A hyperspectral ship detection method and system with a priori coupled iterative RX

Abstract

The present application relates to the technical field of hyperspectral remote sensing image target recognition, and provides a hyperspectral ship detection method and system based on prior coupling iteration RX, which comprises the following steps: obtaining a hyperspectral image to be detected and performing pretreatment, constructing an initial background statistical model, obtaining an initial RX anomaly response map as a current anomaly response map through RX anomaly detection; determining a candidate region ship weak geometric matching degree; constructing a local background difference matching degree; performing normalized weighted fusion on the ship weak geometric matching degree and the local background difference matching degree to obtain a ship prior confidence; calculating RX anomaly statistics based on background modeling weight to update a background mean vector and a background covariance matrix to obtain a ship prior coupling RX anomaly response map; and updating the candidate region iteration output ship detection result according to the ship prior coupling RX anomaly response map to improve the detection directionality and detection accuracy of a ship target in a low spatial resolution hyperspectral image and reduce false alarms caused by a complex sea surface background.

Description

Technical Field

[0001] This invention relates to the field of hyperspectral remote sensing image target recognition technology, and in particular to a hyperspectral ship detection method and system with prior coupled iterative RX. Background Technology

[0002] Hyperspectral remote sensing imaging technology can simultaneously acquire spatial and continuous spectral information of ground objects, and has been widely used in fields such as marine target monitoring, sea surface target identification, and military reconnaissance. Ships, as typical artificial targets on the sea surface, usually appear as small-scale, weakly textured, and easily disturbed by the sea surface background in low spatial resolution hyperspectral images. Therefore, how to achieve stable and reliable ship detection in complex marine backgrounds has always been an important research direction in the field of hyperspectral image processing.

[0003] The field of hyperspectral ship detection has mainly developed three technical approaches. The first is unsupervised anomaly detection methods based on statistical modeling, with the Reed-Xiaoli (RX) algorithm and its various improvements being the most representative. These methods establish a background statistical distribution model and calculate the statistical deviation between pixels and the background to detect anomalous targets. Because they do not require training samples and are suitable for detecting unknown targets, they are widely used in hyperspectral anomaly detection tasks. The second type of method is mainly based on spectral feature analysis, utilizing the spectral differences between ships and seawater, islands, reefs, or other backgrounds to achieve target enhancement and background suppression. The third type of method combines spatial information or local structural features, introducing neighborhood relationships, texture structures, or local consistency constraints to improve the robustness of target detection.

[0004] While the aforementioned methods have improved ship detection performance to some extent, they still have significant shortcomings. First, RX and its related improvements are essentially unsupervised statistical models, relying on accurate modeling of global or local background statistical distributions. However, in complex sea scenarios, strong structural backgrounds such as cloud edges, island / reef boundaries, coastlines, and waves can produce anomalous responses similar to ship targets, leading to numerous false alarms. Second, existing RX and its multi-layered or iterative improvements typically lack clear target orientation; each iteration still focuses on updating background statistical characteristics, making it difficult to gradually converge stably towards the ship target area. Third, existing methods often use spectral features, geometric features, and spatial structural features as independent post-processing constraints or parallel discrimination conditions, failing to achieve deep integration and synergistic optimization with the RX statistical modeling process, thus limiting the improvement in detection performance. Summary of the Invention

[0005] To address the problems of existing hyperspectral ship detection technologies being susceptible to interference from complex ocean backgrounds and generating numerous false alarms, and the lack of prior guidance for ship targets in existing iterative RX methods, this application proposes a hyperspectral ship detection method and system with prior-coupled iterative RX. By constructing a prior confidence level for the ship and using this prior confidence level to iteratively update the background statistical model and RX anomaly response, the ship target response is enhanced and complex background interference is suppressed, thereby improving the accuracy and stability of ship detection. The objective of this invention can be achieved through the following technical solutions: This application provides a hyperspectral ship detection method based on prior coupled iterative RX, comprising the following steps: Step S1: In response to the hyperspectral ship detection task, acquire the hyperspectral image to be detected and preprocess it, construct an initial background statistical model, and obtain the initial RX anomaly response map as the current anomaly response map through RX anomaly detection. Step S2: Determine candidate regions based on the current response map, and construct a weak geometric matching degree for ships based on the scale-constrained features and directional extension features of the candidate regions; Step S3: Determine the neighboring background regions of the candidate region, and construct the local background difference matching degree based on at least one of the spectral differences, abnormal response differences, and neighborhood background consistency between the candidate region and the neighboring background regions; Step S4: Normalize and weightedly fuse the weak geometric matching degree of the ship and the local background difference matching degree to obtain the ship's prior confidence and associate it with the corresponding pixel; Step S5: Determine the background modeling weights based on the ship's prior confidence, update the background mean vector and background covariance matrix based on the background modeling weights and calculate the RX anomaly statistic, modulate the RX anomaly statistic using the ship's prior confidence, and obtain the ship's prior coupled RX anomaly response map. Step S6: Update the candidate region based on the ship's prior coupled RX anomaly response map. If the adjacent iteration results do not meet the preset stability condition and do not reach the preset maximum number of iterations, update the ship's prior coupled RX anomaly response map to the current anomaly response map and repeat steps S2 to S6. If the conditions are met, output the ship detection result.

[0006] Further, step S1 includes, The hyperspectral image to be detected is preprocessed, including radiometric correction and noise suppression, so that the spectral vectors of each pixel in the hyperspectral image to be detected are at a uniform data scale, thus obtaining the preprocessed hyperspectral image. Based on the spectral vectors of each pixel in the preprocessed hyperspectral image, the initial background mean vector and initial background covariance matrix are calculated to obtain the initial background statistical model. Then, based on the initial background statistical model, the RX anomaly statistics of each pixel are calculated using the RX anomaly detection algorithm. The RX anomaly statistics for each pixel are as follows: ;in, Represents the spectral vector of a pixel. Represents the background mean vector. Represents the background covariance matrix. The matrix representing the inverse of the background covariance matrix. Represents a pixel The corresponding anomaly detection statistics; The corresponding RX anomaly statistics are mapped to the image space according to the spatial location of each pixel, and the initial RX anomaly response map is used as the current anomaly response map.

[0007] Further, step S2 includes, The candidate response threshold is determined based on the distribution of abnormal response values ​​of each pixel in the current abnormal response map. Pixels whose abnormal response values ​​reach the candidate response threshold are identified as candidate pixels. Connectivity regions are aggregated based on the spatial adjacency relationship between candidate pixels to obtain at least one candidate region. The number of pixels contained in each candidate region is counted. Based on the spatial resolution of the hyperspectral image and the preset actual size range of the ship, the preset pixel scale range corresponding to the ship is determined. Based on the matching degree between the number of pixels and the preset pixel scale range, the scale-limited features of each candidate region are obtained. The scale-limited features are then quantified to obtain the scale matching degree corresponding to each candidate region. The spatial coordinates of pixels within each candidate region are obtained. A pixel coordinate covariance matrix is ​​constructed based on these coordinates. Eigenvalue decomposition is performed on the pixel coordinate covariance matrix to obtain a first eigenvalue and a second eigenvalue, where the first eigenvalue is greater than the second eigenvalue. The directional extension feature of the candidate region is determined based on the first and second eigenvalues. The directional extension feature is... ;in, Candidate region To prevent small constants with a denominator of zero, The first eigenvalue, It is the second eigenvalue; The directional extension features are normalized to obtain the directional extension matching degree corresponding to each candidate region. The scale matching degree and the directional extension matching degree are weighted and fused according to the corresponding geometric fusion weight to obtain the ship weak geometric matching degree corresponding to each candidate region.

