A Hyperspectral Anomaly Detection Method and System Based on Multi-Feature Intelligent Optimization

Abstract

This invention provides a hyperspectral anomaly detection method and system based on multi-feature intelligent optimization. First, spectral band optimization is performed on the original hyperspectral data by jointly extracting the optimal spectral band set from the original spectral data across subspaces. Then, multi-scale, multi-type attribute filtering and differential operations are performed on each spectral band in the optimal set to extract the spatial spectral feature information of the anomalous target and amplify the difference between the target and the background. Next, a genetic algorithm is used to classify each feature profile and intelligently optimize the set of most representative spatial spectral feature profiles. Finally, the intelligently optimized spatial spectral feature profiles are threshold-weighted and summed to filter out background information and highlight the target location, jointly generating the anomaly detection result. This invention adopts a coarse-to-fine detection strategy, improving accuracy while achieving automated multi-target detection.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a hyperspectral anomaly detection method and system based on multi-feature intelligent optimization. Background Technology

[0002] Hyperspectral imaging organically combines imaging and spectral techniques, enabling the acquisition of not only the morphological features and spatial distribution of ground objects but also the detection of their radiance and spectral information, offering the unique advantages of integrated image and spectrum analysis. Compared to panchromatic or multispectral images, hyperspectral imaging provides refined diagnostic features of ground objects' spectra, allowing for accurate target detection and identification using "fingerprint spectra" that represent the authenticity of targets. It has wide applications in military target reconnaissance, marine scientific research, precision agriculture, environmental monitoring, and mineral exploration. Automatically detecting multiple anomalous targets of varying materials, shapes, and sizes within a scene has become a major challenge and focus of hyperspectral target detection.

[0003] Common hyperspectral anomaly detection methods can be categorized into statistical anomaly detection methods, nonlinear anomaly detection methods, representation-based anomaly detection methods, and unsupervised learning methods. These methods often have some limitations. For example, anomaly detection typically starts with the background, assuming that the background of a hyperspectral image can be represented by a mathematical model. However, anomalies cannot be represented by such a model. Therefore, existing methods are weak in handling complex background distributions in hyperspectral images, and they also have low utilization of spatial information, failing to fully leverage the complementary advantages of spectral and spatial information to effectively uncover the differences between anomalous targets and the background. This is especially true for hyperspectral images containing multiple heterogeneous targets, where detection performance is poor. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing methods in processing various complex background distributions in hyperspectral images and in utilizing image spatial information in a low manner. This invention proposes a hyperspectral anomaly detection method and system based on multi-feature intelligent optimization, which improves accuracy and achieves automated detection of multiple targets.

[0005] To achieve the above objectives, the present invention adopts the following specific technical solution:

[0006] This invention provides a hyperspectral anomaly detection system based on multi-feature intelligent optimization. The system includes a spectral band selection module, a spatial spectral feature extraction module, a spatial spectral feature profile optimization module, and an anomaly detection result generation module.

[0007] The spectral band selection module is used to select spectral bands from the original hyperspectral data through cross-subspace joint processing to form the optimal spectral band set;

[0008] The spatial spectral feature extraction module performs multi-scale and multi-type attribute filtering and difference operations on each spectral band in the optimal spectral band set to extract the spatial spectral feature information of the abnormal target, while suppressing background information and enhancing the difference between the target and the background.

[0009] The spatial spectral feature profile optimization module is used to classify feature profiles obtained with different attributes and different threshold parameters, and automatically and intelligently select the most representative set of spatial spectral feature profiles.

[0010] The anomaly detection result generation module is used to perform threshold-weighted summation of the most representative spatial spectral feature profiles selected by intelligent optimization, so that the target information is continuously superimposed and supplemented, background information is filtered out and the target position is highlighted, and anomaly detection results are jointly generated.

[0011] Preferably, the spectral band selection module includes a correlation coefficient calculation unit, a subspace partitioning unit, an optimal index factor calculation unit, and an optimal spectral band set extraction unit, wherein,

[0012] The correlation coefficient calculation unit uses the correlation coefficient matrix to calculate the correlation between each spectral band;

[0013] Subspace partitioning units are used to divide the original hyperspectral data into multiple subspaces;

[0014] The optimal index factor calculation unit is used to calculate the optimal index factor between each spectral segment in the same subspace and the two adjacent spectral segments, and to find the spectral segment with the largest information content and the smallest correlation to represent the subspace.

