Target Detection Method and Device Based on Hyperspectral Data Cluster Analysis
By using hyperspectral data clustering analysis, the problem of target recognition accuracy under low signal-to-noise ratio conditions in navigation path planning was solved, achieving high-precision and low-cost target detection, which is suitable for unmanned platform navigation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NAT UNIV OF DEFENSE TECH
- Filing Date
- 2026-05-06
- Publication Date
- 2026-07-31
AI Technical Summary
In navigation path planning, traditional manual feature methods and deep learning methods are susceptible to noise interference under low signal-to-noise ratio conditions, which leads to a decrease in target recognition accuracy and affects the detection accuracy of navigation path planning.
A clustering analysis method based on hyperspectral data is adopted. The band dimension is compressed by noise-aware principal component analysis to generate pseudo-RGB images. Contrast-guided linear iterative clustering and peak clustering of effective neighborhood density are used to calculate saliency scores to distinguish between background and target.
It effectively suppresses noise interference, improves detection accuracy in navigation scenarios, reduces data processing volume, adapts to different environments, reduces deployment costs, and achieves high-accuracy target detection.
Smart Images
Figure CN122289665B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of target recognition technology, and in particular to a target detection method and apparatus based on hyperspectral data clustering analysis. Background Technology
[0002] In the field of navigation path planning, significant research achievements have been made in target recognition technology, such as traditional manual feature extraction methods and deep learning methods. Manual feature extraction methods have the advantages of easy deployment and strong interpretability in specific scenarios, while deep learning-based detection methods have achieved significant breakthroughs in detection accuracy and efficiency thanks to their data-driven feature learning capabilities. However, both methods are constrained by image noise and signal-to-noise ratio in practical navigation path planning applications. When the image signal-to-noise ratio is low and noise interference is strong, feature extraction deviations or omissions are prone to occur, which in turn affects the effectiveness of feature representation and limits the overall detection accuracy of target recognition in navigation path planning. Summary of the Invention
[0003] Therefore, it is necessary to provide a target detection method and device based on hyperspectral data clustering analysis to address the aforementioned technical problems.
[0004] A target detection method based on hyperspectral data clustering analysis, the method comprising the following steps: Acquire hyperspectral data in navigation scenarios, compress the band dimension of hyperspectral data through noise-aware principal component analysis, and generate pseudo-RGB images; The pseudo-RGB image is converted to the CIELAB color space, and superpixel segmentation is performed on the pseudo-RGB image using a contrast-guided linear iterative clustering method to generate multiple superpixel blocks; The average spectral vector of each superpixel block is extracted, and the cluster center of the superpixel block is determined by the peak clustering method of effective neighborhood density. The superpixel blocks that are not clustered are assigned to the nearest cluster center to complete the clustering. Based on the extended boundary connectivity score and spatial weight, the saliency score of each cluster is calculated. According to the score results, the clusters are divided into background and target, and the background is removed to output the target detection results.
[0005] In one embodiment, hyperspectral data is acquired, and the band dimension of the hyperspectral data is compressed through noise-aware principal component analysis to generate a pseudo-RGB image, including: Acquire hyperspectral data and calculate the local variance of each pixel in the hyperspectral data; Based on the ratio of the local variance of pixels to the median of the local variance of global pixels, a negative exponential form of noise weight coefficients is constructed, and a weighted covariance matrix is constructed based on the noise weight coefficients. The weighted covariance matrix is decomposed into eigenvalues and sorted in descending order. The first K principal components are selected to form a projection matrix. The hyperspectral data is compressed into a few-band feature data with a band dimension of K through the projection matrix. The value of K is such that the cumulative variance contribution rate of the first K principal components is not less than 95%. Calculate the pixel mean and variance of each band in the few-band feature data, and score the adaptability of each band for the red, green and blue channels based on the mean and variance. Select the band with the highest score for each channel as the candidate band for the corresponding channel to generate a pseudo-RGB image.
