Joint optimization unmanned aerial vehicle hyperspectral remote sensing camouflage target anomaly detection method

Abstract

The invention discloses a joint optimization unmanned aerial vehicle hyperspectral remote sensing camouflage target anomaly detection method. Comprising the following steps: preprocessing hyperspectral data, performing dimension reduction by adopting an extreme point and information entropy, acquiring part of abnormal targets by utilizing an RX anomaly detection method based on Cauchy diagram distance optimization, acquiring the abnormal targets again through a low-rank sparse representation model, fusing multi-metric information based on entropy-sparsity adaptive weighting, and reserving more abnormal targets; through three core innovations of spectral feature enhancement, multi-dimensional anomaly modeling and adaptive fusion, a full-process optimization system from data preprocessing to result output is constructed, the problems of low contrast, strong background interference, high-dimensional calculation bottleneck and the like in camouflage target detection are effectively solved while the advantages of hyperspectral data are maintained, and the detection efficiency is improved. The method has theoretical innovation and engineering practicability, and is suitable for high-precision camouflage target monitoring tasks in a complex environment.

Description

Technical Field

[0001] The present invention relates to the technical field of hyperspectral remote sensing image anomaly detection, and in particular to a jointly optimized unmanned aerial vehicle (UAV) hyperspectral remote sensing camouflaged target anomaly detection method, in particular for anomaly detection and identification of civilian and military camouflaged concealed targets in complex ground scenes. Background Art

[0002] In military confrontation scenarios, camouflage nets, a typical artificial concealed target, are difficult to distinguish from the background through techniques such as spectral simulation and texture fusion. Traditional optical detection methods make it difficult to accurately detect and identify camouflaged targets. Hyperspectral imaging, with its spatial-spectral integration, can capture subtle spectral differences between camouflage nets and the natural background through continuous narrow-band imaging, providing a unique advantage for detecting camouflaged nets as anomalies in complex environments. With breakthroughs in drone platform technology and the advancement of miniaturized and lightweight hyperspectral sensors, drone hyperspectral anomaly detection technology is becoming a crucial tool for detecting camouflaged targets.

[0003] Hyperspectral target detection methods are primarily categorized as target detection based on prior spectral information and anomaly detection that does not require prior information. Target detection involves locating the target in a hyperspectral image, distinguishing it from the background, and ultimately extracting the target, given known spectral characteristics (such as specific minerals, vegetation, or man-made objects). However, the spectral characteristics of camouflage nets are susceptible to environmental degradation, material aging, and other factors, leading to prior spectral mismatch. Furthermore, acquiring the spectral characteristics of enemy camouflage nets is often difficult in real battlefields. Anomaly detection involves identifying pixels that are distinguishable from global or local background pixels through pattern recognition or statistical methods, even without a clear target to be detected. Anomaly detection can be categorized into full-band, band-selective, and deep learning methods. Full-band anomaly detection methods aim to quickly and accurately identify anomalous targets using continuous hyperspectral cube data. Full-band anomaly detection methods include those based on statistical background characteristics, clustering, graph-based, and low-rank sparse representations. Current full-band anomaly detection methods primarily utilize the spectral information of hyperspectral images, rarely utilizing spatial information. These methods are susceptible to false alarms due to background interference in complex backgrounds (those containing numerous objects). Band-selective anomaly detection methods aim to reduce data dimensionality and improve detection efficiency by screening key bands in hyperspectral data (e.g., those with high information content and low redundancy), while also preserving the distinguishing features between anomalies and background. These methods include those based on optimal indexes, information content, band correlation, optimal linear prediction of subspace principal components, and band clustering. Current band selection methods still have some shortcomings. For example, some methods (such as those based on information content and band correlation) suffer from high inter-band correlation, resulting in relatively high computational complexity and slow computation speed. There is still room for improvement in terms of rapid band selection and classification accuracy. Deep learning anomaly detection methods utilize deep neural networks to automatically extract features from hyperspectral data, model the background distribution through unsupervised learning, and identify anomalous pixels that deviate from this distribution. Due to the scarcity of samples and the presence of a large number of redundant bands in hyperspectral images, deep learning anomaly detection methods have problems such as training difficulties, high computational costs, and limited model generalization capabilities.

[0004] Although the above methods do not require prior information, they generally have problems such as sensitivity to noise, reliance on background statistical assumptions, complex parameter tuning, and insufficient adaptability to complex scenarios. In complex terrain, they are easily affected by interference such as vegetation and shadows, resulting in false alarms. Therefore, the robustness of the model can be improved by combining spatial-spectral joint modeling and reducing unnecessary redundant bands. Summary of the Invention