[0015] The optimal spectral band set extraction module is used to extract the spectral bands representing each subspace to obtain the optimal spectral band set representing the original hyperspectral data.

[0016] Preferably, the spatial spectrum feature extraction module includes a spatial structure feature representation unit, a threshold selection and difference operation unit for attribute filters, and a normalization processing unit for spatial spectrum feature profiles, wherein,

[0017] The spatial structure feature characterization unit is used to select four attribute types: area, diagonal length, standard deviation, and moment of inertia, to construct multiple spatial features and fully characterize the spatial structure features of hyperspectral data.

[0018] The threshold selection and differential operation unit for attribute filters is used for threshold selection of different attribute filters and acquisition of corresponding differential profiles. Multiple threshold parameters are used to describe the spatial information of anomalous targets at different scales.

[0019] The difference formula and threshold selection method for area attribute filters are as follows:

[0020] A i =|φ A (fi )-f i |+|f i -γ A (f i )|=φ A (f i )-γ A (f i (1),

[0021]

[0022] Where φ and γ represent the coarsening and thinning operations, respectively, and A i For the area attribute difference profile of the i-th spectral band, the threshold parameter for the area attribute is selected based on the spatial resolution ρ and the image size M×N, a max and a min These represent the upper and lower bounds of the area attribute threshold, respectively. Here, a... min =2 and a max <3×min(M,N) / ρ∩a max <20, δ A Set the step size to 1;

[0023] The difference formula and threshold selection method for the diagonal length attribute filter are as follows:

[0024] B i =|φ B (f i )-f i |+|f i -γ B (f i )|=φ B (f i )-γ B (f i (3),

[0025] λ B ={b min ,b min +δ B ,b min +2δ B ,...,b max} (4),

[0026] Among them, B i For the diagonal length attribute difference profile of the i-th spectral band, the threshold parameter for the diagonal length attribute is selected based on the spatial resolution ρ and the image size M×N, b max and b min These represent the upper and lower bounds of the diagonal length attribute threshold, where b is... min =3 and δ BSet the step size to 2;

[0027] The difference formula and threshold selection method for the standard deviation attribute filter are as follows:

[0028] S i =||φ s (f i )-f i |-|f i -γ s (f i )||(5),

[0029] λ S ={S min ,S min +δ S ,S min +2δ S ,...,S max}(6),

[0030] Among them, S i Let S be the standard deviation attribute difference profile of the i-th spectral band. The threshold parameter of the standard deviation is selected based on the mean of the spectral band. max and S min These represent the upper and lower bounds of the standard deviation attribute threshold, respectively 25% and 5% of the mean, δ. s Set the step size to 2.5% of the mean;

[0031] The difference formula and threshold selection method for the moment of inertia property filter are as follows:

[0032] I i =||φ I (f i )-f i |-|f i -γ I (f i )||(7),

[0033] λ I ={I min ,I min +δ I ,I min +2δ I ,...,I max}(8),

[0034] Among them, I i Let I be the difference profile of the moment of inertia property for the i-th spectral band. max and I min These are the upper and lower bounds of the inertia moment attribute threshold, respectively. Here, I is... max =0.9 and I min =0.1, δ ISet the step size to 0.1;

[0035] The normalization unit for spatial spectral feature profiles normalizes feature profiles of different scales and attributes obtained by the difference operation, which facilitates the analysis and comparison of differences between different feature profiles.

[0036] Preferably, the spatial spectral feature profile optimization module includes a spatial spectral feature profile numbering unit, a difference coefficient calculation unit, and an optimal feature profile screening unit, wherein,

[0037] The spatial spectral feature profile numbering unit is used to number the normalized spatial spectral feature profiles as the encoding value of each chromosome in the genetic algorithm. The chromosomes adopt a real number encoding method, ranging from 1 to the total number of spatial spectral feature profiles q. The number of chromosomes contained in an individual is the expected number of representative feature profiles n.