[0006] In one embodiment, the pixel mean and variance of each band in the few-band feature data are calculated, and the red, green, and blue channels of each band are scored based on the mean and variance. The band with the highest score in each channel is selected as the candidate band for the corresponding channel to generate a pseudo-RGB image. The method also includes: If there is overlap among the candidate bands, the band with the second highest score is selected as the candidate band.
[0007] In one embodiment, the compatibility score for each band across the red, green, and blue channels is calculated using the following formula: ; in, For the first k Each band is used as a score for the red band; For the first k Each band is used as a score for the green band; For the first k Each band is used as a score for the blue band; For the first k The average value of pixel values in each band; For the first k The variance of pixel values in each band; This is the average of the mean values of all selected bands.
[0008] In one embodiment, the pseudo-RGB image is converted to the CIELAB color space, and superpixel segmentation is performed on the pseudo-RGB image using a contrast-guided linear iterative clustering method to generate multiple superpixel blocks, including: The pseudo-RGB image is converted to the CIELAB color space, retaining a five-dimensional feature vector containing luminance, chromaticity, and spatial coordinates; Based on the preset number of superpixels, the cluster centers are uniformly initialized using a grid step size method; Based on the grid step size, the distance between the pixel and the cluster center is calculated using local contrast as a correction factor. The local contrast is obtained by calculating the gradient variance of the pixel's neighborhood. Based on the distance between pixels and cluster centers, the cluster centers are iteratively updated, and each pixel is assigned to the nearest cluster center until the iteration converges, completing the superpixel segmentation and obtaining multiple superpixel blocks.
[0009] In one embodiment, the distance between the cell and the cluster center is calculated according to the following formula: ; in, The distance between the cell and the cluster center; The color distance between the pixel and the cluster center; This refers to the color weighting coefficient; This refers to the contrast weighting coefficient. For local contrast; The spatial distance between a pixel and the cluster center; The grid step size is the cluster center.
[0010] In one embodiment, the average spectral vector of each superpixel block is extracted, and the cluster center of the superpixel block is determined by a peak clustering method based on effective neighborhood density, including: Extract the average spectral vector of each superpixel block, and include neighboring superpixel blocks whose spectral similarity weight exceeds a preset threshold in the improved local density calculation; Calculate the relative distance between superpixel blocks, and construct a normalized decision value by combining local density and relative distance; Sort all superpixel blocks in descending order of decision value, calculate the gradient of decision values between adjacent sorts, and select the superpixel block with the largest gradient as the cluster center.
[0011] In one embodiment, the improved local density is calculated according to the following formula: ; in, For superpixel blocks i Improved local density; For superpixel blocks i With superpixel blocks j Spectral similarity weights; For superpixel blocks i With superpixel blocks j Spatial distance; To cut off the distance; This is an indicator function.
[0012] In one embodiment, based on the extended boundary connectivity score and spatial weights, a saliency score is calculated for each cluster. Based on the score, the clusters are divided into background and target, and the background is removed to output the target detection result, including: Calculate the extended boundary connectivity score for each cluster: ; in, For clusters R Extended boundary connectivity score: Spatial weights; For clusters R The number of pixels overlapping with the image edges; For clusters R Total number of pixels; The significance score for each cluster is calculated based on the extended boundary connectivity score: ; in, The significance score for cluster R; To adjust the parameters, the significance score of each cluster is calculated based on the extended boundary connectivity score. The significance score ranges from [0,1]. The closer the value is to 0, the more likely the corresponding cluster is to be the background. The closer the value is to 1, the more likely the corresponding cluster is to be the target. Using a preset threshold as the cluster classification line, clusters are divided into background and target, and background is removed to output the target detection result.