[0005] In order to reduce redundant bands, make full use of hyperspectral spatial-spectral information, and improve the model's adaptability to complex backgrounds, the present invention designs a jointly optimized method for detecting camouflaged targets in unmanned aerial vehicle (UAV) hyperspectral remote sensing. First, in order to effectively reduce the redundant bands of hyperspectral images and reduce the amount of computation, a hybrid band selection algorithm is proposed. Combining the two strategies of spectral extreme point detection and information entropy screening, the most representative bands are screened through multi-dimensional information. Secondly, in order to make full use of the spectral and spatial information of the hyperspectral image, an RX (Reed-Xiaoli, RX) anomaly detection method based on Cauchy diagram distance optimization is proposed. This method uses the Cauchy diagram distance to quantify the degree of deviation between pixels and background in the spectral dimension, thereby reducing the impact of the target on the background, replacing the covariance matrix inversion to reduce the amount of computation. At the same time, the graph Laplacian matrix and morphological opening operation are introduced in the spatial dimension to enhance the target features and preliminarily obtain sparse targets. Then, the low-rank and sparse representation model (LRASR) is used to construct the background dictionary, the low-rank coefficient matrix of the background dictionary, and the sparse anomaly matrix to obtain abnormal targets. Finally, in order to make full use of the spatial spectrum information, the entropy-sparsity-based adaptive weighted fusion of multiple metric information is adopted to reduce false positives and retain more correct targets.

[0006] Specifically, the technical solution provided by the present invention is: a jointly optimized UAV hyperspectral remote sensing camouflaged target anomaly detection method comprises the following steps:

[0007] S1: Hyperspectral data preprocessing:

[0008] The hyperspectral remote sensing images of camouflaged targets are collected using UAV hyperspectral technology, and the hyperspectral remote sensing image data collected by the UAV are normalized and preprocessed to eliminate baseline drift;

[0009] S2: Dimensionality reduction using extreme points and information entropy:

[0010] Calculate the extreme points in the spectral curve of the hyperspectral data preprocessed in step S1, mark the bands with significant spectral changes in all pixels, and capture the characteristic absorption or reflection positions of the material;

[0011] On the basis of selecting the extreme point band, the information entropy is further used to screen the remaining bands to retain more band information. The first b bands are selected in descending order of entropy value to form complementary bands with the extreme point band. The bands are fused to obtain the recombined band B after dimensionality reduction, and the recombined new data is obtained.

[0012] S3: Obtain some abnormal targets using the RX anomaly detection method based on Cauchy graph distance optimization:

[0013] Calculate the local contrast map of the hyperspectral image, perform edge detection to obtain the edge map, and then weightedly fuse the local contrast map and the edge map to obtain the saliency map;

[0014] The Cauchy kernel function is used to construct a nonlinear similarity graph to suppress the influence of noise, and the graph Laplacian matrix is used to describe the local geometric structure of the data to enhance the consistency of the local background.

[0015] The Mahalanobis distance extended by the graph structure is used to measure the degree of deviation of pixels in the feature space and initially obtain partial abnormal information D1;

[0016] S4: Obtain abnormal targets again through low-rank sparse representation model:

[0017] The reorganized new hyperspectral data is decomposed into background and anomaly parts, and the low-rank sparse representation model is used to represent the background and anomaly as a low-rank matrix and a sparse matrix respectively. A background dictionary is constructed by a K-means clustering method, and then the adaptive penalty linearization alternating direction method is used to solve the low-rank sparse representation model to obtain the anomaly information D2.

[0018] S5: Adaptively weighted fusion of multiple metrics based on entropy-sparsity to retain more abnormal targets:

[0019] The entropy-sparsity adaptive weighted fusion of abnormal information D1 and D2 is used to obtain the abnormal target detection result D.

[0020] The advantages of the present invention compared with the prior art are:

[0021] (1) The present invention effectively eliminates spectral noise caused by factors such as UAV platform vibration and atmospheric interference through normalization and baseline drift correction, ensuring that subsequent analysis is based on stable and reliable hyperspectral data, laying the foundation for high-precision detection.

[0022] (2) The present invention focuses on the extreme point bands with significant absorption / reflection characteristics in the spectral curve, directly capturing the specific spectral response of the substance to avoid the omission of key bands by traditional dimensionality reduction methods; based on the extreme point bands, the high-information bands are further screened through information entropy, which not only highlights the spectral differences of the camouflaged targets, but also retains the overall structural information of the data, solving the "dimensionality disaster" problem of hyperspectral data while maximizing information utilization.

[0023] (3) The present invention constructs a nonlinear similarity graph based on the Cauchy kernel function, which can effectively suppress non-Gaussian noise and complex background interference in hyperspectral data, and combines the graph Laplacian matrix to characterize the local geometric structure and enhance the consistency representation of background pixels; by fusing local contrast, edge saliency and the Mahalanobis distance of graph structure expansion, the abnormal target that deviates from the background spectrum / spatial structure can be accurately captured during the first detection, which is particularly suitable for weak signal detection of camouflaged targets.

[0024] (4) This paper uses a low-rank sparse model to separate background and anomalies, combines it with K-means clustering to build a background dictionary, and adaptively fits the low-rank characteristics of complex backgrounds, showing strong robustness to large areas of similar backgrounds. Through the adaptive penalty algorithm, it can effectively extract sparsely distributed camouflaged targets, compensating for the RX algorithm's missed detection problem in low-contrast scenes.