[0038] When the profile combination is represented by the fitness function as the optimal feature profile, the formula for calculating the ratio of intra-class to inter-class dissimilarity coefficients is as follows:

[0039]

[0040] Among them, f ci and f cj For different optimal feature profiles, f h For a member of class ci, C ci This represents the number of members belonging to class ci. D NMI The coefficient of variation represents the difference between two feature profiles. A smaller coefficient indicates a greater difference between the two profiles. NMI The calculation formula is as follows:

[0041]

[0042] Among them, G i and G j f i and f j The set of gray values, g x and g y These are the individual grayscale values, p(g) x g y Then it is the joint probability mass function;

[0043] The optimal feature profile selection unit performs selection, crossover, and mutation operations, repeatedly calculates the fitness function of each individual in the population, and intelligently selects the feature profile combination with the smallest intra-class difference and the largest inter-class difference, which is the optimal feature profile set representing the spatial-spectral characteristics of hyperspectral data.

[0044] Preferably, the anomaly detection result generation module includes a feature profile threshold calculation unit and a detection result acquisition unit, wherein,

[0045] The feature profile threshold calculation unit is used to perform thresholding on the most representative spatial spectral feature profiles selected by intelligent optimization. The feature profile threshold is calculated using the Otsu algorithm with maximum inter-class variance, and the calculation formula is as follows:

[0046] B i =Thresh(f i ,θ)(11),

[0047] Where Thresh is the threshold calculation symbol, and θ is the feature profile threshold;

[0048] The detection result acquisition unit is used to determine the number of members C of each selected feature profile in the genetic algorithm classification. ci The ratio of the number of spatial spectral feature profiles to the total number of feature profiles q is used as a weight. The optimal feature profiles after thresholding are then summed using a weighted average to obtain the final detection result. x and y are the horizontal and vertical coordinates of the image, respectively. The calculation formula for the detection result is as follows:

[0049]

[0050] A hyperspectral anomaly detection method based on multi-feature intelligent optimization specifically includes the following steps:

[0051] S1. Perform spectral band optimization on the original hyperspectral data and extract the optimal spectral band set from the original spectral data through cross-subspace joint extraction;

[0052] S2. Perform multi-scale and multi-type attribute filtering and difference operations on each spectral band in the optimal spectral band set to extract the spatial spectral feature information of the abnormal target, suppress background information, and enhance the difference between the target and the background.

[0053] S3. Use a genetic algorithm to classify each feature profile and intelligently select the most representative set of spatial spectral feature profiles.

[0054] S4. The most representative spatial spectral feature profiles selected by intelligent optimization are subjected to threshold weighted summation to filter out background information and highlight the target location, and jointly generate anomaly detection results.

[0055] Preferably, step S1 specifically includes the following steps:

[0056] S11. Calculate the correlation between each spectral band using the correlation coefficient matrix, and divide the hyperspectral data into multiple subspaces;

[0057] S12. Calculate the optimal index factor between each spectral segment in the same subspace and the two adjacent spectral segments, and find the spectral segment with the maximum information content and the minimum correlation to represent the subspace.

[0058] S13. Extract the spectral segments representing each subspace to obtain the optimal set of spectral segments that can represent the original hyperspectral data.

[0059] Preferably, step S2 specifically includes the following steps:

[0060] S21. Select four attribute types: area, diagonal length, standard deviation, and moment of inertia to construct multiple spatial features and fully characterize the spatial structure features of hyperspectral data.

[0061] S22. For different attribute types, multiple threshold parameters are used to describe the spatial information of anomalous targets at different scales, among which,

[0062] The difference formula and threshold selection method for area attribute filters are as follows:

[0063] A i =|φ A (f i )-f i |+|f i -γ A (f i )|=φ A (f i )-γ A (f i (1),

[0064]

[0065] Where φ and γ represent the coarsening and thinning operations, respectively, and A i For the area attribute difference profile of the i-th spectral band, the threshold parameter for the area attribute is selected based on the spatial resolution ρ and the image size M×N, a max and a min These represent the upper and lower bounds of the area attribute threshold, respectively. Here, a... min =2 and a max <3×min(M,N) / ρ∩a max <20, δ A Set the step size to 1;