[0013] A target detection device based on hyperspectral data clustering analysis, the device comprising: The compression and dimensionality reduction module is used to acquire hyperspectral data in navigation scenarios. It compresses the band dimension of the hyperspectral data through noise-aware principal component analysis and generates a pseudo-RGB image. The superpixel segmentation module is used to convert pseudo-RGB images to the CIELAB color space. It performs superpixel segmentation on the pseudo-RGB images using a contrast-guided linear iterative clustering method to generate multiple superpixel blocks. The clustering module is used to extract the average spectral vector of each superpixel block, determine the cluster center of the superpixel block through the peak clustering method of effective neighborhood density, and assign the superpixel blocks that are not clustered to the nearest cluster center to complete the clustering. The target detection module is used to calculate the saliency score of each cluster based on the extended boundary connectivity score and spatial weights. Based on the score results, the clusters are divided into background and target, and the background is removed to output the target detection results.
[0014] The aforementioned target detection method and apparatus based on hyperspectral data clustering analysis reduces the dimensionality of hyperspectral data by compressing the band dimension through noise-aware principal component analysis, thereby selecting the main band components of the hyperspectral data and reducing the amount of hyperspectral data processing. A contrast-guided linear iterative clustering method is used to perform superpixel segmentation on pseudo-RGB images, ensuring that the superpixel segmentation maintains boundary sensitivity and preserves the image structural features of targets in navigation scenarios. A peak clustering method based on effective neighborhood density is used to determine the cluster centers of superpixel blocks, effectively suppressing the contamination of the density field by noise points. Based on extended boundary connectivity scores and spatial weights, a saliency score is calculated for each cluster, thereby distinguishing between background and target based on the score and outputting the target detection result after background removal, ensuring detection accuracy in navigation scenarios and providing reliable support for path planning.
[0015] This invention possesses strong anti-interference capabilities. Through principal component analysis for noise perception dimensionality reduction and peak clustering based on effective neighborhood density, it suppresses strong noise interference in navigation scenarios throughout the entire process, achieving high-accuracy target detection. Furthermore, this invention eliminates the need for collecting large amounts of hyperspectral data for navigation scenario annotation and model training, allowing for rapid adaptation to different navigation environments and reducing deployment costs and time for unmanned platform navigation. By employing dimensionality reduction and superpixel segmentation, this invention significantly compresses the amount of hyperspectral data processing, improving computational efficiency and meeting the processing requirements of real-time navigation for unmanned platforms. This invention requires no auxiliary light source or human intervention, passively completing environmental perception and target detection in navigation scenarios, exhibiting strong concealment and a wide range of applications. Attached Figure Description
[0016] Figure 1 This is a flowchart illustrating a target detection method based on hyperspectral data clustering analysis in one embodiment; Figure 2 This is a structural block diagram of a target detection device based on hyperspectral data clustering analysis in one embodiment. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0018] The target detection method based on hyperspectral data clustering analysis provided in this application can be applied to autonomous navigation scenarios, including but not limited to civilian unmanned vehicles and drones.
[0019] In one embodiment, such as Figure 1 As shown, a target detection method based on hyperspectral data clustering analysis is provided, the method comprising the following steps: Step 201: Obtain hyperspectral data in the navigation scenario, compress the band dimension of the hyperspectral data through noise-aware principal component analysis, and generate a pseudo-RGB image.
[0020] Step 202: Convert the pseudo-RGB image to the CIELAB color space, and perform superpixel segmentation on the pseudo-RGB image using a contrast-guided linear iterative clustering method to generate multiple superpixel blocks.
[0021] Step 203: Extract the average spectral vector of each superpixel block, determine the cluster center of the superpixel block by the peak clustering method of effective neighborhood density, and assign the superpixel blocks that are not clustered to the nearest cluster center to complete the clustering.
[0022] Step 204: Based on the extended boundary connectivity score and spatial weight, calculate the saliency score of each cluster. According to the score results, the clusters are divided into background and target, and the background is removed to output the target detection results.