[0008] Further, step S3 includes, Based on the spatial location of each candidate region, the preset neighborhood range is expanded outward, and the pixels outside the candidate regions within the preset neighborhood range are determined as neighborhood background pixels, forming the corresponding neighborhood background region. The spectral vectors of each pixel in the candidate region and the neighboring background region are obtained respectively. The first average spectral vector is calculated based on the spectral vectors of each pixel in the candidate region, and the second average spectral vector is calculated based on the spectral vectors of each pixel in the neighboring background region. The spectral distance or spectral angle between the first average spectral vector and the second average spectral vector is calculated to obtain the spectral difference between the candidate region and the neighboring background region. The abnormal response values ​​of each pixel in the candidate region and the neighboring background region in the current abnormal response map are obtained respectively. The average abnormal response values ​​corresponding to the candidate region and the neighboring background region are calculated respectively, and the abnormal response difference is determined based on the difference between the two average abnormal response values. The abnormal response variance is calculated based on the abnormal response values ​​of each pixel in the neighborhood background area, or the spectral variance is calculated based on the spectral vectors of each pixel in the neighborhood background area, and the neighborhood background consistency is determined based on the abnormal response variance or the spectral variance. Normalization is performed on at least one of spectral differences, abnormal response differences, and neighborhood background consistency, and then weighted fusion is performed according to preset difference fusion weights to obtain the local background difference matching degree corresponding to each candidate region.

[0009] Further, step S4 includes, The ship weak geometric matching degree and local background difference matching degree corresponding to each candidate region are obtained. The ship weak geometric matching degree and local background difference matching degree are normalized respectively and mapped to a unified numerical range to obtain the normalized ship weak geometric matching degree and normalized local background difference matching degree. Based on preset first and second weighting coefficients, the normalized weak geometric matching degree of ships and the normalized local background difference matching degree are weighted and fused to obtain the prior confidence of ships for each candidate region. The prior confidence of ships is: , Candidate region Candidate region The corresponding prior confidence level of the ship. Candidate region The corresponding normalized weak geometric matching degree of the ships, This corresponds to the normalized local background difference matching degree. , These are the corresponding weight coefficients; The prior confidence of the ship corresponding to the candidate region is associated with each pixel in the candidate region to determine the prior confidence of the ship for the corresponding pixel.

[0010] Further, step S5 includes, Obtain the prior confidence of the ship corresponding to each pixel in the current iteration round, take the confidence complement of each ship prior confidence relative to the single-location confidence, and determine the confidence complement as the background modeling weight of the corresponding pixel; The spectral vectors of each pixel are weighted and summed using their corresponding background modeling weights. The sum of these weighted sums is then normalized to obtain the prior weighted background mean vector for the current iteration. The prior weighted background mean vector is: ,in, This is the prior weighted background mean vector. For the first The spectral vector of each pixel, For the pixel indexes involved in background modeling, The first iteration in the previous iteration The background modeling weights corresponding to each pixel This refers to the current iteration round; Calculate the deviation vector between the spectral vector of each pixel and the prior weighted background mean vector. Multiply each deviation vector by its corresponding transpose to obtain the corresponding deviation outer product matrix. Then, perform a weighted summation of these deviation outer product matrices using the corresponding background modeling weights. Finally, normalize the weighted summation using the sum of all background modeling weights to obtain the prior weighted background covariance matrix. The prior weighted background covariance matrix is ​​as follows: ,in, This represents the prior weighted background covariance matrix. The first iteration in the previous iteration The background modeling weights corresponding to each pixel Indicates the first The spectral vector of each pixel, This is the prior weighted background mean vector for the current iteration round. This indicates the cell index that participates in background modeling, and T represents the transpose operation; Using the prior weighted background mean vector and the prior weighted background covariance matrix, the RX anomaly statistic for each pixel is calculated. A prior modulation term is constructed using the prior ship confidence and prior modulation coefficients for each pixel. Multiplying the prior modulation term by the corresponding RX anomaly statistic yields the ship prior coupled RX anomaly statistic. The ship prior coupled RX anomaly statistic is... ;in, This represents the a priori coupling RX anomaly statistic for ships. For the spectral vector of a pixel, This is the prior weighted background mean vector. The prior weighted background covariance matrix, Let T be the inverse of the prior weighted background covariance matrix, and let T denote the transpose operation. For pixels The corresponding prior confidence level of the ship, where λ represents the prior modulation coefficient; Based on the spatial location of each pixel, the corresponding ship prior coupled RX anomaly statistics are mapped to the image space to obtain the ship prior coupled RX anomaly response map.

[0011] Further, step S6 includes, Obtain the ship prior coupled RX anomaly response map generated in the current iteration round, and count the anomaly response value of each pixel. Calculate the mean and standard deviation of the anomaly response values, and combine the mean and standard deviation according to the threshold adjustment coefficient to obtain the candidate response threshold corresponding to the current iteration round. Cells whose abnormal response values ​​reach the candidate response threshold are identified as update candidate cells, and connected regions are aggregated according to the spatial adjacency relationship between each update candidate cell to obtain the update candidate region corresponding to the current iteration round. The ship prior coupled RX anomaly response map and update candidate region corresponding to the current iteration are compared with the anomaly response map and candidate region corresponding to the previous iteration to obtain the iterative change results corresponding to the adjacent iterations. Based on the iterative change results, it is determined whether the adjacent iteration results have reached the preset stability condition. When the results of adjacent iterations do not reach the preset stability condition and the current iteration number does not reach the preset maximum iteration number, the ship prior coupled RX abnormal response map generated in the current iteration round is used as the current abnormal response map of the next iteration round, and steps S2 to S6 are repeated. When the results of adjacent iterations reach the preset stability condition, or when the current iteration number reaches the preset maximum iteration number, the iteration is terminated, and the ship detection result is output according to the updated candidate region corresponding to the current iteration round.

[0012] Furthermore, the preset stability conditions include at least one of the following: the change in abnormal response between the ship prior coupled RX abnormal response maps corresponding to adjacent iteration rounds is not greater than a preset response change threshold; the regional overlap between candidate regions corresponding to adjacent iteration rounds is not less than a preset overlap threshold; and the change in confidence between the ship prior confidence corresponding to adjacent iteration rounds is not greater than a preset confidence change threshold.

[0013] Based on the same inventive concept, this application also provides a hyperspectral ship detection system using a priori coupled iterative RX, employing the hyperspectral ship detection method described above, including: The response module is used to respond to the hyperspectral ship detection task, acquire the hyperspectral image to be detected and preprocess it, build an initial background statistical model, and obtain the initial RX anomaly response map as the current anomaly response map through RX anomaly detection. The matching module is used to determine candidate regions based on the current response map, construct a weak geometric matching degree for the ship based on the scale-constrained features and directional extension features of the candidate regions, determine the neighboring background regions of the candidate regions, and construct a local background difference matching degree based on at least one of the spectral differences, anomalous response differences, and neighborhood background consistency between the candidate regions and the neighboring background regions, and normalize and weightedly fuse the ship's weak geometric matching degree and the local background difference matching degree to obtain the ship's prior confidence and associate it with the corresponding pixel. The prior coupling detection module is used to determine the background modeling weights based on the ship's prior confidence, update the background mean vector and background covariance matrix based on the background modeling weights and calculate the RX anomaly statistic, and modulate the RX anomaly statistic using the ship's prior confidence to obtain the ship's prior coupled RX anomaly response map. The result output module is used to update the candidate region based on the ship's prior coupled RX anomaly response map. When the adjacent iteration results do not reach the preset stability condition and do not reach the preset maximum number of iterations, the ship's prior coupled RX anomaly response map is updated to the current anomaly response map, and the matching module, prior coupling detection module and iterative update module are controlled to repeatedly execute the corresponding processing; if the conditions are met, the ship detection result is output.