[0066] The difference formula and threshold selection method for the diagonal length attribute filter are as follows:

[0067] B i =|φ B (f i )-f i |+|f i -γB (f i )|=φ B (f i )-γ B (f i (3),

[0068] λ B ={b min ,b min +δ B ,b min +2δ B ,...,b max} (4),

[0069] Among them, B i For the diagonal length attribute difference profile of the i-th spectral band, the threshold parameter for the diagonal length attribute is selected based on the spatial resolution ρ and the image size M×N, b max and b min These represent the upper and lower bounds of the diagonal length attribute threshold, where b is... min =3 and δ B Set the step size to 2;

[0070] The difference formula and threshold selection method for the standard deviation attribute filter are as follows:

[0071] S i =||φ s (f i )-f i |-|f i -γ s (f i )||(5),

[0072] λ S ={S min ,S min +δ S ,S min +2δ S ,...,S max}(6),

[0073] Among them, S i Let S be the standard deviation attribute difference profile of the i-th spectral band. The threshold parameter of the standard deviation is selected based on the mean of the spectral band. max and S min These represent the upper and lower bounds of the standard deviation attribute threshold, respectively 25% and 5% of the mean, δ. s Set the step size to 2.5% of the mean;

[0074] The difference formula and threshold selection method for the moment of inertia property filter are as follows:

[0075] Ii =||φ I (f i )-f i |-|f i -γ I (f i )||(7),

[0076] λ I ={I min ,I min +δ I ,I min +2δ I ,...,I max}(8),

[0077] Among them, I i Let I be the difference profile of the moment of inertia property for the i-th spectral band. max and I min These are the upper and lower bounds of the inertia moment attribute threshold, respectively. Here, I is... max =0.9 and I min =0.1, δ I Set the step size to 0.1;

[0078] S23. By normalizing the feature profiles of different scales and attributes obtained by the difference operation, it is easier to analyze and compare the differences between different feature profiles.

[0079] Preferably, step S3 specifically includes the following steps:

[0080] S31. Number the normalized spatial spectral feature profiles and use them as the encoding values ​​of each chromosome in the genetic algorithm. Chromosomes are encoded using real numbers, ranging from 1 to the total number of spatial spectral feature profiles q. The number of chromosomes in an individual is the expected number of representative feature profiles n.

[0081] S32. The ratio of the difference coefficients between within and between classes when this profile combination is used as the optimal feature profile is expressed by the fitness function, and the formula is as follows:

[0082]

[0083] Among them, f ci and f cj For different optimal feature profiles, f h For a member of class ci, C ci This represents the number of members belonging to class ci. D NMI The coefficient of difference represents the difference between two feature profiles. The smaller the value, the greater the difference between the two feature profiles. The formula for calculating it is as follows:

[0084]

[0085] Among them, G i and G j f i and f j The set of gray values, g x and g y These are the individual grayscale values, p(g) x g y Then it is the joint probability mass function;

[0086] S33. By performing selection, crossover, and mutation operations, the fitness function of each individual in the population is repeatedly calculated, and the feature profile combination with the smallest intra-class difference and the largest inter-class difference is intelligently selected, which is the optimal feature profile set representing the spatial-spectral characteristics of hyperspectral data.

[0087] Preferably, step S4 specifically includes the following steps:

[0088] S41. The most representative spatial spectral feature profiles selected by intelligent optimization are thresholded. The Otsu algorithm with maximum inter-class variance is used to calculate the feature profile threshold. The calculation formula is as follows:

[0089] B i =Thresh(f i ,θ)(11),

[0090] Where Thresh is the threshold calculation symbol, and θ is the feature profile threshold;

[0091] S42. The detection result acquisition unit will determine the number of in-class members C of each selected feature profile in the genetic algorithm classification. ci The ratio of the number of spatial spectral feature profiles to the total number of feature profiles q is used as a weight. The optimal feature profiles after thresholding are then summed using a weighted average to obtain the final detection result. x and y are the horizontal and vertical coordinates of the image, respectively. The calculation formula for the detection result is as follows:

[0092]

[0093] The present invention can achieve the following technical effects:

[0094] 1. While ensuring high accuracy and low false alarm rate, improve the accuracy of detecting multi-scale abnormal targets.