[0023] In the aforementioned target detection method based on hyperspectral data clustering analysis, dimensionality reduction is achieved by compressing the band dimension of the hyperspectral data through noise-aware principal component analysis, thereby selecting the main band components of the hyperspectral data and reducing the amount of hyperspectral data processing. A contrast-guided linear iterative clustering method is used to perform superpixel segmentation on the pseudo-RGB image, ensuring that the superpixel segmentation maintains boundary sensitivity and preserves the image structural features of the target in the navigation scene. A peak clustering method based on effective neighborhood density is used to determine the cluster centers of superpixel blocks, effectively suppressing the contamination of the density field by noise points. Based on the extended boundary connectivity score and spatial weights, a saliency score for each cluster is calculated, thereby distinguishing between the background and the target based on the score and outputting the target detection result after removing the background. This ensures the detection accuracy in the navigation scene and provides reliable support for path planning.
[0024] In one embodiment, hyperspectral data in a navigation scenario is acquired, and the band dimension of the hyperspectral data is compressed through noise-aware principal component analysis to generate a pseudo-RGB image, including: Acquire hyperspectral data in navigation scenarios and calculate the local variance of each pixel in the hyperspectral data; Based on the ratio of the local variance of pixels to the median of the local variance of global pixels, a negative exponential form of noise weight coefficients is constructed, and a weighted covariance matrix is constructed based on the noise weight coefficients. The weighted covariance matrix is decomposed into eigenvalues and sorted in descending order. The first K principal components are selected to form a projection matrix. The hyperspectral data is compressed into a few-band feature data with a band dimension of K through the projection matrix. The value of K is such that the cumulative variance contribution rate of the first K principal components is not less than 95%. Calculate the pixel mean and variance of each band in the few-band feature data, and score the adaptability of each band for the red, green and blue channels based on the mean and variance. Select the band with the highest score for each channel as the candidate band for the corresponding channel to generate a pseudo-RGB image.
[0025] It should be noted that hyperspectral images are represented as ,in, For spatial coordinates, The data is indexed by bands; the local variance of each pixel is calculated within a local sliding window. In hyperspectral data, the local variance of each pixel characterizes the pixel's noise level, with higher noise regions corresponding to higher local variances. The constructed negative exponential form of noise weight coefficients assigns high weights to low-noise regions and low weights to high-noise regions, thus constructing a weighted covariance matrix. This weighted allocation automatically suppresses the interference of high-noise regions on the dimensionality reduction process, enhancing the contribution of the effective signal. The data also includes the pixel mean and variance for each band in the few-band feature data, where the mean characterizes the overall brightness of the band, and the variance characterizes the richness and contrast of band information.
[0026] In this embodiment, the first K principal components are selected to form a projection matrix, and the hyperspectral data is compressed into fewer band feature data with a band dimension of K through the projection matrix, thereby completing the data dimensionality reduction.
[0027] In one embodiment, the pixel mean and variance of each band in the few-band feature data are calculated, and the red, green, and blue channels of each band are scored based on the mean and variance. The band with the highest score in each channel is selected as the candidate band for the corresponding channel to generate a pseudo-RGB image. The method also includes: If there is overlap among the candidate bands, the band with the second highest score is selected as the candidate band.
[0028] In one embodiment, the compatibility score for each band across the red, green, and blue channels is calculated using the following formula: ; in, For the first k Each band is used as a score for the red band; For the first k Each band is used as a score for the green band; For the first k Each band is used as a score for the blue band; For the first k The average value of pixel values in each band; For the first k The variance of pixel values in each band; This is the average of the mean values of all selected bands.
[0029] In one embodiment, the pseudo-RGB image is converted to the CIELAB color space, and superpixel segmentation is performed on the pseudo-RGB image using a contrast-guided linear iterative clustering method to generate multiple superpixel blocks, including: The pseudo-RGB image is converted to the CIELAB color space, retaining a five-dimensional feature vector containing luminance, chromaticity, and spatial coordinates; Based on the preset number of superpixels, the cluster centers are uniformly initialized using a grid step size method; Based on the grid step size, the distance between the pixel and the cluster center is calculated using local contrast as a correction factor. The local contrast is obtained by calculating the gradient variance of the pixel's neighborhood. Based on the distance between pixels and cluster centers, the cluster centers are iteratively updated, and each pixel is assigned to the nearest cluster center until the iteration converges, completing the superpixel segmentation and obtaining multiple superpixel blocks.