[0025] (5) The present invention dynamically adjusts the weights of D1 and D2 through entropy and sparsity to achieve complementary advantages of the two detection methods; the fusion strategy avoids the limitations of a single method and improves the integrity and reliability of the detection results, which is particularly suitable for complex scenarios where the camouflaged target has "partial spectral matching and local structural abnormalities".

[0026] (6) The present invention forms a closed-loop collaboration through preprocessing, dimensionality reduction, multi-method detection, and fusion: dimensionality reduction reduces the computational burden for subsequent detection and improves feature purity; the two detection methods model anomalies from the perspectives of "local deviation" and "global sparsity" respectively, covering different types of camouflaged targets; adaptive fusion further balances the recall rate and accuracy, reduces the false alarm rate, and performs more robustly under multi-angle and multi-lighting conditions of UAV dynamic imaging.

[0027] Through three core innovations, namely spectral feature enhancement, multi-dimensional anomaly modeling, and adaptive fusion, this invention constructs a full-process optimization system from data preprocessing to result output. While maintaining the advantages of hyperspectral data, it effectively solves problems such as low contrast, strong background interference, and high-dimensional computing bottlenecks in camouflaged target detection. It has both theoretical innovation and engineering practicality, and is suitable for high-precision camouflaged target monitoring tasks in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flow chart of the present invention.

[0029] Figure 2 This is a flow chart of the band selection dimensionality reduction based on extreme information entropy constraints of the present invention.

[0030] Figure 3 This is a flow chart of obtaining abnormal information based on the saliency weighted graph of the present invention.

[0031] Figure 4This is a flow chart of the adaptive weighted fusion of multi-metric information to obtain abnormal targets in the present invention.

[0032] Figure 5 The pseudo-color images, true value images, and detection effect images of different methods in the present invention on the military green camouflage net dataset;

[0033] Among them, (a) is pseudo color; (b) is the true value map; (c) is RX; (d) is RPCA-RX; (e) is LRASR; (f) is LSMAD; (g) is LRSNCR; (h) is Proposed;

[0034] Figure 6 The pseudo-color images, true value images, and detection effect images of the three-color camouflage net dataset using different methods in the present invention are shown;

[0035] Among them, (a) is pseudo color; (b) is the true value map; (c) is RX; (d) is RPCA-RX; (e) is LRASR; (f) is LSMAD; (g) is LRSNCR; (h) is Proposed;

[0036] Figure 7 The ROC diagrams of different methods in the present invention on two data sets are shown below;

[0037] Among them, (a) is the military green camouflage net dataset; (b) is the three-color camouflage net dataset; DETAILED DESCRIPTION

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0039] Example:

[0040] This paper designs a joint optimization method for detecting abnormalities of UAV hyperspectral remote sensing camouflaged targets. The flow chart is as follows: Figure 1As shown in the figure, the hyperspectral image data is first preprocessed, such as by normalizing the data. The preprocessed data is then reconstructed using band extreme points and information entropy to reduce its dimensionality and generate new hyperspectral data. Local contrast maps and edge maps of the reconstructed new data are then calculated, and a saliency map is obtained using weighted fusion. The Cauchy kernel function and graph Laplacian matrix are then used to construct a nonlinear similarity map and smooth local features, respectively. Partial anomaly information is then obtained using the Mahalanobis distance. The reconstructed new data is then decomposed into low-rank background and sparse target components. A background dictionary is constructed using the K-means clustering method, and partial anomaly information is then obtained using the adaptive penalty linearization alternating direction method. Finally, the multi-metric information is fused to obtain the anomaly target.

[0041] S1: Use UAV hyperspectral technology to collect hyperspectral remote sensing images of camouflaged targets, and then perform preprocessing on the original images such as normalization and elimination of baseline drift.

[0042] S101: First, the collected hyperspectral data needs to be standardized using the maximum and minimum method. The formula is:

[0043]

[0044] Among them, X0 represents the original hyperspectral data collected, and X represents the standardized data.

[0045] S102: Due to the discrete nature of hyperspectral data, in order to eliminate baseline drift (such as illumination changes) and enhance local changes in spectral features (such as absorption peaks and reflection valleys), the spectral difference is used as a finite approximation of the derivative of the spectral curve. The first-order derivative calculation formula is:

[0046]

[0047] Where f(λb) is the spectral curve, λb is the central wavelength of the b-th band, b is the band index, R(λb) is the reflectivity in band b, and Δλ is the wavelength interval between adjacent bands.

[0048] S2: Construct a band selection method based on extreme information entropy constraints. The method flow chart is as follows: Figure 2 Band selection can significantly reduce the redundancy of hyperspectral data, shorten the calculation time of anomaly detection, and improve detection efficiency. The specific steps are as follows.