[0095] 2. Adopting a coarse-to-fine detection strategy, we fully utilize the complementary technical advantages of spectral and spatial information. Through multi-scale and multi-type attribute filtering and differential operations, we can fully characterize the spatial structure features of hyperspectral data and enhance the ability to distinguish between targets and background.

[0096] 3. It adopts a genetic algorithm to automatically optimize and reduce the amount of data to be processed and redundant information, making it suitable for multi-target anomaly detection, thus making it highly intelligent and more robust. Attached Figure Description

[0097] Figure 1 This is a schematic diagram of the module composition of a hyperspectral anomaly detection system based on multi-feature intelligent optimization according to an embodiment of the present invention.

[0098] Figure 2 This is a schematic diagram of the structural units of the spectral selection module provided in an embodiment of the present invention.

[0099] Figure 3 This is a schematic diagram of the structural units of the spatial spectrum feature extraction module provided in an embodiment of the present invention.

[0100] Figure 4 This is a schematic diagram of the structural unit of the spatial spectrum feature profile optimization module provided in an embodiment of the present invention.

[0101] Figure 5 This is a schematic diagram of the structural units of the anomaly detection result generation module provided in an embodiment of the present invention.

[0102] Figure 6a This is a schematic diagram of a two-dimensional correlation coefficient matrix provided according to an embodiment of the present invention.

[0103] Figure 6b This is a schematic diagram of a three-dimensional correlation coefficient matrix provided according to an embodiment of the present invention.

[0104] Figure 7 This is a flowchart of a hyperspectral anomaly detection method based on multi-feature intelligent optimization provided by an embodiment of the present invention.

[0105] Figure 8a It is an RGB pseudo-color image of the HYDICE dataset provided according to an embodiment of the present invention.

[0106] Figure 8b This is a truth graph of the HYDICE dataset provided according to an embodiment of the present invention.

[0107] Figure 8c This is a schematic diagram illustrating the detection performance of the HYDICE dataset provided in an embodiment of the present invention. Detailed Implementation

[0108] In the following description, embodiments of the invention will be described with reference to the accompanying drawings. In the description below, the same modules are denoted by the same reference numerals. Where the same reference numerals are used, their names and functions are also the same. Therefore, their detailed description will not be repeated.

[0109] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not constitute a limitation thereof.

[0110] Figure 1 The module composition of a hyperspectral anomaly detection system based on multi-feature intelligent optimization, provided according to an embodiment of the present invention, is illustrated below. Figure 1 The module units of the hyperspectral anomaly detection system based on multi-feature intelligent optimization proposed in this invention will be described in detail.

[0111] like Figure 1 As shown, the hyperspectral anomaly detection system based on multi-feature intelligent optimization includes a spectral band selection module, a spatial spectral feature extraction module, a spatial spectral feature profile optimization module, and an anomaly detection result generation module.

[0112] The spectral selection module is used to filter the original hyperspectral data by cross-subspace joint spectral selection to form the optimal spectral set. Figure 2 The structural unit of the spectral band selection module provided according to an embodiment of the present invention is shown, such as Figure 2 As shown, the spectral band selection module includes a correlation coefficient calculation unit, a subspace partitioning unit, an optimal index factor calculation unit, and an optimal spectral band set extraction unit.

[0113] Figure 6a and Figure 6b The correlation coefficient matrix provided according to the embodiments of the present invention is displayed from two dimensions, as follows: Figure 6a and Figure 6b As shown, the correlation coefficient calculation unit uses the correlation coefficient matrix to calculate the correlation between each spectral band.

[0114] Subspace partitioning units are used to divide the original hyperspectral data into multiple subspaces.

[0115] The optimal index factor calculation unit is used to calculate the optimal index factor between each spectral segment in the same subspace and the two adjacent spectral segments, and to find the spectral segment with the maximum information content and the minimum correlation to represent the subspace.

[0116] The optimal spectral band set extraction unit is used to extract the spectral bands representing each subspace to obtain the optimal spectral band set representing the original hyperspectral data.