[0030] It should be noted that the five-dimensional feature vector is represented as follows: ,in, Let be the brightness of the i-th pseudo-RGB image; Let be the chromaticity of the i-th pseudo-RGB image; Let represent the spatial coordinates of the i-th pseudo-RGB image; a five-dimensional feature vector containing brightness, chromaticity, and spatial coordinates is retained, taking into account both the material spectral characteristics and spatial location characteristics of the target in the navigation scene. The grid step size is calculated based on the height and width of the navigation scene image, specifically according to the following formula: ; in, S For grid step size, P That is the expected number of superpixels. H , W These are the height and width of the image, respectively.
[0031] Local contrast The gradient variance of the 3×3 neighborhood of a cell is obtained by calculating: ; in, The gradient map for the luminance component is calculated using the following formula: ; in, The neighborhood is a 3×3 area centered on the current pixel; The mean value of the brightness gradient magnitude within a 3×3 neighborhood; The coordinates of any pixel within a 3×3 neighborhood centered on the current pixel; After converting a pseudo-RGB image to the CIELAB color space, the coordinates are... The brightness component at that location; for along x The first-order partial derivative of the brightness of the axis; for along y The first-order partial derivative of the brightness of the axis; coordinates The brightness gradient magnitude at that location; The variance of the brightness gradient magnitude within the neighborhood is used to characterize the local contrast of the current pixel.
[0032] In this embodiment, local contrast is introduced as a correction factor when calculating the distance between a pixel and the cluster center, thereby ensuring the sensitivity of superpixels to target boundaries in the navigation scene. Furthermore, the local contrast is obtained by calculating the gradient variance of the pixel's neighborhood, which can characterize the edge boundary features at that location. Through the correction of the contrast term, superpixel segmentation can accurately match the edges of obstacles and terrain changes in the navigation scene, avoiding target misjudgment and deviations in traversable area identification caused by blurred boundaries.
[0033] In one embodiment, the distance between the cell and the cluster center is calculated according to the following formula: ; in, The distance between the cell and the cluster center; The color distance between the pixel and the cluster center; This refers to the color weighting coefficient; This refers to the contrast weighting coefficient. For local contrast; The spatial distance between a pixel and the cluster center; The grid step size is the cluster center.
[0034] In one embodiment, the average spectral vector of each superpixel block is extracted, and the cluster center of the superpixel block is determined by a peak clustering method based on effective neighborhood density, including: Extract the average spectral vector of each superpixel block, and include neighboring superpixel blocks whose spectral similarity weight exceeds a preset threshold in the improved local density calculation; Calculate the relative distance between superpixel blocks, and construct a normalized decision value by combining local density and relative distance; Sort all superpixel blocks in descending order of decision value, calculate the gradient of decision values between adjacent sorts, and select the superpixel block with the largest gradient as the cluster center.
[0035] It should be noted that the relative distance between superpixel blocks is the minimum distance from a given superpixel block to all points with higher local density. A higher decision value indicates that the superpixel block possesses more characteristics of a cluster center. The constructed normalized decision value is: ; ; in, For normalized decision values, For superpixel blocks i Improved local density, The relative distance between superpixel blocks; The relative distance between superpixel blocks is defined as the minimum distance to all points with higher density. ; For the point with the highest density, its relative distance is defined as: .
[0036] Sort all superpixel blocks in descending order of decision value, calculate the gradient of decision values between adjacent sorting positions, and select the superpixel block with the largest gradient as the cluster center. Specifically, sort all superpixel blocks in descending order of normalized decision value, and calculate the gradient of decision values between adjacent sorting positions. : ; Find the location of the maximum gradient If the first p superpixel blocks have significantly higher decision values than other points, they are the true cluster centers. Starting from the (p+1)th point, the decision values drop sharply, indicating that these points do not possess the characteristics of cluster centers. Therefore, the first p points are selected as cluster centers.
