthermore, in step one, the selection of recognition features for three modes—laser micro-motion, polarization, and infrared—is specifically as follows:

[0067] (1) Regarding the identification features of the laser micro-motion mode (e.g.) Figure 1 (as shown)

[0068] For a single-stage flapping-wing UAV, the selection of bi-peak consistency without wing sweeping frequency shift is used as the discrimination index for a single-stage flapping-wing UAV.

[0069] For two-stage flapping-wing UAVs, the absence of bi-peak consistency and no sweeping wing frequency shift is selected as the discrimination index for two-stage flapping-wing UAVs.

[0070] For birds, the non-double-peak consistency with swept-wing frequency shift is selected as the bird discrimination index;

[0071] (4) Regarding the identification features of the polarization mode (e.g.) Figure 2 (as shown)

[0072] Natural biological materials with a deviance value higher than 0.5 (such as pheasant feathers, guinea fowl feathers, silver pheasant feathers, goose feathers, hen feathers, or duck feathers) were selected as the criteria for distinguishing flying birds.

[0073] Artificial composite materials with a depolarization value of less than 0.5 (such as epoxy carbon fiber or peptide carbon fiber) were selected as the discrimination index for flapping-wing UAVs; the depolarization value of 0.5 was obtained from the analysis of experimental data.

[0074] (5) Regarding the identification features of the infrared mode (e.g.) Figure 3 (as shown)

[0075] The texture structure features of the target surface are selected as the discrimination index between birds and flapping-wing UAVs. The infrared image is preprocessed (to reduce the background influence and facilitate subsequent processing) and the target is extracted using a target locking and image cropping method based on brightness maximization.

[0076] Furthermore, the selection of the target surface texture structure features as the discrimination index between birds and flapping-wing UAVs, and the use of a target locking and image cropping method based on maximizing brightness to preprocess the infrared image and extract the target, specifically:

[0077] (3.1) First, read the experimental image data and set the target size, size tolerance, brightness threshold and center cropping expansion according to the user-defined target size, size tolerance, brightness threshold and center cropping expansion amount; then convert the image to grayscale, find all the brightest points, calculate the average of the row coordinates of these brightest points, and round it to the nearest integer as the row coordinates of the center point. Similarly, calculate the average of the column coordinates of all the brightest points and round it to the nearest integer as the column coordinates of the center point. Thus, the average center of the brightest point is obtained.

[0078] (3.2) Locate the center of the average brightest point, obtain the original image size, determine the cropping range and adjust the boundary based on the above center cropping expansion amount, and complete the image cropping centered on the center of the average brightest point;

[0079] (3.3) Target localization: First, adaptive histogram equalization is used to enhance contrast; then dynamic thresholding is performed: the top 10% of pixels with the highest image brightness are selected, and the threshold is set to the minimum value of these 10% of pixels, and the threshold is not less than 200, generating a binary mask; then morphological processing is performed, and noise is removed and the boundaries are smoothed by combining opening and closing operations; finally, region analysis and screening are performed on images whose size is within ±50% of the target size and whose aspect ratio is between 0.5 and 2; the target size is 120×120 pixels;

[0080] Morphological opening and closing operations are existing algorithms. Reference: Zhang Ying. Research on the application of opening and closing operations in eliminating image noise [J]. Journal of Weifang University, 2002(2).

[0081] (3.4) Finally, target extraction is performed. The candidate region closest to the center of the image is selected, and the bounding box is expanded (left 50px, top 40px, width and height each +100px) to protect the boundary and prevent it from going out of bounds. The final cropping is then performed.

[0082] Furthermore, the texture structure features of the target surface include contrast, inverse difference moment, energy, and entropy;

[0083] For birds, the mean contrast ratio ranges from 0.269 to 0.312, the mean inverse moment ranges from 0.931 to 0.946, the mean energy ranges from 0.398 to 0.538, and the mean entropy ranges from 1.683 to 2.225.

[0084] For flapping-wing UAVs, the average contrast ratio ranges from 0.027 to 0.068, the average inverse moment ranges from 0.969 to 0.989, the average energy ranges from 0.599 to 0.916, and the average entropy ranges from 0.391 to 1.122.