[0014] Furthermore, the prior coupling detection module includes, The prior background modeling weight determination unit is used to obtain the prior confidence of the ship corresponding to each pixel in the current iteration round, take the confidence complement of each ship's prior confidence relative to the single-position confidence, and determine the confidence complement as the background modeling weight of the corresponding pixel; The prior weighted background mean calculation unit is used to perform weighted summation using the spectral vector of each pixel and the corresponding background modeling weight, and to normalize the weighted summation result using the sum of the background modeling weights to obtain the prior weighted background mean vector of the current iteration. The prior weighted background covariance calculation unit calculates the deviation vector between the spectral vector of each pixel and the prior weighted background mean vector. Each deviation vector is multiplied by its corresponding transpose to obtain the corresponding deviation outer product matrix. The deviation outer product matrices are then weighted and summed using the corresponding background modeling weights. Finally, the sum of all background modeling weights is used to normalize the weighted sum, yielding the prior weighted background covariance matrix. The prior coupled RX calculation unit is used to calculate the RX anomaly statistics corresponding to each pixel using the prior weighted background mean vector and the prior weighted background covariance matrix, and to construct a prior modulation term using the ship prior confidence and prior modulation coefficient corresponding to each pixel. The prior modulation term is multiplied with the corresponding RX anomaly statistics to obtain the ship prior coupled RX anomaly statistics. The prior response map generation unit is used to map the corresponding ship prior coupled RX anomaly statistics to the image space according to the spatial location of each pixel, so as to obtain the ship prior coupled RX anomaly response map.

[0015] Compared with the prior art, the present invention has at least one of the following technical advantages: This application improves the detection directionality and accuracy of ship targets in low spatial resolution hyperspectral images by iteratively coupling ship prior information with RX anomaly detection, and reduces false alarms caused by complex sea backgrounds. By constructing a weak geometric matching degree for ships and a candidate region-neighborhood background local background difference matching degree, and fusing the two to form a ship prior confidence, the ship target's scale-limited, directional extension, local prominence, and neighborhood background differences in low spatial resolution hyperspectral images can participate in RX background statistical modeling and anomaly statistics modulation. This transforms ship prior information from traditional post-processing screening conditions into coupled constraint variables in the RX detection process. Simultaneously, by using a prior weighted background statistical model and ship prior coupled RX anomaly statistics, the influence of high-confidence ship candidate regions on the estimation of background mean vector and background covariance matrix is ​​reduced, enhancing the ship target directionality in the RX detection process. Combined with iterative feedback and convergence cutoff mechanisms, the detection process gradually focuses on the ship target region, reducing false alarms caused by complex backgrounds such as cloud edges, islands, reefs, and coastlines. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below: Figure 1 This is a flowchart illustrating the steps of the hyperspectral ship detection method using prior coupled iterative RX in an embodiment of the present invention. Figure 2 This is a diagram illustrating the process of constructing the prior confidence level of a ship according to an embodiment of the present invention. Figure 3 This is a diagram illustrating the process of generating a priori coupled RX abnormal response of a ship according to an embodiment of the present invention. Figure 4 This is a diagram illustrating the iterative feedback and convergence judgment process for the candidate ship region in an embodiment of the present invention. Figure 5 These are examples of the original and result images for hyperspectral image ship detection according to an embodiment of the present invention. Figure 5 Image (a) is the original image of the second band of the hyperspectral spectrum. Figure 5 (b) in the image represents the initial RX detection result. Figure 5 (c) in the figure represents the RX detection result of the first iteration. Figure 5 (d) in the figure represents the RX detection result of the second iteration. Figure 5 (e) in the figure represents the RX detection result of the third iteration. Figure 5(f) in the figure represents the RX detection result of the fourth iteration. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. It should be noted that the specific embodiments described herein are merely some embodiments of the present application, not all embodiments, and are only used to explain the present application and are not intended to limit the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application.

[0018] First Embodiment In low spatial resolution hyperspectral remote sensing images, ship targets typically occupy only a small number of pixels. Their complete outlines, textures, and local structures are easily weakened due to spatial resolution limitations. However, certain spectral differences, anomalous response differences, and scale and directional extension features can still be preserved between the ship and the surrounding seawater background. Therefore, for this type of hyperspectral ship detection scenario, it is necessary to utilize the spectral dimension information of the hyperspectral data to identify the ship's anomaly relative to the background, while also making full use of the weak geometric information and local spatial difference information that can still be preserved under low spatial resolution conditions to reduce the interference of complex sea surface background on the ship anomaly detection results.

[0019] In existing hyperspectral ship detection methods, the RX anomaly detection method typically calculates the background mean vector and background covariance matrix based on the pixel spectral vectors in the hyperspectral image to be detected, and uses the Mahalanobis distance of the pixel relative to the background statistical model to determine the corresponding anomalous response. This type of method does not require pre-acquiring ship training samples and can detect unknown anomalous targets, thus it is suitable for preliminary search of maritime targets. However, the judgment of RX detection mainly comes from the degree of deviation of the pixel from the background statistical distribution, and it does not distinguish whether the object generating the anomalous response is a ship or other high-contrast background areas such as cloud edges, island / reef boundaries, coastlines, and ocean waves. When complex backgrounds also deviate from the main seawater background in spectral space, they are more likely to obtain higher RX anomaly responses, forming anomalous candidate regions that compete with ship targets.

[0020] Existing iterative RX or background update methods typically filter out some anomalous pixels based on the results of the previous round of anomaly detection, or gradually improve the background model by re-estimating background statistical parameters. While these methods can reduce the contamination of background statistical results by anomalous targets to some extent, their iterative process still mainly revolves around updating "which pixels are more suitable as background samples," lacking prior descriptions of the ship target itself. Especially under low spatial resolution conditions, ship targets and backgrounds such as island edges and cloud edges may all have high anomalous responses. Relying solely on the magnitude of the anomalous response or background statistical deviation cannot further determine whether a candidate region is more consistent with the characteristics of a ship target. Therefore, the iterative process is prone to convergence to general anomalous regions, rather than ensuring continuous focus on the true ship target region.

[0021] Furthermore, while existing technologies can post-process RX detection results using spectral differences, target size, length and width features, or local neighborhood relationships, this type of processing typically occurs after the RX anomaly response has been calculated, primarily serving to further filter existing candidate targets. In other words, the ship-related spectral, geometric, and local spatial difference information does not directly participate in the calculation of the background mean vector, background covariance matrix, and RX anomaly statistics. Therefore, even if some false alarm regions can be removed during post-processing, it cannot change the fact that the front-end RX background statistical model has already been affected by ship candidate regions or complex anomaly backgrounds, and it is difficult to proactively adjust the anomaly response distribution in subsequent iterations using prior ship information.