[0049] S201: Calculate the extreme points in the spectral curve of the preprocessed data, mark the bands with significant spectral changes in all pixels, and capture the locations of the material's characteristic absorption or reflection. The extreme points correspond to the key bands of the material's characteristic absorption or reflection. Selecting these bands can preserve important information in the hyperspectral data and reduce the data dimension. The extreme point detection conditions are:

[0050]

[0051] Among them, M(b) is the extreme point of the b-th band. The extreme bands are counted to screen out the bands that contribute significantly to the global detection and avoid local noise interference. The formula is:

[0052]

[0053] Among them, B1(n) is the extreme band number of the nth pixel selected by the extreme point, n is the pixel number index, and N is the total number of pixels.

[0054] S202: Based on the selection of extreme point bands, the remaining bands are further screened using information entropy to retain more band information. High entropy bands contain more information (such as complex backgrounds and multi-category mixed areas), which supplement the global features that may be missed by extreme points. The formula is:

[0055]

[0056] Where B2(Nb) is the number of bands filtered by information entropy, Nb is the number of pixels in the bth band, and K is the number of bins into which the continuous spectral reflectance values are divided during the spectral data discretization (binning). For example, if the reflectance range of a band is [0, 1], it is divided into 256 equal-width bins, then K = 256. k = 0, 1, ..., K is the index after the spectral data is discretized, and pk is the probability that the pixel value falls into the kth bin, which is calculated as:

[0057]

[0058] Where Nk is the number of pixels in the kth interval, and N is the total number of pixels.

[0059] Select the first b (specified value) bands in descending order of entropy value to form complementary bands with the extreme point bands, and fuse the bands to obtain the recombined bands after dimensionality reduction. The formula is:

[0060] B=B1(j)∪B2(N i ) (7)

[0061] Among them, B is the recombined band after band selection.

[0062] S3: Construct a method based on saliency weighted graph to obtain some abnormal information. The method flow chart is as follows Figure 3As shown. In practical applications, traditional RX anomaly detection algorithms are prone to problems such as insufficient number of samples in the background covariance matrix, which makes it impossible to accurately describe the background model, resulting in reduced detection performance. To suppress the impact of outliers on background statistics, this paper uses significance weighting, kernel methods, and graph structures to improve performance, enhance the robustness of background modeling, enhance adaptability to non-Gaussian backgrounds, introduce morphological opening operations to remove small noise, and enhance true anomaly targets. The specific steps are as follows.

[0063] S301: Perform saliency detection on the hyperspectral data to highlight abnormal areas. Assume that the input hyperspectral image is X∈XN×B, where N is the total number of pixels and B is the number of bands. First, perform a centering operation on the hyperspectral image to eliminate the mean shift of the spectral data. The formula is:

[0064]

[0065] in, is the centered data, μ is the mean of each band. To calculate the single-channel image, the formula is:

[0066]

[0067] Where SCI is a single-channel image obtained by adding all bands of the multi-band image X. k is the band index.

[0068] The local contrast reflects the intensity of spectral changes in the pixel neighborhood. The larger the standard deviation, the more likely the area is to contain abnormal targets. Define the window size w. For each pixel (i, j), the local contrast calculation formula is:

[0069] C(i,j)=std(region(i:i+w-1,j:j+w-1)) (10)

[0070] Where C is the local contrast map, region is the w×w neighborhood area centered on pixel (i, j), and std represents the calculated standard deviation.

[0071] S302: Use Laplace Gaussian (LoG) operator to perform edge detection,

[0072] E=edge(SCI,'log') (11)

[0073] Where E is the edge map, edge is the edge detection function, and 'log' indicates the use of the Log operator.

[0074] The saliency map is obtained by weighted fusion of the local contrast map and the edge map:

[0075] S=0.9C+0.1E (12)

[0076] Among them, S is the saliency map.

[0077] To avoid overfitting, the low-saliency region (background) is given a high weight to suppress its contribution, and the weight of the high-saliency region (anomaly) is reduced, and the saliency map is reconstructed by weighted data. Through saliency adaptive weighting, the influence of abnormal pixels on the background statistic estimation is reduced. The formula is:

[0078]

[0079] Where W is the weight matrix used to construct a statistical model dominated by local background. ε = 0.001 prevents division by zero. T is the transposed sign.

[0080] S303: Use the Cauchy kernel function to construct a nonlinear similarity graph, which is less sensitive to long-distance differences and enhances local background consistency. The expression for calculating the similarity between pixels using the Cauchy kernel function is:

[0081]

[0082] Where Aij is the similarity matrix, representing the similarity measure between pixel i and pixel j. i,j∈{1,2,...,N} are the pixel indices in the hyperspectral image, where N is the total number of pixels. XXi,XXj are the weighted feature vectors of pixels i and j, weighted by the saliency map, and are the mean statistics. ‖·‖ is the Euclidean distance, which quantifies the difference between the two vectors. a is the scale parameter, which is the mean of the weighted feature vectors of all pixels and is used to adaptively adjust the similarity decay rate.