[0085] The infrared images obtained in the experiment were processed using a method based on the gray-level co-occurrence matrix (a current technology) to obtain the corresponding values ​​for the above features. It can be seen that there are significant differences between the four texture features of birds and flapping-wing UAVs.

[0086] The method based on the gray-level co-occurrence matrix is ​​cited in this paper: Liu Guangyu, Huang Yi, Cao Yu, et al. Research on image texture feature extraction based on gray-level co-occurrence matrix [J]. Science and Technology Wind, 2021, (12): 61-64.

[0087] Furthermore, the formula for calculating the deflection degree is as follows:

[0088] (1)

[0089] In equation (1), Indicates the degree of deflection. Indicates degree of polarization; polarization degree (2)

[0090] In equation (2), S0 represents the total incident light intensity, S1 represents the linearly polarized light intensity in the 0° and 90° directions, S2 represents the linearly polarized light intensity in the 45° and 135° directions, and S3 represents the circularly polarized light intensity in both left and right directions. Most of the circularly polarized components can be ignored within the instrument's detection range, so the circularly polarized components are taken as 0 in multi-beam imaging polarization detection.

[0091] Furthermore, (Contrast) (Inverse moment) (energy), The formulas for calculating (entropy) are as follows:

[0092] (3)

[0093] (4)

[0094] ; (5)

[0095] (6)

[0096] In equations (3)-(6), i is the gray value of the reference pixel or the starting pixel, and j is the gray value of the neighboring pixel paired with the reference pixel i under a specific spatial relationship. Represents pixel pairs in the grayscale matrix The probability of.

[0097] Furthermore, in step two, the multimodal feature fusion criterion (a target classification criterion based on a feature hierarchical fusion strategy) is proposed, and weights are assigned to the features; specifically:

[0098] By combining the features of the above modalities with machine learning, and adopting a multi-dimensional information fusion criterion with target micro-motion modal features as the main factor and polarization and infrared features as auxiliary factors, efficient and accurate target classification is achieved; and weight allocation is performed based on the physical salience of the features.

[0099] Furthermore, the weight allocation is based on the physical saliency of the features, combined with machine learning algorithm training. The weight allocation principle is as follows:

[0100] The total weight of the micro-motion feature is 0.70, of which the weight of the bimodal consistency is 0.55 and the weight of the swept-wing frequency shift is 0.15;

[0101] The polarization feature weight is 0.10;

[0102] The total weight of image texture features is 0.20, of which the weight of contrast is 0.15, the weight of inverse moment is 0.02, the weight of energy is 0.01, and the weight of entropy is 0.02.

[0103] The above data was obtained and allocated based on the following: based on its significant physical characteristics and combined with the training set and dataset, the weights of each feature corresponding to the best recognition accuracy were found by training with SVM, MLP, and RF algorithms.

[0104] Furthermore, in step three, the construction of the dataset and the combination of machine learning algorithms for target recognition specifically involves:

[0105] First, import the training dataset containing features of birds, single-stage drones, and two-stage drones. Randomize the data by shuffling the order, normalize the data, and set feature weights of 0.55, 0.15, 0.10, 0.15, 0.02, 0.01, and 0.02. Apply the weights to the normalized features, import the test set, process the test set using the same normalization parameters as the training set, and apply the same feature weights for preprocessing. Next, train three algorithm models: MLP, RF, and SVM (all well-known algorithms). Finally, use the trained models to predict the test set and calculate the accuracy using the following formula.

[0106] Correct predictions / total number of samples = classification accuracy;

[0107] Both the number of correct predictions and the total number of samples can be reflected in the program. The total number of samples is the total number of samples in the test set, and the number of correct predictions is obtained in the training process of the trained model predicting the test set.

[0108] Example 1:

[0109] like Figure 5 As shown in the figure, this embodiment discloses a multimodal feature fusion method for detecting and identifying flapping-wing targets based on laser micro-motion features and polarization-infrared features. The method includes the following steps:

[0110] First, the characteristics of the micro-motion, polarization, and infrared modes are selected, and the selection criteria are as follows:

[0111] (1) Micro-motion modal characteristics: The specific characteristics are shown in the time-frequency diagram. Figure 1 As shown, based on the results of the characteristic time-frequency maps, bimodal consistency and swept-wing frequency shift were selected as the discrimination indicators between birds and flapping-wing UAVs. Specifically, birds do not have bimodal consistency but have swept-wing frequency shift, while one-stage flapping-wing UAVs have bimodal consistency but no swept-wing frequency shift, and two-stage flapping-wing UAVs have neither bimodal consistency nor swept-wing frequency shift.