[0022] Based on the above problems, this embodiment considers that although it is difficult to obtain a complete and reliable ship outline under low spatial resolution conditions, the candidate ship region still has certain scale-limited features and directional extension features. At the same time, real ships usually exhibit certain spectral differences and anomalous response differences relative to the surrounding seawater background. Therefore, this embodiment does not use the above information only as a screening condition after RX detection is completed. Instead, it first constructs a weak geometric matching degree for the ship based on the scale-limited features and directional extension features of the candidate region, and then constructs a local background difference matching degree by combining the spectral differences, anomalous response differences, and neighborhood background consistency between the candidate region and its neighboring background region. The two are then fused to form the ship's prior confidence. The ship's prior confidence is further utilized in the RX detection process: on the one hand, the background modeling weights of each pixel in the calculation of the background mean vector and background covariance matrix are adjusted based on the ship's prior confidence to reduce the influence of pixels with a high ship probability on the background statistical model; on the other hand, the ship's prior confidence is used to modulate the RX anomaly statistics to enhance the anomalous response that conforms to the ship's prior characteristics. The specific implementation method is as follows: like Figure 1As shown, this application provides a hyperspectral ship detection method with prior coupled iterative RX. Without relying on additional labeled samples, it achieves a synergistic optimization mechanism for stable ship target prominence and background suppression in low spatial resolution hyperspectral images, thereby improving the reliability and robustness of ship detection in complex marine environments. The steps include: Step S1: In response to the hyperspectral ship detection task, acquire the hyperspectral image to be detected and preprocess it, construct an initial background statistical model, calculate the initial anomaly detection statistics of each pixel through RX anomaly detection, and obtain the initial RX anomaly response map as the current anomaly response map. Step S2: Determine candidate regions based on the current response map, and construct a weak geometric matching degree for ships based on the scale-constrained features and directional extension features of the candidate regions; Step S3: Determine the neighboring background region of the candidate region. Based on at least one of the spectral differences, abnormal response differences, and neighborhood background consistency between the candidate region and the neighboring background region, determine the local background difference matching degree between the candidate region and the neighboring background region. Step S4: Normalize and weightedly fuse the weak geometric matching degree of the ship and the local background difference matching degree to obtain the ship's prior confidence and associate it with the corresponding pixel; Step S5: Determine the background modeling weight by weighting the background modeling samples according to the ship's prior confidence. Update the background mean vector and background covariance matrix based on the background modeling weight and calculate the RX anomaly statistic. Modulate the RX anomaly statistic using the ship's prior confidence to obtain the ship's prior coupled RX anomaly response map. Step S6: Update the candidate region based on the ship's prior coupled RX anomaly response map. If the adjacent iteration results do not meet the preset stability condition and do not reach the preset maximum number of iterations, update the ship's prior coupled RX anomaly response map to the current anomaly response map and repeat steps S2 to S6. If the conditions are met, output the ship detection result.

[0023] Further, step S1 includes, The hyperspectral image to be detected is preprocessed, including radiometric correction and noise suppression, so that the spectral vectors of each pixel in the hyperspectral image to be detected are at a uniform data scale, thus obtaining the preprocessed hyperspectral image. Based on the spectral vectors of each pixel in the preprocessed hyperspectral image, the initial background mean vector and initial background covariance matrix are calculated to obtain the initial background statistical model. Then, based on the initial background statistical model, the statistical deviation of the spectral vector of each pixel relative to the initial background statistical model is calculated using the RX anomaly detection algorithm. The statistical deviation is determined as the RX anomaly statistic for the corresponding pixel. The RX anomaly statistics for each pixel are as follows: ;in, Represents the spectral vector of a pixel. Represents the background mean vector. Represents the background covariance matrix. The matrix representing the inverse of the background covariance matrix. Represents a pixel The corresponding anomaly detection statistics; The corresponding RX anomaly statistics are mapped to the image space according to the spatial location of each pixel, and the initial RX anomaly response map is used as the current anomaly response map.

[0024] Further, step S2 includes the following: the weak geometric matching degree of the ship is not determined based on the complete ship outline corresponding to the candidate region, but based on the weak geometric information retained by the candidate region in the low spatial resolution hyperspectral image. The weak geometric information includes at least the number of pixels, scale range, and principal axis extension degree of the candidate region. The ship's weak geometric matching degree is determined based on the scale matching degree and the directional extension matching degree. The scale matching degree is determined based on the matching relationship between the number of pixels in the candidate region and the preset pixel number range corresponding to the ship target. The preset pixel number range is determined based on the spatial resolution of the hyperspectral image to be detected and the preset actual size range of the ship. The directional extension matching degree is determined based on the principal axis extension degree of the pixel coordinate distribution in the candidate region.

[0025] The candidate response threshold is determined based on the distribution of abnormal response values ​​of each pixel in the current abnormal response map. Pixels whose abnormal response values ​​reach the candidate response threshold are identified as candidate pixels. Connectivity regions are aggregated based on the spatial adjacency relationship between candidate pixels to obtain at least one candidate region. The number of pixels contained in each candidate region is counted. Based on the spatial resolution of the hyperspectral image and the preset actual size range of the ship, a preset pixel scale range corresponding to the ship is determined. The scale-constrained features of each candidate region are obtained based on the degree of matching between the number of pixels and the preset pixel scale range. These scale-constrained features indicate whether the number of pixels contained in the candidate region meets the scale constraints corresponding to the ship target. The scale-constrained features are quantized to obtain the scale matching degree corresponding to each candidate region. Specifically, when the number of pixels contained in a candidate region is within the preset pixel count range, the candidate region is determined to meet the ship scale constraints. When the number of pixels contained in a candidate region is less than the lower limit of the preset pixel count range or greater than the upper limit of the preset pixel count range, the candidate region is determined not to meet the ship scale constraints.

[0026] Specifically, the candidate regions extracted from the initial RX anomaly response map or the prior coupled RX anomaly response map of the current layer ship are: Let the number of pixels contained in the candidate region be... The ship size range is The scale matching degree can be expressed as: ; in, Indicates candidate region Scale matching degree, Candidate region Candidate region Number of pixels included. and This refers to the minimum and maximum number of pixels corresponding to the ship target, determined based on the spatial resolution of the hyperspectral image to be detected and a preset range of actual ship sizes. For example, the spatial resolution of the hyperspectral image is... meters per pixel, the actual length of the vessel ranges from [missing information]. Width range is The actual size range of the ship is converted to a pixel scale range. , ; The spatial coordinates of pixels within each candidate region are obtained. A pixel coordinate covariance matrix is ​​constructed based on the spatial coordinates. Eigenvalue decomposition is performed on the pixel coordinate covariance matrix to obtain the first eigenvalue and the second eigenvalue, where the first eigenvalue is greater than the second eigenvalue. The directional extension features of the candidate region are determined based on the first eigenvalue and the second eigenvalue.

[0027] Specifically, ships appear as weakly extended structures distributed along a certain direction in hyperspectral images, and candidate regions... The set of internal cell coordinates is: ,in, For the candidate region The spatial coordinates of a pixel This indicates the number of pixels within the candidate region.

[0028] Based on coordinate set Construct the pixel coordinate covariance matrix and calculate the first eigenvalue. Second eigenvalue ,in ; The degree of directional extension of the candidate region is, ;in, Candidate region To prevent small constants with a denominator of zero, The larger the value, the more directional the expansion characteristic of the candidate region. The first eigenvalue, The second feature value is used; the directional extension feature is normalized to obtain the directional extension matching degree corresponding to each candidate region. The scale matching degree and the directional extension matching degree are weighted and fused according to the corresponding geometric fusion weight to obtain the weak geometric matching degree of the ship corresponding to each candidate region.

[0029] The directional extension matching degree is, ,in, To extend the matching degree in a directional manner, To pre-determine the upper limit of normalization of the extension degree, R) represents directional extension characteristics; The ship's weak geometric matching degree is ,in, Indicates candidate region The ship's weak geometric matching degree, Indicates the degree of scale matching. Indicates the degree of directional extension matching. Let be the weighting coefficient, satisfying .