[0083] S304: Use the graph Laplacian matrix to describe the local geometric structure of the data, suppress the influence of noise, and describe the local smoothness of pixels in the graph structure. The graph Laplacian matrix L is defined as:

[0084] L=G -1 / 2 (GA)G -1 / 2 (15)

[0085] Where G is the degree matrix, Gii = ∑jAij.

[0086] whereGisthedegreematrixandGii=∑jAij.

[0087] The Mahalanobis distance extended by the graph structure is used to measure the degree of deviation of pixels in the feature space, and the abnormal pixels are counted to obtain the abnormality score:

[0088]

[0089] Among them, R(n) represents the degree of deviation of pixel n in the graph structure. The larger the value, the higher the possibility of abnormality. is the centered n-th pixel spectrum vector.

[0090] S305: Using morphological opening operation on the abnormal score image to eliminate small area noise and retain the connected abnormal area. The opening operation formula is:

[0091]

[0092] Where R is the anomaly score image. I is the structuring element, a small image that defines the neighborhood for the morphological operation. θ represents the erosion operation. ⊕ represents the dilation operation. ○ represents the opening operation.

[0093] Therefore, the initial abnormal information D1 is:

[0094] D1=R○I (18)

[0095] S4: Build a matrix decomposition model based on low-rank sparse representation (LRASR) to obtain some abnormal information again. The specific steps are as follows:

[0096] S401: In hyperspectral images, there is usually a strong correlation between background pixels, that is, a background pixel can be represented by some other background pixels. This means that the image matrix X can be decomposed into a background part and an anomaly part, as follows:

[0097] X=HY+Z (19)

[0098] Where X∈RB×N represents the hyperspectral image, B represents the number of bands, and N represents the total number of pixels. H = [h1,h2,...,hM] represents the background dictionary, and M represents the number of subspaces. Y = [y1,y2,...,yN] represents the low-rank coefficient matrix of the background dictionary. Z = [z1,z2,...,zN] represents the sparse anomaly matrix.

[0099] The proportion of abnormal pixels in the overall image is small, which shows sparsity. The vast majority of pixels are background, which shows low rank. The spectral vectors of background pixels have high interdependence and can be linearly represented by their adjacent pixels. Formula (19) can be expressed equivalently as:

[0100] min Y,Z ‖Y‖ * +β‖Y‖1+λ‖Z‖ 2,1 ,

[0101] s,t X=HY+Z (20)

[0102] Where ‖·‖2,1 represents the matrix l2,1 norm, β represents the low-rank part balance parameter, and λ represents the sparse part balance parameter. To solve the LRASR model, we first need to solve the background dictionary and then approximate the matrix solution as a convex optimization problem.

[0103] S402: In the low-rank sparse representation-based anomaly detection framework, the background dictionary is constructed using the K-means clustering algorithm. This method treats each cluster center as a dictionary atom and optimizes the dictionary structure by minimizing the sum of the distances from the cluster samples to the center. Specifically, the algorithm divides the data into M subspaces, ensuring that the data within each subspace is highly homogeneous while there are significant differences between different subspaces, thereby covering all background object categories in the image. The loss function is:

[0104]

[0105] Among them, cj is the center of the jth cluster, that is, the mean of each cluster.

[0106] In hyperspectral target detection, the background dictionary construction method based on K-means clustering can be decomposed into the following core steps: (1) Subspace partitioning: The hyperspectral image pixels are clustered into subspaces using the K-means algorithm. Each subspace corresponds to a ground feature category (such as vegetation, water, soil, etc.), and its spectral characteristics have high intra-class similarity. (2) Atom screening criteria: In each subspace, the Mahalanobis distance (MahalanobisDistance) of all pixel spectral vectors to the cluster center is calculated. This distance is normalized by the covariance matrix and can effectively eliminate the influence of inter-band correlation. It is more statistically robust than the Euclidean distance. (3) Background dictionary construction: The first q pixels with the smallest Mahalanobis distance (i.e., the samples closest to the subspace center) are selected from each subspace as background atoms. Finally, the q atoms of M subspaces are fused to form a background dictionary H with a dimension of M×q. This dictionary covers all background feature patterns and provides a complete representation basis for subsequent sparse reconstruction and anomaly detection.

[0107] S403: After the background dictionary is obtained, the Linearized Alternating Direction Method with Adaptive Penalty (LADMAP) is used to solve the low-rank sparse representation model. This method introduces linearization technology based on the alternating direction multiplier method and further accelerates the convergence of the algorithm by adding an approximation term to the quadratic term of the linearization. By introducing the auxiliary variable V1, formula (20) is replaced by:

[0108] min Y,Z ‖Y‖ * +β‖V1‖1+λ‖Z‖2,1 ,

[0109] s,t X=HY+Z, Y=V1 (22)

[0110] The Lagrangian function is constructed as:

[0111]

[0112] Where F1 and F2 are Lagrange multipliers, and τ is the penalty term coefficient. The augmented Lagrangian function guarantees the global optimality of the iterative solution for each convex function and can be divided into subproblems with multiple variables for solution.