[0112] (2) Polarization mode characteristics: Depolarization results for different materials are as follows Figure 2As shown, a debiasing value of 0.5 was selected as the discrimination index between birds and flapping-wing UAVs based on the debiasing results. The specific difference is that the debiasing of natural biological materials (such as birds) is generally higher than 0.5, while the debiasing of artificial composite materials (such as flapping-wing UAVs) is generally lower than 0.5.

[0113] (3) Infrared modal features: Texture feature parameters of real birds and bionic birds at different distances, such as Figure 3 As shown, the texture structure features of the target surface, namely contrast, inverse moment, energy, and entropy, are selected as the discrimination indicators between birds and flapping-wing UAVs. Specifically, birds exhibit high contrast, low inverse moment, low energy, and high entropy, reflecting the complexity and dynamism of natural textures. In contrast, flapping-wing UAVs exhibit low contrast, high inverse moment, and low entropy, reflecting the uniformity and static nature of artificial materials.

[0114] Next, a target classification criterion based on a feature-layered fusion strategy is proposed. This criterion combines the features of the aforementioned modalities with machine learning, employing a multimodal information fusion method that prioritizes target micro-motion features and supplements them with polarization and infrared features. This achieves efficient and accurate target classification. The micro-motion layer implies the physical unforgeability of motion dynamics features, the polarization layer reflects the differences in material properties, and the infrared layer improves the accuracy of the discrimination.

[0115] Based on the physical saliency of features, weights are assigned and trained using machine learning algorithms. The optimal weight allocation is as follows: Micro-motion features have a total weight of 0.70, as they directly determine the primary classification and possess irreplaceable physical mechanisms; their sub-feature bimodal consistency has a weight of 0.55, as it strongly represents the periodic motion of mechanical structures and directly determines the classification of one-piece UAVs; the swept-wing frequency shift has a weight of 0.15, because it effectively separates birds from two-piece UAVs by quantifying the spectral differences between biological flexible motion and mechanical rigid body motion; polarization features have a weight of 0.10, as polarization, as an intrinsic material property, shows significant separability between biological feathers and polymer materials in laboratory tests; image texture features have a total weight of 0.20, with contrast having a weight of 0.15, describing the sharpness of target contours through differences in grayscale gradient distribution, clearly distinguishing birds from flapping-wing UAVs; inverse moment of difference has a weight of 0.02, quantifying differences in surface uniformity; energy has a weight of 0.01, reflecting the periodicity of texture; and entropy has a weight of 0.02, reflecting the statistical characteristics of random distribution.

[0116] Example 2:

[0117] like Figure 5 As shown in the figure, this embodiment provides a multimodal feature fusion method for flapping-wing target detection and recognition based on laser micro-motion features and polarization-infrared features. The specific implementation steps are as follows:

[0118] First, a hybrid dataset was constructed based on a combination of physical mechanism modeling and experimental data acquisition. The training set generated 1000 sets of numerical simulation samples for each of the three target classes (one-piece flapping-wing UAV, two-piece flapping-wing UAV, and birds) through numerical simulation (totaling 3000 sets), and randomized the samples to eliminate order bias. The test set contained 200 sets of samples for each class (totaling 600 sets), where micro-motion feature data were obtained through model simulation, and deflection and infrared texture features were derived from experimental data in a laboratory environment.

[0119] Secondly, weights are assigned based on the physical saliency of the features. Combined with machine learning training, the optimal weights are as follows: the total weight of micro-Doppler features is 0.70, of which the weight of bimodal consistency is 0.55 and the weight of swept-wing frequency shift is 0.15; the weight of polarization features is 0.10; the total weight of image texture features is 0.20, of which the weight of contrast is 0.15, the weight of inverse moment difference is 0.02, the weight of energy is 0.01, and the weight of entropy is 0.02.