[0030] Further, step S3 includes, Based on the spatial location of each candidate region, the preset neighborhood range is expanded outward, and the pixels outside the candidate regions within the preset neighborhood range are determined as neighborhood background pixels, forming the corresponding neighborhood background region. The spectral vectors of each pixel in the candidate region and the neighboring background region are obtained respectively. The first average spectral vector is calculated based on the spectral vectors of each pixel in the candidate region, and the second average spectral vector is calculated based on the spectral vectors of each pixel in the neighboring background region. The spectral distance or spectral angle between the first average spectral vector and the second average spectral vector is calculated to obtain the spectral difference between the candidate region and the neighboring background region. The abnormal response values ​​of each pixel in the candidate region and the neighboring background region in the current abnormal response map are obtained respectively. The average abnormal response values ​​corresponding to the candidate region and the neighboring background region are calculated respectively, and the abnormal response difference is determined based on the difference between the two average abnormal response values. The abnormal response variance is calculated based on the abnormal response values ​​of each pixel in the neighborhood background area, or the spectral variance is calculated based on the spectral vectors of each pixel in the neighborhood background area, and the neighborhood background consistency is determined based on the abnormal response variance or the spectral variance. Normalization is performed on at least one of spectral differences, abnormal response differences, and neighborhood background consistency, and then weighted fusion is performed according to preset difference fusion weights to obtain the local background difference matching degree corresponding to each candidate region.

[0031] Specifically, candidate regions are set. The first average spectral vector is Its outer neighboring background area The second average spectral vector is The spectral angle between the two can be expressed as, , Used to represent the spectral morphological differences between candidate regions and neighboring background regions.

[0032] The response differences between the candidate region and its neighborhood background are as follows: ,in, This represents the mean of abnormal responses within the candidate region. This represents the mean of abnormal responses within the outer neighborhood background region. This represents the standard deviation of the abnormal response within the outer neighborhood background region. To prevent constants with a denominator of zero, this index is used to characterize the degree of unusual prominence of a candidate region relative to its surrounding background region.

[0033] Neighborhood background consistency is determined based on the variance of anomalous responses or spectral variance within the outer neighborhood background region. Candidate vessel region. The corresponding neighborhood background consistency term is calculated using the following formula: ,in, Indicates the candidate ship area The outer neighborhood background region, Indicates the background area of ​​the outer neighborhood. The set of abnormal response values ​​corresponding to each pixel within the set. Indicates the background area of ​​the outer neighborhood. The standard deviation of the internal anomaly response value is used to characterize the degree of fluctuation of the anomaly response within the neighborhood background region.

[0034] Based on the aforementioned spectral differences, anomalous response differences, and neighborhood background consistency, a candidate region-neighborhood background local background difference matching degree can be constructed. After normalizing the above three components, a weighted fusion is performed: , This indicates the matching degree of local background differences between the candidate region and its neighboring background. This represents the normalized spectral difference term. This represents the difference in abnormal responses after normalization. This represents the normalized neighborhood background consistency term. This represents the weighting coefficient.

[0035] Furthermore, such as Figure 2 As shown, step S4 includes, The ship weak geometric matching degree and local background difference matching degree corresponding to each candidate region are obtained. The ship weak geometric matching degree and local background difference matching degree are normalized respectively and mapped to a unified numerical range to obtain the normalized ship weak geometric matching degree and normalized local background difference matching degree. Based on preset first and second weighting coefficients, the normalized weak geometric matching degree of ships and the normalized local background difference matching degree are weighted and fused to obtain the prior confidence of ships for each candidate region. The prior confidence of ships is: , Candidate region Candidate region The corresponding prior confidence level of the ship. Candidate region The corresponding normalized weak geometric matching degree of the ships, This corresponds to the normalized local background difference matching degree. , These are the corresponding weight coefficients, satisfying... ; The prior confidence of the ship corresponding to the candidate region is associated with each pixel in the candidate region to determine the prior confidence of the ship for the corresponding pixel.

[0036] Furthermore, such as Figure 3 , 4 As shown, step S5 includes, When the prior confidence of a ship corresponding to a pixel or candidate region is high, its background modeling weight in the background statistical model estimation is reduced; when the prior confidence of a ship corresponding to a pixel or candidate region is low, its background modeling weight in the background statistical model estimation is increased. This reduces the impact of high-confidence ship candidate regions on the estimation of the background mean vector and background covariance matrix.

[0037] Obtain the prior confidence of the ship corresponding to each pixel in the current iteration round, take the confidence complement of each ship prior confidence relative to the single-location confidence, and determine the confidence complement as the background modeling weight of the corresponding pixel; The spectral vectors of each pixel are weighted and summed using their corresponding background modeling weights. The sum of these weighted sums is then normalized to obtain the prior weighted background mean vector for the current iteration. The prior weighted background mean vector is: ,in, This is the prior weighted background mean vector. For the first The spectral vector of each pixel, For the pixel indexes involved in background modeling, The first iteration in the previous iteration The background modeling weights corresponding to each pixel This refers to the current iteration round; Calculate the deviation vector between the spectral vector of each pixel and the prior weighted background mean vector. Multiply each deviation vector by its corresponding transpose to obtain the corresponding deviation outer product matrix. Then, perform a weighted summation of these deviation outer product matrices using the corresponding background modeling weights. Finally, normalize the weighted summation using the sum of all background modeling weights to obtain the prior weighted background covariance matrix. The prior weighted background covariance matrix is ​​as follows: ,in, This represents the prior weighted background covariance matrix. The previous iteration round, i.e., the th The first in the layer The background modeling weights corresponding to each pixel Indicates the first The spectral vector of each pixel, This is the prior weighted background mean vector for the current iteration round. This indicates the cell index that participates in background modeling, and T represents the transpose operation; Using the prior weighted background mean vector and the prior weighted background covariance matrix, the RX anomaly statistic for each pixel is calculated. A prior modulation term is constructed using the prior ship confidence and prior modulation coefficients for each pixel. Multiplying the prior modulation term by the corresponding RX anomaly statistic yields the ship prior coupled RX anomaly statistic. The ship prior coupled RX anomaly statistic is... ;in, This represents the a priori coupling RX anomaly statistic for ships. For the spectral vector of a pixel, This is the prior weighted background mean vector. The prior weighted background covariance matrix, Let T be the inverse of the prior weighted background covariance matrix, and let T denote the transpose operation. For pixels The corresponding prior confidence level of the ship, where λ represents the prior modulation coefficient; Based on the spatial location of each pixel, the corresponding ship prior coupled RX anomaly statistics are mapped to the image space to obtain the ship prior coupled RX anomaly response map.

[0038] In the above formula, Modulation is used to measure the anomalous statistics of pixels relative to a prior weighted background statistical model; This is used to modulate the anomaly statistics based on the prior confidence of the ship. When a candidate region has both a high RX anomaly response and a high prior confidence of the ship, its response is amplified in the current layer; when a candidate region has a high RX anomaly response but does not meet the weak geometric characteristics of the ship or the candidate region-neighborhood background difference characteristics, its response will not be continuously amplified.