[0037] In this embodiment, the average spectral vector of each superpixel block is extracted as the feature representation of the superpixel block, thereby completely preserving the hyperspectral material feature information after dimensionality reduction. By introducing a spectral similarity weight based on a Gaussian kernel, the spectral feature difference between two superpixel blocks is quantified (the higher the weight, the more similar the materials of the two superpixel blocks are), thereby distinguishing targets and terrains of different materials in the navigation scene.
[0038] In one embodiment, the improved local density is calculated according to the following formula: ; in, For superpixel blocks i Improved local density; For superpixel blocks i With superpixel blocks j Spectral similarity weights; For superpixel blocksi With superpixel blocks j Spatial distance; To cut off the distance; This is an indicator function; it is only valid if the value inside the parentheses is... If the value is 1, then the value is 0; otherwise, the value is 0.
[0039] In this embodiment, by using spectral similarity constraints, neighboring points with large material differences are filtered out, effectively suppressing the contamination of the density field by isolated noise points and clutter in the navigation scene, and ensuring the accuracy of local density calculation.
[0040] In one embodiment, based on the extended boundary connectivity score and spatial weights, a saliency score is calculated for each cluster. Based on the score, the clusters are divided into background and target, and the background is removed to output the target detection result, including: Calculate the extended boundary connectivity score for each cluster: ; in, For clusters R Extended boundary connectivity score: Spatial weights; For clusters R The number of pixels overlapping with the image edges; For clusters R Total number of pixels; The significance score for each cluster is calculated based on the extended boundary connectivity score: ; in, The significance score for cluster R; To adjust the parameters, the significance score of each cluster is calculated based on the extended boundary connectivity score. The significance score ranges from [0,1]. The closer the value is to 0, the more likely the corresponding cluster is to be the background. The closer the value is to 1, the more likely the corresponding cluster is to be the target. Using a preset threshold as the cluster classification line, clusters are divided into background and target, and background is removed to output the target detection result.
[0041] It should be noted that in navigation scenarios, passable background areas are usually connected to image edges, while obstacles and objects to be detected are mostly located within the image's interior. Spatial weights are constructed by combining the overall size of the navigation scene image and the size of the bounding rectangle of the clusters. Specifically, spatial weights are defined as: ; in The width of the original image. The height of the original image. Includes evaluating the minimum width of the clustered region. This represents the maximum height of the region containing the cluster to be evaluated. This is the sensitivity coefficient of EBC, used to control the influence of size differences on the weights. The extended boundary connectivity score of the cluster is calculated by the number of pixels overlapping between the cluster and the image edge, and the total number of pixels in the cluster.
[0042] In this embodiment, spatial weights are introduced to correct the extended boundary connectivity score of clusters, thereby controlling the impact of size differences on weights and adapting to navigation scenario requirements with different field of view and different imaging distances.
[0043] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.
[0044] In one embodiment, such as Figure 2 As shown, a target detection device based on hyperspectral data clustering analysis is provided, the device comprising: The compression and dimensionality reduction module 901 is used to acquire hyperspectral data in navigation scenarios, compress the band dimension of the hyperspectral data through noise-aware principal component analysis, and generate a pseudo-RGB image.
[0045] The superpixel segmentation module 902 is used to convert the pseudo-RGB image to the CIELAB color space and perform superpixel segmentation on the pseudo-RGB image using a contrast-guided linear iterative clustering method to generate multiple superpixel blocks.
[0046] Clustering module 903 is used to extract the average spectral vector of each superpixel block, determine the cluster center of the superpixel block through the peak clustering method of effective neighborhood density, and assign the superpixel blocks that are not clustered to the nearest cluster center to complete the clustering.
[0047] The target detection module 904 is used to calculate the saliency score of each cluster based on the extended boundary connectivity score and spatial weight, divide the cluster into background and target based on the score results, remove the background and output the target detection results.
[0048] Specific limitations regarding the target detection device based on hyperspectral data clustering analysis can be found in the limitations of the target detection method based on hyperspectral data clustering analysis described above, and will not be repeated here. Each module in the aforementioned target detection device based on hyperspectral data clustering analysis can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.