[0120] Next, to compare the recognition accuracy under a single micro-Doppler mode, the micro-Doppler feature weights were first set to 1. The recognition accuracies of each algorithm under a single micro-Doppler mode are as follows: RF algorithm: 58.83%; SVM algorithm: 59.83%; MLP algorithm: 57.5%. Based on multi-modal feature fusion, the recognition accuracies of each algorithm are as follows: RF algorithm: 91.83%; SVM algorithm: 93.17%; MLP algorithm: 93.50%, showing a significant improvement in accuracy. Further combining the confusion matrix, the confusion matrix after three-dimensional feature fusion is as follows: Figure 4 As shown in the analysis, the confusion matrix reveals that most identification errors occur when two-segment drones are mistakenly identified as one-segment drones. The main reason for this is likely the overlap in features between the two types of targets. In the micro-Doppler mode, neither one-segment nor two-segment drones exhibit wing-sweeping frequency shifts, but both are made of man-made materials and are similar in polarization modes. MLP enhances feature fusion through deep nonlinear modeling, SVM strengthens the classification boundary using kernel functions, while RF suffers from decreased accuracy due to its random feature sampling, which ignores feature differences.

[0121] Bimodal consistency: This means that in the micro-Doppler time-frequency plot, the micro-Doppler frequency shift values ​​of the main peak and the secondary peak are consistent.

[0122] Sweep-wing frequency shift: In a micro-Doppler time-frequency diagram, the frequency shift value below the fundamental frequency corresponding to time 0 is represented.

[0123] Depolarization: A physical quantity representing the degree of change in polarization state of light during propagation or scattering. It measures the extent to which light transitions from a fully polarized state to an unpolarized state, and its value ranges from 0 to 1. Its relationship with the degree of polarization is as follows:

[0124] (1)

[0125] In equation (1), Indicates the degree of deflection. Indicates the degree of polarization.

[0126] polarization degree (2)

[0127] In equation (2), S0 represents the total incident light intensity, S1 represents the linearly polarized light intensity in the 0° and 90° directions, S2 represents the linearly polarized light intensity in the 45° and 135° directions, and S3 represents the circularly polarized light intensity in both left and right directions. Most of the circularly polarized components can be ignored within the instrument's detection range, so they are taken as 0 in multi-beam imaging polarization detection.

[0128] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for detecting and identifying flapping-wing targets based on the combination of laser micro-motion and polarization-infrared features, characterized in that: The implementation steps of the method are as follows: Step 1: Select the recognition features of three modes: laser micro-motion, polarization, and infrared; the specific process is as follows: (1) Regarding the identification characteristics of the laser micro-motion mode; For a single-stage flapping-wing UAV, the selection of bi-peak consistency without wing sweeping frequency shift is used as the discrimination index for a single-stage flapping-wing UAV. For two-stage flapping-wing UAVs, the absence of bi-peak consistency and no sweeping wing frequency shift is selected as the discrimination index for two-stage flapping-wing UAVs. For birds, the non-double-peak consistency with swept wing frequency shift is selected as the bird discrimination index; (2) Regarding the identification features of the polarization mode; Natural biological materials with a debiasing value higher than 0.5 were selected as the discrimination index for birds. Artificial composite materials with a deflection value of less than 0.5 were selected as the discrimination index for flapping-wing UAVs; (3) Regarding the identification features of the infrared mode; The texture structure features of the target surface are selected as the discrimination index between birds and flapping-wing UAVs. The infrared image is preprocessed and the target is extracted using a target locking and image cropping method based on brightness maximization. The formula for calculating the deflection degree is as follows: （1） In equation (1), Indicates the degree of deflection. Indicates degree of polarization; polarization degree (2) In equation (2), This represents the total intensity of the incident light. This represents the intensity of linearly polarized light in the 0° and 90° directions. This indicates the intensity of linearly polarized light at 45° and 135°. Indicates the intensity of circularly polarized light that is either left-handed or right-handed; Step 2: Propose a multimodal feature fusion criterion and assign weights to the features; specifically: By combining the features of the above modalities with machine learning, and adopting a multi-dimensional information fusion criterion with target micro-motion modal features as the main factor and polarization and infrared features as auxiliary factors, efficient and accurate target classification is achieved; and weight allocation is performed based on the physical saliency of the features. The weight allocation is based on the physical saliency of the features, combined with machine learning algorithm training. The weight allocation principle is as follows: The total weight of the micro-motion feature is 0.70, of which the weight of the bimodal consistency is 0.55 and the weight of the swept-wing frequency shift is 0.15; The polarization feature weight is 0.10; The total weight of image texture features is 0.20, of which the contrast weight is 0.15, the inverse moment weight is 0.02, the energy weight is 0.01, and the entropy weight is 0.