[0039] Further, step S6 includes, Obtain the ship prior coupled RX anomaly response map generated in the current iteration round, and count the anomaly response value of each pixel. Calculate the mean and standard deviation of the anomaly response values, and combine the mean and standard deviation according to the threshold adjustment coefficient to obtain the candidate response threshold corresponding to the current iteration round. Cells whose abnormal response values ​​reach the candidate response threshold are identified as update candidate cells, and connected regions are aggregated according to the spatial adjacency relationship between each update candidate cell to obtain the update candidate region corresponding to the current iteration round. The ship prior coupled RX anomaly response map and update candidate region corresponding to the current iteration are compared with the anomaly response map and candidate region corresponding to the previous iteration to obtain the iterative change results corresponding to the adjacent iterations. Based on the iterative change results, it is determined whether the adjacent iteration results have reached the preset stability condition. When the results of adjacent iterations do not reach the preset stability condition and the current iteration number does not reach the preset maximum iteration number, the ship prior coupled RX abnormal response map generated in the current iteration round is used as the current abnormal response map of the next iteration round, and steps S2 to S6 are repeated. When the results of adjacent iterations reach the preset stability condition, or when the current iteration number reaches the preset maximum iteration number, the iteration is terminated, and the ship detection result is output according to the updated candidate region corresponding to the current iteration round.

[0040] Furthermore, the preset stability conditions include at least one of the following: the change in abnormal response between the ship prior coupled RX abnormal response maps corresponding to adjacent iteration rounds is not greater than a preset response change threshold; the regional overlap between candidate regions corresponding to adjacent iteration rounds is not less than a preset overlap threshold; and the change in confidence between the ship prior confidence corresponding to adjacent iteration rounds is not greater than a preset confidence change threshold.

[0041] In one embodiment, the criterion for determining the change in anomaly response maps between adjacent layers is: Let the first The a priori coupling RX anomaly response diagram of the layered ship is as follows: , No. The a priori coupling RX anomaly response diagram of the layered ship is as follows: .in, Indicates the location of a pixel in an image. This represents the set of all pixels in the hyperspectral image to be detected. This indicates the total number of pixels.

[0042] The change in response plots between two adjacent layers can be defined as: When the following conditions are met: At that time, the abnormal response graph is judged to tend to stabilize. Among them, This indicates the preset response change threshold.

[0043] The overlap between candidate regions of two adjacent ship layers is represented by the cross-union ratio:

[0044] in, Indicates the first Layered ship candidate area, Indicates the first Layered ship candidate area, To prevent constants with a denominator of zero, when the overlap between candidate regions of two adjacent ship layers exceeds a preset overlap threshold... At that time, it was determined that the candidate regions of two adjacent layers were basically the same.

[0045] The criterion for the change in prior confidence between two adjacent ship decks is as follows: Let the first Layer Pixel The corresponding prior confidence level for the ship is: , No. Layer Pixel The corresponding prior confidence level for the ship is: .

[0046] The change in prior confidence between two adjacent ship decks is located as follows: When the following conditions are met: At that time, it was determined that the prior confidence level of the ship tended to stabilize. Among them, This represents the threshold for changes in the confidence level of the preset value.

[0047] Criterion for maximum number of iterations: Let the current iteration level be... The preset maximum number of iterations is When satisfied Stop iterating when the time is right.

[0048] The final output of the hyperspectral image ship detection results includes at least one of the following: ship target location, ship candidate region range, ship prior coupled RX anomaly response map, and ship target detection marker.

[0049] Through the above steps, the weak geometric matching degree of the ship and the local background difference matching degree of the candidate region-neighborhood background are unified into the ship prior confidence degree. The ship prior confidence degree is then used to participate in the background modeling sample weighting, the prior weighted background statistical model update, and the modulation of RX anomaly statistics, so that the RX detection process has the detection orientation towards ship targets.

[0050] Specific results are as follows Figure 5 As shown, Figure 5 (a) is a schematic diagram of the hyperspectral image to be detected and the region where the ship target is located; Figure 5 (b) shows the initial candidate regions determined based on the initial RX anomaly response map. It can be seen that in addition to the high anomaly response to the ship target, the initial RX anomaly detection also generates a lot of anomaly candidate regions for sea surface texture, local bright areas and other complex backgrounds. Figure 5 (c) to Figure 5 (e) shows the changes in candidate regions obtained by redetermining candidate regions based on the current anomaly response map in different iteration rounds, and updating the ship's prior confidence using the scale matching degree, directional extension matching degree, and local background difference matching degree between the candidate region and the neighboring background. As the ship's prior confidence continues to participate in the determination of background modeling weights and the modulation of RX anomaly statistics, some anomaly regions that do not conform to the ship's weak geometric features or have insufficient differences from the neighboring background are gradually suppressed. Figure 5 (f) To meet the preset stability conditions or reach the preset maximum number of iterations, the abnormal response caused by the complex sea surface background is further weakened compared with the initial candidate region, and the target region that conforms to the ship size and directional extension characteristics is preserved, so that the detection results are gradually concentrated in the ship target region.

[0051] Second Embodiment Based on the same inventive concept, this application also provides a hyperspectral ship detection system using a priori coupled iterative RX, employing the hyperspectral ship detection method described above, including: The response module is used to respond to the hyperspectral ship detection task, acquire the hyperspectral image to be detected and preprocess it, build an initial background statistical model, and obtain the initial RX anomaly response map as the current anomaly response map through RX anomaly detection. The matching module is used to determine candidate regions based on the current response map, construct a weak geometric matching degree for the ship based on the scale-constrained features and directional extension features of the candidate regions, determine the neighboring background regions of the candidate regions, and construct a local background difference matching degree based on at least one of the spectral differences, anomalous response differences, and neighborhood background consistency between the candidate regions and the neighboring background regions, and normalize and weightedly fuse the ship's weak geometric matching degree and the local background difference matching degree to obtain the ship's prior confidence and associate it with the corresponding pixel. The prior coupling detection module is used to determine the background modeling weights based on the ship's prior confidence, update the background mean vector and background covariance matrix based on the background modeling weights and calculate the RX anomaly statistic, and modulate the RX anomaly statistic using the ship's prior confidence to obtain the ship's prior coupled RX anomaly response map. The result output module is used to update the candidate region based on the ship's prior coupled RX anomaly response map. When the adjacent iteration results do not reach the preset stability condition and do not reach the preset maximum number of iterations, the ship's prior coupled RX anomaly response map is updated to the current anomaly response map, and the matching module, prior coupling detection module and iterative update module are controlled to repeatedly execute the corresponding processing; if the conditions are met, the ship detection result is output.

[0052] Furthermore, the prior coupling detection module includes, The prior background modeling weight determination unit is used to obtain the prior confidence of the ship corresponding to each pixel in the current iteration round, take the confidence complement of each ship's prior confidence relative to the single-position confidence, and determine the confidence complement as the background modeling weight of the corresponding pixel; The prior weighted background mean calculation unit is used to perform weighted summation using the spectral vector of each pixel and the corresponding background modeling weight, and to normalize the weighted summation result using the sum of the background modeling weights to obtain the prior weighted background mean vector of the current iteration. The prior weighted background covariance calculation unit calculates the deviation vector between the spectral vector of each pixel and the prior weighted background mean vector. Each deviation vector is multiplied by its corresponding transpose to obtain the corresponding deviation outer product matrix. The deviation outer product matrices are then weighted and summed using the corresponding background modeling weights. Finally, the sum of all background modeling weights is used to normalize the weighted sum, yielding the prior weighted background covariance matrix. The prior coupled RX calculation unit is used to calculate the RX anomaly statistics corresponding to each pixel using the prior weighted background mean vector and the prior weighted background covariance matrix, and to construct a prior modulation term using the ship prior confidence and prior modulation coefficient corresponding to each pixel. The prior modulation term is multiplied with the corresponding RX anomaly statistics to obtain the ship prior coupled RX anomaly statistics. The prior response map generation unit is used to map the corresponding ship prior coupled RX anomaly statistics to the image space according to the spatial location of each pixel, so as to obtain the ship prior coupled RX anomaly response map.