[0049] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0050] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.
Claims
1. A target detection method based on hyperspectral data clustering analysis, characterized in that, The method includes the following steps: Acquire hyperspectral data in navigation scenarios, compress the band dimension of the hyperspectral data through noise-aware principal component analysis, and generate pseudo-RGB images, including: Acquire hyperspectral data in navigation scenarios and calculate the local variance of each pixel in the hyperspectral data; Based on the ratio of the local variance of pixels to the median of the local variance of global pixels, a negative exponential form of noise weight coefficients is constructed, and a weighted covariance matrix is constructed based on the noise weight coefficients. The weighted covariance matrix is decomposed into eigenvalues and sorted in descending order. The first K principal components are selected to form a projection matrix. The hyperspectral data is compressed into a few-band feature data with a band dimension of K through the projection matrix. The value of K is such that the cumulative variance contribution rate of the first K principal components is not less than 95%. Calculate the pixel mean and variance of each band in the few-band feature data, and score the adaptability of each band for the red, green and blue channels based on the mean and variance. Select the band with the highest score for each channel as the candidate band corresponding to the channel to generate a pseudo-RGB image. The pseudo-RGB image is converted to the CIELAB color space, and superpixel segmentation is performed on the pseudo-RGB image using a contrast-guided linear iterative clustering method to generate multiple superpixel blocks; The average spectral vector of each superpixel block is extracted, and the cluster center of the superpixel block is determined by the peak clustering method of effective neighborhood density. The superpixel blocks that are not clustered are assigned to the nearest cluster center to complete the clustering. Based on the extended boundary connectivity score and spatial weights, a saliency score is calculated for each cluster. Based on the score, clusters are divided into background and target clusters. Background clusters are removed, and the target detection results are output, including: Calculate the extended boundary connectivity score for each cluster: in, For clusters R Extended boundary connection fraction; Spatial weights; For clusters R The number of pixels overlapping with the image edges; For clusters R Total number of pixels; The significance score for each cluster is calculated based on the extended boundary connectivity score: in, The significance score for cluster R; To adjust the parameters; the significance score ranges from [0,1]. The closer the value is to 0, the more likely the corresponding cluster is to be the background; the closer the value is to 1, the more likely the corresponding cluster is to be the target. Using a preset threshold as the cluster classification line, clusters are divided into background and target, and background is removed to output the target detection result.
2. The target detection method based on hyperspectral data clustering analysis according to claim 1, characterized in that, The algorithm calculates the pixel mean and variance of each band in the few-band feature data, and scores the adaptability of each band for the red, green, and blue channels based on the mean and variance. It then selects the band with the highest score for each channel as the candidate band for that channel, generating a pseudo-RGB image. The algorithm also includes: If there is overlap among the candidate bands, the band with the second highest score is selected as the candidate band.
3. The target detection method based on hyperspectral data clustering analysis according to claim 1, characterized in that, For each band, the compatibility score for the red, green, and blue channels is calculated using the following formula: in, For the first k Each band is used as a score for the red band; For the first k Each band is used as a score for the green band; For the first k Each band is used as a score for the blue band; For the first k The average value of pixel values in each band; For the first k The variance of pixel values in each band; This is the average of the mean values of all selected bands.
4. The target detection method based on hyperspectral data clustering analysis according to claim 1, characterized in that, The pseudo-RGB image is converted to the CIELAB color space, and superpixel segmentation is performed on the pseudo-RGB image using a contrast-guided linear iterative clustering method to generate multiple superpixel blocks, including: The pseudo-RGB image is converted to the CIELAB color space, retaining a five-dimensional feature vector containing luminance, chromaticity, and spatial coordinates; Based on the preset number of superpixels, the cluster centers are uniformly initialized using a grid step size method; Based on the grid step size, the distance between the pixel and the cluster center is calculated using local contrast as a correction factor. The local contrast is obtained by calculating the gradient variance of the pixel's neighborhood. Based on the distance between pixels and cluster centers, the cluster centers are iteratively updated, and each pixel is assigned to the nearest cluster center until the iteration converges, completing the superpixel segmentation and obtaining multiple superpixel blocks.