02. Step 3: Construct a dataset and combine it with machine learning algorithms for target recognition.

2. The flapping-wing target detection and identification method based on laser micro-motion and polarization-infrared feature combination as described in claim 1, characterized in that: The textural features of the target surface are selected as the discrimination index between birds and flapping-wing UAVs. A target locking and image cropping method based on maximizing brightness is used to preprocess the infrared image and extract the target. Specifically: (3.1) First, read the experimental image data and set the target size, size tolerance, brightness threshold and center cropping expansion according to the user-defined target size, size tolerance, brightness threshold and center cropping expansion amount; then convert the image to grayscale, find all the brightest points, calculate the average of the row coordinates of these brightest points, and round it to the nearest integer as the row coordinates of the center point. Similarly, calculate the average of the column coordinates of all the brightest points and round it to the nearest integer as the column coordinates of the center point. Thus, the average center of the brightest point is obtained. (3.2) Locate the center of the average brightest point, obtain the original image size, determine the cropping range and adjust the boundary based on the above center cropping expansion amount, and complete the image cropping centered on the center of the average brightest point; (3.3) Target localization: First, adaptive histogram equalization is used to enhance contrast; then dynamic thresholding is performed: the top 10% of pixels with the highest image brightness are selected, and the threshold is set to the minimum value of these 10% of pixels, and the threshold is not less than 200, generating a binary mask; then morphological processing is performed, and noise is removed and the boundaries are smoothed by combining opening and closing operations; finally, region analysis and screening are performed on images whose size is within ±50% of the target size and whose aspect ratio is between 0.5 and 2; the target size is 120×120 pixels; (3.4) Finally, target extraction is performed, the candidate region closest to the center of the image is selected, the bounding box is expanded, boundary protection is performed to prevent cross-boundary, and the final cropping is performed.

3. The flapping-wing target detection and identification method based on laser micro-motion and polarization-infrared feature combination as described in claim 1 or 2, characterized in that: The texture structure features of the target surface include contrast, inverse moment, energy, and entropy; For birds, the mean contrast ratio ranges from 0.269 to 0.312, the mean inverse moment ranges from 0.931 to 0.946, the mean energy ranges from 0.398 to 0.538, and the mean entropy ranges from 1.683 to 2.

225. For flapping-wing UAVs, the average contrast ratio ranges from 0.027 to 0.068, the average inverse moment ranges from 0.969 to 0.989, the average energy ranges from 0.599 to 0.916, and the average entropy ranges from 0.391 to 1.

122.

4. The flapping-wing target detection and identification method based on laser micro-motion and polarization-infrared feature combination as described in claim 3, characterized in that: , , , The calculation formulas are as follows: （3） （4） ； （5） （6） In equations (3) to (6), The grayscale value of the reference pixel, or starting pixel. Let be the grayscale value of the neighboring pixel paired with the reference pixel i under a certain spatial relationship. Represents pixel pairs in the grayscale matrix The probability of.

5. The flapping-wing target detection and identification method based on laser micro-motion and polarization-infrared feature combination according to claim 1, characterized in that: In step three, the construction of the dataset and the combination of machine learning algorithms for target recognition specifically involve: First, import the training dataset containing features of birds, single-stage drones, and two-stage drones. Randomize the data by shuffling the order, normalize the data, and set feature weights of 0.55, 0.15, 0.10, 0.15, 0.02, 0.01, and 0.

02. Apply the weights to the normalized features, import the test set, process the test set using the same normalization parameters as the training set, and apply the same feature weights for preprocessing. Next, train three algorithm models: MLP, RF, and SVM. Finally, use the trained models to predict the test set and calculate the accuracy using the following formula. Correct predictions / total number of samples = classification accuracy; Both the number of correct predictions and the total number of samples can be reflected in the program. The total number of samples is the total number of samples in the test set, and the number of correct predictions is obtained in the training process of the trained model predicting the test set.

Images