[0053] In one embodiment, the present invention also provides an electronic device, the electronic device including a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, it implements the hyperspectral image ship detection method based on ship prior coupling iterative RX as described in any of the above embodiments.

[0054] In one embodiment, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the hyperspectral image ship detection method based on ship prior coupling iterative RX as described in any of the above embodiments.

[0055] In the above embodiments, the prior confidence of the ship is not an independent post-processing screening condition, but a coupled variable involved in background modeling sample weighting, prior weighted background statistical model updating, and RX anomaly statistic modulation. Therefore, the traditional RX detection process is transformed into an iterative detection process constrained by the ship's prior information, making it applicable to ship target detection in low spatial resolution hyperspectral images.

[0056] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the foregoing claims.

[0057] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0058] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A hyperspectral ship detection method with prior coupled iterative RX, characterized in that, The steps include, Step S1: In response to the hyperspectral ship detection task, acquire the hyperspectral image to be detected and preprocess it, construct an initial background statistical model, and obtain the initial RX anomaly response map as the current anomaly response map through RX anomaly detection. Step S2: Determine candidate regions based on the current response map, and construct a weak geometric matching degree for ships based on the scale-constrained features and directional extension features of the candidate regions; Step S3: Determine the neighboring background region of the candidate region, and construct a local background difference matching degree based on at least one of the spectral difference, abnormal response difference and neighborhood background consistency between the candidate region and the neighboring background region; Step S4: Normalize and weightedly fuse the weak geometric matching degree of the ship and the local background difference matching degree to obtain the ship's prior confidence and associate it with the corresponding pixel; Step S5: Determine the background modeling weights based on the prior confidence of the ship, update the background mean vector and background covariance matrix based on the background modeling weights and calculate the RX anomaly statistic, modulate the RX anomaly statistic using the prior confidence of the ship to obtain the ship prior coupled RX anomaly response map; Step S6: Update the candidate region according to the ship prior coupled RX anomaly response map. When the adjacent iteration results do not reach the preset stability condition and do not reach the preset maximum number of iterations, update the ship prior coupled RX anomaly response map to the current anomaly response map, and repeat steps S2 to S6; if satisfied, output the ship detection result.

2. The hyperspectral ship detection method based on prior coupled iterative RX according to claim 1, characterized in that, Step S1 includes, The hyperspectral image to be detected is preprocessed, including radiometric correction and noise suppression, so that the spectral vectors of each pixel in the hyperspectral image to be detected are at a uniform data scale, thus obtaining a preprocessed hyperspectral image. Based on the spectral vectors of each pixel in the preprocessed hyperspectral image, an initial background mean vector and an initial background covariance matrix are calculated to obtain the initial background statistical model. Then, based on the initial background statistical model, the RX anomaly statistics of each pixel are calculated using an RX anomaly detection algorithm. The RX anomaly statistics for each pixel are as follows: ;in, The spectral vector representing the pixel, This represents the background mean vector. Represents the background covariance matrix. The matrix representing the inverse of the background covariance matrix. Represents the pixel The corresponding anomaly detection statistics; The corresponding RX anomaly statistics are mapped to the image space according to the spatial location of each pixel to obtain the initial RX anomaly response map as the current anomaly response map.

3. The hyperspectral ship detection method based on prior coupled iterative RX according to claim 2, characterized in that, Step S2 includes, Based on the distribution of abnormal response values ​​of each pixel in the current abnormal response map, a candidate response threshold is determined. Pixels whose abnormal response values ​​reach the candidate response threshold are identified as candidate pixels. Connectivity regions are aggregated based on the spatial adjacency relationship between the candidate pixels to obtain at least one candidate region. The number of pixels contained in each candidate region is counted. Based on the spatial resolution of the hyperspectral image and the preset actual size range of the ship, the preset pixel scale range corresponding to the ship is determined. Based on the matching degree between the number of pixels and the preset pixel scale range, the scale-limited features of each candidate region are obtained. The scale-limited features are then quantified to obtain the scale matching degree corresponding to each candidate region. The spatial coordinates of the pixels within each candidate region are obtained. A pixel coordinate covariance matrix is ​​constructed based on the spatial coordinates. Eigenvalue decomposition is performed on the pixel coordinate covariance matrix to obtain a first eigenvalue and a second eigenvalue, where the first eigenvalue is greater than the second eigenvalue. The directional extension feature of the candidate region is determined based on the first eigenvalue and the second eigenvalue. The directional extension feature is... ;in, This refers to the candidate region. To prevent small constants with a denominator of zero, The first feature value, This is the second eigenvalue; The directional extension features are normalized to obtain the directional extension matching degree corresponding to each candidate region. The scale matching degree and the directional extension matching degree are weighted and fused according to the corresponding geometric fusion weight to obtain the weak geometric matching degree of the ship corresponding to each candidate region.

4. The hyperspectral ship detection method based on prior coupled iterative RX according to claim 3, characterized in that, Step S3 includes, Based on the spatial location of each candidate region, a preset neighborhood range is expanded outward, and the pixels within the preset neighborhood range other than the candidate regions are determined as neighborhood background pixels, forming a corresponding neighborhood background region. The spectral vectors of each pixel in the candidate region and the neighboring background region are obtained respectively. A first average spectral vector is calculated based on the spectral vectors of each pixel in the candidate region, and a second average spectral vector is calculated based on the spectral vectors of each pixel in the neighboring background region. The spectral distance or spectral angle between the first average spectral vector and the second average spectral vector is calculated to obtain the spectral difference between the candidate region and the neighboring background region. The abnormal response values ​​of each pixel in the candidate region and the neighboring background region in the current abnormal response map are obtained respectively. The average abnormal response values ​​corresponding to the candidate region and the neighboring background region are calculated respectively, and the abnormal response difference is determined based on the difference between the two average abnormal response values. The abnormal response variance is calculated based on the abnormal response value of each pixel in the neighborhood background region, or the spectral variance is calculated based on the spectral vector of each pixel in the neighborhood background region, and the neighborhood background consistency is determined based on the abnormal response variance or the spectral variance. Normalize at least one of the spectral differences, the abnormal response differences, and the neighborhood background consistency, and perform weighted fusion according to a preset difference fusion weight to obtain the local background difference matching degree corresponding to each candidate region.

5. The hyperspectral ship detection method based on prior coupled iterative RX according to claim 4, characterized in that, Step S4 includes, The ship weak geometric matching degree and the local background difference matching degree corresponding to each candidate region are obtained. The ship weak geometric matching degree and the local background difference matching degree are normalized respectively, and the two are mapped to a unified numerical range to obtain the normalized ship weak geometric matching degree and the normalized local background difference matching degree. Based on preset first and second weighting coefficients, the normalized weak geometric matching degree of the vessel and the normalized local background difference matching degree are weighted and fused to obtain the prior confidence degree of the vessel corresponding to each candidate region. The prior confidence degree of the vessel is... , This refers to the candidate region. For the candidate region The corresponding prior confidence level of the ship. For the candidate region The corresponding normalized weak geometric matching degree of the ships, For the corresponding normalized local background difference matching degree, , These are the corresponding weight coefficients; The prior confidence of the ship corresponding to the candidate region is associated with each pixel in the candidate region to determine the prior confidence of the ship corresponding to the pixel.