[0117] The spatial spectral feature extraction module performs multi-scale and multi-type attribute filtering and differential operations on each spectral band in the optimal spectral band set to extract the spatial spectral feature information of the abnormal target, while suppressing background information and enhancing the difference between the target and the background. Figure 3 The structural units of the spatial spectral feature extraction module provided according to an embodiment of the present invention are shown, such as... Figure 3As shown, the spatial spectrum feature extraction module includes a spatial structure feature representation unit, a threshold selection and difference operation unit for attribute filters, and a normalization processing unit for spatial spectrum feature profiles.

[0118] The spatial structure feature characterization unit is used to select four attribute types: area, diagonal length, standard deviation, and moment of inertia, to construct multiple spatial features and fully characterize the spatial structure features of hyperspectral data.

[0119] The threshold selection and differential operation unit for attribute filters is used for threshold selection of different attribute filters and acquisition of corresponding differential profiles. Multiple threshold parameters are used to describe the spatial information of anomalous targets at different scales.

[0120] The difference formula and threshold selection method for area attribute filters are as follows:

[0121] A i =|φ A (f i )-f i |+|f i -γ A (f i )|=φ A (f i )-γ A (f i (1),

[0122]

[0123] Where φ and γ represent the coarsening and thinning operations, respectively, and A i For the area attribute difference profile of the i-th spectral band, the threshold parameter for the area attribute is selected based on the spatial resolution ρ and the image size M×N, a max and a min These represent the upper and lower bounds of the area attribute threshold, respectively. Here, a... min =2 and a max <3×min(M,N) / ρ∩a max <20, δ A Set the step size to 1.

[0124] The difference formula and threshold selection method for the diagonal length attribute filter are as follows:

[0125] B i =|φ B (f i )-f i |+|f i -γ B (f i )|=φ B (f i )-γ B(f i (3),

[0126] λ B ={b min ,b min +δ B ,b min +2δ B ,...,b max} (4),

[0127] Among them, B i For the diagonal length attribute difference profile of the i-th spectral band, the threshold parameter for the diagonal length attribute is selected based on the spatial resolution ρ and the image size M×N, b max and b min These represent the upper and lower bounds of the diagonal length attribute threshold, where b is... min =3 and δ B Set the step size to 2.

[0128] The difference formula and threshold selection method for the standard deviation attribute filter are as follows:

[0129] S i =||φ s (f i )-f i |-|f i -γ s (f i )||(5),

[0130] λ S ={S min ,S min +δ S ,S min +2δ S ,...,S max}(6),

[0131] Among them, S i Let S be the standard deviation attribute difference profile of the i-th spectral band. The threshold parameter of the standard deviation is selected based on the mean of the spectral band. max and S min These represent the upper and lower bounds of the standard deviation attribute threshold, respectively 25% and 5% of the mean, δ. s Set the step size to 2.5% of the mean.

[0132] The difference formula and threshold selection method for the moment of inertia property filter are as follows:

[0133] I i =||φ I (f i )-f i |-|fi -γ I (f i )||(7),

[0134] λ I ={I min ,I min +δ I ,I min +2δ I ,...,I max}(8),

[0135] Among them, I i Let I be the difference profile of the moment of inertia property for the i-th spectral band. max and I min These are the upper and lower bounds of the inertia moment attribute threshold, respectively. Here, I is... max =0.9 and I min =0.1, δ I Set the step size to 0.1.

[0136] The normalization unit of the spatial spectrum feature profile normalizes feature profiles of different scales and attributes obtained by difference operation, which facilitates the analysis and comparison of differences between different feature profiles.

[0137] The spatial spectral feature profile optimization module is used to classify feature profiles obtained with different attributes and different threshold parameters, and automatically and intelligently select the most representative set of spatial spectral feature profiles. Figure 4 The structural unit of the spatial spectrum feature profile optimization module provided according to an embodiment of the present invention is shown, such as... Figure 4 As shown, the spatial spectral feature profile optimization module includes a spatial spectral feature profile numbering unit, a difference coefficient calculation unit, and an optimal feature profile selection unit, wherein,

[0138] The spatial spectral feature profile numbering unit is used to number the normalized spatial spectral feature profiles as the encoding values ​​of each chromosome in the genetic algorithm. The chromosomes adopt a real number encoding method, ranging from 1 to the total number of spatial spectral feature profiles q. The number of chromosomes contained in an individual is the expected number of representative feature profiles n.