[0113] Then, fix the variables (V1, Z), optimize the variable Y, fix the variables (Y, Z), optimize the variable V1, fix the variables (Y, V1), optimize the variable Z, and iteratively optimize each variable. When the objective function reaches the optimal solution, the abnormal part is solved as follows:

[0114] D2(y i )=||[Z * ] :,i ||2,i=1,2,...,N (24)

[0115] Formula (24) represents the l2,1 norm of the sparse abnormal part. When it is greater than the set value (self-set), the pixel point is abnormal information.

[0116] S5: Construct an adaptive weighted fusion model based on entropy-sparseness ratio to fuse multiple metric information to obtain abnormal targets. The method flow chart is as follows: Figure 4 shown.

[0117] S501: Calculate the information entropy E of the abnormal information D1 detected by significance weighting, using the following formula:

[0118] E=-∑p k log p k (25)

[0119] Where pk is the probability of pixel value distribution in D1.

[0120] S502: Calculate the sparsity S of the abnormal information D2 detected by low-rank sparse representation. The formula is as follows:

[0121]

[0122] Where ||D2||0 represents the number of non-zero elements in D2, and N is the total number of pixels.

[0123] S503: Normalize E and S to the range [0,1], and directly calculate the fusion weight coefficient α of D1 based on the relative ratio of entropy value and sparsity:

[0124]

[0125] The two information acquisition methods are weighted and fused to obtain the anomaly detection target. The formula is as follows:

[0126] D=αD1+(1-α)D2 (28)

[0127] Where D is the anomaly detection target and α∈[0,1] is the D1 weight coefficient.

[0128] To verify the feasibility of the proposed method, simulation experiments were conducted on two camouflage hyperspectral datasets collected using drone hyperspectral remote sensing technology. The drone was a DJI M300RTK multi-rotor drone, and the hyperspectral sensor was an Anzhou Technology CubertX20P airborne hyperspectral imager. The X20P has a spectral range of 350-1000 nanometers, a spectral resolution of 4 nanometers, and a spatial resolution of 6.79 centimeters at a collection altitude of 200 meters. The datasets consisted of a mixed background of grass, shrubs, and bare soil, and the targets were military green camouflage netting and three-color camouflage netting.

[0129] Figure 5 and Figure 6 The following are the detection effect diagrams of different methods in the military green camouflage net dataset and the three-color camouflage net dataset, where "Proposed" represents the anomaly detection method proposed by the present invention. The brighter the pixels in the figure, the more target pixels there are. Figure 5 and Figure 6 It can be seen that the method proposed in the present invention has the best effect on target extraction. Figure 7 Figure 1 shows the ROC curves of different methods for the military green camouflage net dataset and the three-color camouflage net dataset, respectively. "Proposed" indicates the proposed anomaly detection method. The closer the curve is to the upper left corner, the higher the detection rate for the same false alarm rate (or the lower the false alarm rate for the same detection rate). The ROCA curves of the two datasets show that the proposed method is in the upper left corner, overshadowing the other curves, indicating that the proposed method performs better than other methods.

[0130] Table 1 shows the AUC values of different methods. The bold numbers are the best values in this row. The closer the AUC value is to 1, the better the detection performance of the method. As can be seen from Table 1, the proposed method has the highest AUC value in both datasets, indicating that the proposed method has superior detection performance compared to other methods.

[0131] Table 1. AUC values of different methods

[0132]

[0133] The present invention and its embodiments are described above. This description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs structures and embodiments similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.

Claims

1. A jointly optimized UAV hyperspectral remote sensing camouflaged target anomaly detection method, characterized by: The following steps are involved: S1: Hyperspectral data preprocessing: The hyperspectral remote sensing images of camouflaged targets are collected using UAV hyperspectral technology, and the hyperspectral remote sensing image data collected by the UAV are normalized and preprocessed to eliminate baseline drift; S2: Dimensionality reduction using extreme points and information entropy: Calculate the extreme points in the spectral curve of the hyperspectral data preprocessed in step S1, mark the bands with significant spectral changes in all pixels, and capture the characteristic absorption or reflection positions of the material; On the basis of selecting the extreme point band, the information entropy is further used to screen the remaining bands to retain more band information. The first b bands are selected in descending order of entropy value to form complementary bands with the extreme point band. The bands are fused to obtain the recombined band B after dimensionality reduction, and the recombined new data is obtained. S3: Obtain some abnormal targets using the RX anomaly detection method based on Cauchy graph distance optimization: Calculate the local contrast map of the hyperspectral image, perform edge detection to obtain the edge map, and then weightedly fuse the local contrast map and the edge map to obtain the saliency map; The Cauchy kernel function is used to construct a nonlinear similarity graph to suppress the influence of noise, and the graph Laplacian matrix is used to describe the local geometric structure of the data to enhance the consistency of the local background. The Mahalanobis distance extended by the graph structure is used to measure the degree of deviation of pixels in the feature space and initially obtain partial abnormal information D1; S4: Obtain abnormal targets again through low-rank sparse representation model: The reorganized new hyperspectral data is decomposed into background and anomaly parts, and the low-rank sparse representation model is used to represent the background and anomaly as a low-rank matrix and a sparse matrix respectively. A background dictionary is constructed by a K-means clustering method, and then the adaptive penalty linearization alternating direction method is used to solve the low-rank sparse representation model to obtain the anomaly information D2. S5: Adaptively weighted fusion of multiple metrics based on entropy-sparsity to retain more abnormal targets: The entropy-sparsity adaptive weighted fusion of abnormal information D1 and D2 is used to obtain the abnormal target detection result D.