6. The hyperspectral ship detection method based on prior coupled iterative RX according to claim 5, characterized in that, Step S5 includes, Obtain the prior confidence of the ship corresponding to each of the pixels in the current iteration round, take the confidence complement of each ship prior confidence relative to the single-location confidence, and determine the confidence complement as the background modeling weight of the corresponding pixel; The spectral vectors of each pixel and their corresponding background modeling weights are used to perform a weighted summation. The summation result is then normalized using the sum of all background modeling weights to obtain the prior weighted background mean vector for the current iteration. The prior weighted background mean vector is... ,in, Let the prior weighted background mean vector be... For the first The spectral vector of each pixel, For the pixel indexes involved in background modeling, The first iteration mentioned in the previous iteration round The background modeling weights corresponding to each pixel. The current iteration round; Calculate the deviation vector between the spectral vector of each pixel and the prior weighted background mean vector. Multiply each deviation vector by its corresponding transpose to obtain a corresponding deviation outer product matrix. Summate each deviation outer product matrix using the corresponding background modeling weights, and normalize the sum of the background modeling weights to obtain the prior weighted background covariance matrix. The prior weighted background covariance matrix is... ,in, Let the prior weighted background covariance matrix be represented. The first iteration in the previous iteration round The background modeling weights corresponding to each pixel Indicates the first The spectral vector of each pixel, The prior weighted background mean vector for the current iteration round. This represents the cell index that participated in the background modeling, and T represents the transpose operation; Using the prior weighted background mean vector and the prior weighted background covariance matrix, the RX anomaly statistic corresponding to each pixel is calculated. A prior modulation term is constructed using the prior ship confidence and prior modulation coefficients corresponding to each pixel. The prior modulation term is then multiplied by the corresponding RX anomaly statistic to obtain the ship prior coupled RX anomaly statistic. The ship prior coupled RX anomaly statistic is... ;in, This represents the a priori coupling RX anomaly statistic for ships. The spectral vector of the pixel. This is the prior weighted background mean vector. Let be the prior weighted background covariance matrix. Let T be the inverse of the prior weighted background covariance matrix, and let T denote the transpose operation. For the pixel The corresponding prior confidence level of the ship, λ represents the prior modulation coefficient; According to the spatial location of each pixel, the corresponding ship prior coupled RX anomaly statistics are mapped to the image space to obtain the ship prior coupled RX anomaly response map.

7. The hyperspectral ship detection method based on prior coupled iterative RX according to claim 6, characterized in that, Step S6 includes, Obtain the ship's prior coupled RX anomaly response map generated in the current iteration round, and count the anomaly response value of each pixel. Calculate the average value and standard deviation of the anomaly response value, and combine the average value and standard deviation according to the threshold adjustment coefficient to obtain the candidate response threshold corresponding to the current iteration round. The pixels whose abnormal response values ​​reach the candidate response threshold are identified as update candidate pixels, and connected regions are aggregated according to the spatial adjacency relationship between each update candidate pixel to obtain the update candidate region corresponding to the current iteration round. The a prior coupled RX anomaly response map and the update candidate region corresponding to the current iteration round are compared with the anomaly response map and the candidate region corresponding to the previous iteration round, respectively, to obtain the iterative change results corresponding to the adjacent iteration rounds, and based on the iterative change results, it is determined whether the adjacent iteration results have reached the preset stability condition. When the results of the adjacent iterations do not reach the preset stability condition and the current iteration number does not reach the preset maximum iteration number, the ship prior coupled RX abnormal response map generated in the current iteration round is used as the current abnormal response map in the next iteration round, and steps S2 to S6 are repeated. When the results of adjacent iterations reach the preset stability condition, or when the current iteration number reaches the preset maximum iteration number, the iteration is terminated, and the ship detection result is output according to the update candidate region corresponding to the current iteration round.

8. The hyperspectral ship detection method based on prior coupled iterative RX according to claim 7, characterized in that, The preset stability conditions include at least one of the following: the change in abnormal response between the ship's prior coupled RX abnormal response maps corresponding to adjacent iteration rounds is not greater than a preset response change threshold; the regional overlap between the candidate regions corresponding to adjacent iteration rounds is not less than a preset overlap threshold; and the change in confidence between the ship's prior confidence corresponding to adjacent iteration rounds is not greater than a preset confidence change threshold.

9. A hyperspectral ship detection system with prior coupled iterative RX, employing the hyperspectral ship detection method with prior coupled iterative RX as described in any one of claims 1 to 8, characterized in that, include, The response module is used to respond to the hyperspectral ship detection task, acquire the hyperspectral image to be detected and preprocess it, build an initial background statistical model, and obtain the initial RX anomaly response map as the current anomaly response map through RX anomaly detection. The matching module is used to determine candidate regions based on the current response map and to construct a weak geometric matching degree for ships based on the scale-constrained features and directional extension features of the candidate regions. Determine the neighboring background regions of the candidate region, and construct a local background difference matching degree based on at least one of the spectral differences, abnormal response differences, and neighborhood background consistency between the candidate region and the neighboring background regions; The weak geometric matching degree of the vessel and the local background difference matching degree are normalized and weighted and fused to obtain the prior confidence degree of the vessel and associate it with the corresponding pixel. The prior coupling detection module is used to determine the background modeling weights based on the prior confidence of the ship, update the background mean vector and background covariance matrix based on the background modeling weights and calculate the RX anomaly statistic, and modulate the RX anomaly statistic using the prior confidence of the ship to obtain the ship prior coupled RX anomaly response map. The result output module is used to update the candidate region based on the ship's prior coupled RX anomaly response map. When the adjacent iteration results do not reach the preset stability condition and do not reach the preset maximum number of iterations, the ship's prior coupled RX anomaly response map is updated to the current anomaly response map, and the matching module, the prior coupled detection module and the iterative update module are controlled to repeatedly execute the corresponding processing; if the conditions are met, the ship detection result is output.

10. The hyperspectral ship detection system with prior coupled iterative RX according to claim 9, characterized in that, The prior coupling detection module includes, The prior background modeling weight determination unit is used to obtain the prior confidence of the ship corresponding to each of the pixels in the current iteration round, take the confidence complement value of each ship prior confidence value relative to the single position confidence value, and determine the confidence complement value as the background modeling weight of the corresponding pixel; The prior weighted background mean calculation unit is used to perform a weighted summation using the spectral vector of each pixel and the corresponding background modeling weight, and to normalize the weighted summation result using the sum of each background modeling weight to obtain the prior weighted background mean vector of the current iteration. The prior weighted background covariance calculation unit is used to calculate the deviation vector between the spectral vector of each pixel and the prior weighted background mean vector. Each deviation vector is multiplied by its corresponding transpose to obtain a corresponding deviation outer product matrix. The corresponding background modeling weights are used to perform a weighted summation of the deviation outer product matrices, and the summation of the background modeling weights is used to normalize the weighted summation result to obtain the prior weighted background covariance matrix. The prior coupled RX calculation unit is used to calculate the RX anomaly statistics corresponding to each pixel using the prior weighted background mean vector and the prior weighted background covariance matrix, and to construct a prior modulation term using the ship prior confidence and prior modulation coefficient corresponding to each pixel. The prior modulation term is multiplied by the corresponding RX anomaly statistics to obtain the ship prior coupled RX anomaly statistics. The prior response map generation unit is used to map the corresponding ship prior coupled RX anomaly statistics to the image space according to the spatial position of each pixel, so as to obtain the ship prior coupled RX anomaly response map.

Images