5. The target detection method based on hyperspectral data clustering analysis according to claim 4, characterized in that, The distance between the pixel and the cluster center is calculated using the following formula: in, The distance between the cell and the cluster center; The color distance between the pixel and the cluster center; This refers to the color weighting coefficient; This refers to the contrast weighting coefficient. For local contrast; The spatial distance between a pixel and the cluster center; The grid step size is the cluster center.
6. The target detection method based on hyperspectral data clustering analysis according to claim 4, characterized in that, The average spectral vector of each superpixel block is extracted, and the cluster centers of the superpixel blocks are determined by peak clustering of effective neighborhood density, including: Extract the average spectral vector of each superpixel block, and include neighboring superpixel blocks whose spectral similarity weight exceeds a preset threshold in the improved local density calculation; Calculate the relative distance between superpixel blocks, and construct a normalized decision value by combining local density and relative distance; Sort all superpixel blocks in descending order of decision value, calculate the gradient of decision values between adjacent sorts, and select the superpixel block with the largest gradient as the cluster center.
7. The target detection method based on hyperspectral data clustering analysis according to claim 6, characterized in that, The improved local density is calculated according to the following formula: in, For superpixel blocks i Improved local density; For superpixel blocks i With superpixel blocks j Spectral similarity weights; For superpixel blocks i With superpixel blocks j Spatial distance; To cut off the distance; This is an indicator function.
8. A target detection device based on hyperspectral data clustering analysis, characterized in that, The device includes: The compression and dimensionality reduction module is used to acquire hyperspectral data in navigation scenarios. It compresses the band dimension of the hyperspectral data through noise-aware principal component analysis to generate a pseudo-RGB image, including: Acquire hyperspectral data in navigation scenarios and calculate the local variance of each pixel in the hyperspectral data; Based on the ratio of the local variance of pixels to the median of the local variance of global pixels, a negative exponential form of noise weight coefficients is constructed, and a weighted covariance matrix is constructed based on the noise weight coefficients. The weighted covariance matrix is decomposed into eigenvalues and sorted in descending order. The first K principal components are selected to form a projection matrix. The hyperspectral data is compressed into a few-band feature data with a band dimension of K through the projection matrix. The value of K is such that the cumulative variance contribution rate of the first K principal components is not less than 95%. Calculate the pixel mean and variance of each band in the few-band feature data, and score the adaptability of each band for the red, green and blue channels based on the mean and variance. Select the band with the highest score for each channel as the candidate band corresponding to the channel to generate a pseudo-RGB image. The superpixel segmentation module is used to convert pseudo-RGB images to the CIELAB color space. It performs superpixel segmentation on the pseudo-RGB images using a contrast-guided linear iterative clustering method to generate multiple superpixel blocks. The clustering module is used to extract the average spectral vector of each superpixel block, determine the cluster center of the superpixel block through the peak clustering method of effective neighborhood density, and assign the superpixel blocks that are not clustered to the nearest cluster center to complete the clustering. The object detection module calculates the saliency score for each cluster based on the extended boundary connectivity score and spatial weights. Based on the score, it divides the clusters into background and target clusters, removes the background clusters, and outputs the object detection results, including: Calculate the extended boundary connectivity score for each cluster: in, For clusters R Extended boundary connection fraction; Spatial weights; For clusters R The number of pixels overlapping with the image edges; For clusters R Total number of pixels; The significance score for each cluster is calculated based on the extended boundary connectivity score: in, The significance score for cluster R; To adjust the parameters; the significance score ranges from [0,1]. The closer the value is to 0, the more likely the corresponding cluster is to be the background; the closer the value is to 1, the more likely the corresponding cluster is to be the target. Using a preset threshold as the cluster classification line, clusters are divided into background and target, and background is removed to output the target detection result.