2. The method for detecting anomaly of a camouflaged target using a joint optimization method of UAV hyperspectral remote sensing according to claim 1 is characterized in that: Step S1 specifically includes the following contents: S101: First, the collected hyperspectral data needs to be standardized using the maximum and minimum method. The formula is: Among them, X0 represents the original hyperspectral data collected, and X represents the standardized data; S102: Due to the discreteness of hyperspectral data, in order to eliminate baseline drift and enhance local changes in spectral features, the spectral difference is used as a finite approximation of the spectral curve derivative. The first-order derivative calculation formula is: Where f(λb) is the spectral curve, λb is the central wavelength of the b-th band, b is the band index, R(λb) is the reflectivity in band b, and Δλ is the wavelength interval between adjacent bands.

3. The method for detecting anomaly of a camouflaged target using joint optimization of UAV hyperspectral remote sensing according to claim 1 is characterized in that: Step S2 specifically includes the following contents: S201: The extreme point detection condition uses the following formula: Among them, M(b) is the extreme point of the b-th band. The extreme bands are counted to screen out the bands that contribute significantly to the global detection and avoid local noise interference. The formula is: Among them, B1(n) is the extreme band number of the nth pixel selected by the extreme point, n is the pixel number index, and N is the total number of pixels. S202: Use the following formula to supplement the global features that may be missed by extreme points: Among them, B2(Nb) is the number of bands screened by information entropy, Nb is the number of pixels in the b-th band, and K is the number of intervals into which the continuous spectral reflectance value is divided during the discretization of spectral data. For example, if the reflectance range of a band is [0,1], it is divided into 256 equal-width intervals, then K = 256, k = 0, 1, ..., K is the index after discretization of spectral data, and pk is the probability of a pixel value falling into the k-th interval, which is calculated as follows: Where Nk is the number of pixels in the kth interval, and N is the total number of pixels; Select the first b bands in descending order of entropy value to form complementary bands with the extreme point bands, and fuse the bands to obtain the recombined bands after dimensionality reduction. The formula is: B=B1(j)∪B2(N i ) (7) Among them, B is the recombined band after band selection.

4. The method for detecting anomaly of a camouflaged target using joint optimization of UAV hyperspectral remote sensing according to claim 1 is characterized in that: Step S3 specifically Includes the following: S301: Assume that the input hyperspectral image is X∈XN×B, where N is the total number of pixels and B is the number of bands; First, the hyperspectral image is centered to eliminate the mean shift of the spectral data. The formula is: in, is the centered data, μ is the mean of each band; the formula for calculating the single-channel image is: Where SCI is a single-channel image obtained by adding all bands of the multi-band image X; k is the band index; The local contrast reflects the intensity of spectral changes in the pixel neighborhood. The larger the standard deviation, the more likely the area is to contain abnormal targets. Define the window size w. For each pixel (i, j), the local contrast calculation formula is: C(i,j)=std(region(i:i+w-1,j:j+w-1)) (10) Where C is the local contrast map, region is the w×w neighborhood area centered on pixel (i, j), and std represents the calculated standard deviation; S302: Use the Laplace Gaussian operator to perform edge detection. The formula is: E=edge(SCI,'log') (11) Where E is the edge map, edge is the edge detection function, and 'log' indicates the use of the Log operator. The saliency map is obtained by weighted fusion of the local contrast map and the edge map: S=0.9C+0.1E (12) Among them, S is the saliency map; In order to avoid overfitting, low-significance regions are given high weights to suppress their contributions, and the weights of high-significance regions are reduced. The saliency map is reconstructed by weighted data. By adaptively weighting the saliency, the influence of abnormal pixels on the background statistic estimation is reduced. The formula is: Where W is the weight matrix, which constructs the statistical model dominated by local background; ε = 0.001 to prevent division by zero; T is the transposition sign; S303: The expression for calculating the similarity between pixels using the Cauchy kernel function is: Where Aij is the similarity matrix, representing the similarity measure between pixel i and pixel j, i,j∈{1,2,...,N} are the indices of the pixels in the hyperspectral image, where N is the total number of pixels. XXi,XXj are the weighted feature vectors of pixels i and j, which are weighted by the saliency map and are the mean statistics; ‖·‖ is the Euclidean distance, which is used to quantify the difference between the two vectors; a is the scale parameter, which is the mean of XX, that is, the mean of the weighted feature vectors of all pixels, and is used to adaptively adjust the similarity decay rate. S304: Use the graph Laplacian matrix to describe the local geometric structure of the data, suppress the influence of noise, and describe the local smoothness of pixels in the graph structure. The graph Laplacian matrix L is defined as: L=G -1 / 2 (G-A)G -1 / 2 (15) Where G is the degree matrix, Gii=∑j Aij; The Mahalanobis distance extended by the graph structure is used to measure the degree of deviation of pixels in the feature space, and the abnormal pixels are counted to obtain the abnormality score: Among them, R(n) represents the degree of deviation of pixel n in the graph structure. The larger the value, the higher the possibility of abnormality. is the centered n-th pixel spectrum vector; S305: Using morphological opening operation on the abnormal score image to eliminate small area noise and retain the connected abnormal area. The opening operation formula is: Where R is the anomaly score image; I is the structural element, which is a small image that defines the neighborhood of the morphological operation; represents the corrosion operation; ⊕ represents the expansion operation; ○ represents the opening operation; Therefore, the initial abnormal information D1 is:

5. The method for detecting anomaly of a camouflaged target using joint optimization of UAV hyperspectral remote sensing according to claim 1 is characterized in that: Step S4 specifically includes the following contents: S401: In hyperspectral images, there is usually a strong correlation between background pixels, that is, a background pixel can be represented by some other background pixels, which means that the image matrix X can be decomposed into a background part and an anomaly part, as follows: X=HY+Z (19) Where X∈RB×N represents the hyperspectral image, B represents the number of bands, and N represents the total number of pixels; H = [h1,h2,...,hM] represents the background dictionary, and M represents the number of subspaces; Y = [y1,y2,...,yN] represents the low-rank coefficient matrix of the background dictionary; Z = [z1,z2,...,zN] represents the sparse anomaly matrix; The proportion of abnormal pixels in the whole image is small, showing sparsity. The vast majority of pixels are background, showing low rank. The spectral vectors of background pixels have high interdependence and can be linearly represented by their adjacent pixels. Formula (19) can be equivalently expressed as: minutes Y,Z ‖Y‖ * +β‖Y‖1+λ‖Z‖ 2,1 , s,t X=HY+Z (20) Among them, ‖·‖2,1 represents the matrix l2,1 norm, β represents the low-rank part balance parameter, and λ represents the sparse part balance parameter. In order to solve the LRASR model, we first need to solve the background dictionary, and then approximate the matrix solution for the convex optimization problem. S402: In the anomaly detection framework based on low-rank sparse representation, the background dictionary is constructed using the K-means clustering algorithm. The algorithm divides the data into M subspaces, ensuring that the data in each subspace is highly homogeneous, while there are significant differences between different subspaces, thereby covering all background object categories in the image. The loss function is: Among them, cj is the center of the jth cluster, that is, the mean of each cluster; S403: After the background dictionary is obtained, the low-rank sparse representation model is solved by the linearized alternating direction method with adaptive penalty. An approximation term is added to the quadratic term of the linearization to further accelerate the convergence of the algorithm. An auxiliary variable V1 is introduced, and formula (20) is replaced by: minutes Y,Z "Y" * +β‖V1‖1+λ‖Z‖ 2,1 , s,t X=HY+Z, Y=V1 (22) The Lagrangian function is constructed as: Among them, F1, F2 are Lagrange multipliers, τ is the penalty term coefficient; the augmented Lagrangian function can ensure the global optimality of the solution after iteration for each convex function, and can be divided into sub-problems of multiple variables for solution; Then, fix the variables (V1, Z), optimize the variable Y, fix the variables (Y, Z), optimize the variable V1, fix the variables (Y, V1), optimize the variable Z, and iteratively optimize each variable; when the objective function reaches the optimal solution, the abnormal part is solved as follows: D2(y i )=||[Z * ] :,i ||2,i=1,2,...,N (24) Formula (24) represents the l2,1 norm of the sparse abnormal part. When it is greater than the set value (self-set), the pixel point is abnormal information.

6. The jointly optimized UAV hyperspectral remote sensing camouflaged target anomaly detection method according to claim 1 is characterized by: Step S5 specifically includes the following contents: S501: Calculate the information entropy E of the abnormal information D1 detected by significance weighting, using the following formula: E=-∑p k logp k (25) Where pk is the probability of pixel value distribution in D1. S502: Calculate the sparsity S of the abnormal information D2 detected by low-rank sparse representation. The formula is as follows: Where ||D2||0 represents the number of non-zero elements in D2, and N is the total number of pixels. S503: Normalize E and S to the range [0,1], and directly calculate the fusion weight coefficient α of D1 based on the relative ratio of entropy value and sparsity: The two information acquisition methods are weighted and fused to obtain the anomaly detection target. The formula is as follows: D=αD1+(1-α)D2 (28) Where D is the anomaly detection target and α∈[0,1] is the D1 weight coefficient.

Images