[0139] When the profile combination is represented by the fitness function as the optimal feature profile, the formula for calculating the ratio of intra-class to inter-class dissimilarity coefficients is as follows:

[0140]

[0141] Among them, f ci and f cj For different optimal feature profiles, f h For a member of class ci, C ciThis represents the number of members belonging to class ci. D NMI The coefficient of variation represents the difference between two feature profiles. A smaller coefficient indicates a greater difference between the two profiles. NMI The calculation formula is as follows:

[0142]

[0143] Among them, G i and G j f i and f j The set of gray values, g x and g y These are the individual grayscale values, p(g) x g y Then ) is the joint probability mass function.

[0144] The optimal feature profile selection unit performs selection, crossover, and mutation operations, repeatedly calculates the fitness function of each individual in the population, and intelligently selects the feature profile combination with the smallest intra-class difference and the largest inter-class difference, which is the optimal feature profile set representing the spatial-spectral characteristics of hyperspectral data.

[0145] The anomaly detection result generation module is used to perform threshold-weighted summation on the most representative optimal feature profiles selected by intelligent optimization, so that the target information is continuously superimposed and supplemented, background information is filtered out and the target position is highlighted, and anomaly detection results are jointly generated. Figure 5 The structural units of the anomaly detection result generation module provided according to an embodiment of the present invention are shown, such as... Figure 5 As shown, the anomaly detection result generation module includes a feature profile threshold calculation unit and a detection result acquisition unit, wherein,

[0146] The feature profile threshold calculation unit is used to perform thresholding on the most representative spatial spectral feature profiles selected by intelligent optimization. The feature profile threshold is calculated using the Otsu algorithm with maximum inter-class variance, and the calculation formula is as follows:

[0147] B i =Thresh(f i ,θ)(11),

[0148] Where Thresh is the threshold calculation symbol, and θ is the feature profile threshold.

[0149] The detection result acquisition unit will determine the number of in-class members C of each selected feature profile in the genetic algorithm classification. ci The ratio of the number of spatial spectral feature profiles to the total number of feature profiles q is used as a weight. The optimal feature profiles after thresholding are then summed using a weighted average to obtain the final detection result. x and y are the horizontal and vertical coordinates of the image, respectively. The calculation formula for the detection result is as follows:

[0150]

[0151] Figure 7 The specific flow of the hyperspectral anomaly detection method based on multi-feature intelligent optimization provided by an embodiment of the present invention is shown, such as... Figure 7 As shown, the specific operation steps are as follows:

[0152] S1. Perform spectral band optimization on the original hyperspectral data and extract the optimal spectral band set from the original spectral data through cross-subspace joint extraction.

[0153] S11. Calculate the correlation between each spectral band using the correlation coefficient matrix, and divide the hyperspectral data into multiple subspaces.

[0154] S12, such as Figure 6a and Figure 6b As shown, the optimal index factor between each spectral segment in the same subspace and the two adjacent spectral segments is calculated, and the spectral segment with the maximum information content and the minimum correlation is found to represent the subspace.

[0155] S13. Extract the spectral segments representing each subspace to obtain the optimal set of spectral segments that can represent the original hyperspectral data.

[0156] S2. Perform multi-scale and multi-type attribute filtering and differential operations on each spectral band in the optimal spectral band set to extract the spatial spectral feature information of the abnormal target, suppress background information, and enhance the difference between the target and the background.

[0157] S21. Select four attribute types: area, diagonal length, standard deviation, and moment of inertia, to construct multiple spatial features and fully characterize the spatial structure features of hyperspectral data.

[0158] S22. For different attribute types, multiple threshold parameters are used to describe the spatial information of anomalous targets at different scales, among which,

[0159] The difference formula and threshold selection method for area attribute filters are as follows:

[0160] A i =|φ A (f i )-f i |+|f i -γ A (f i )|=φ A (f i )-γ A (f i (1),

[0161]

[0162] Where φ and γ represent the coarsening and thinning operations, respectively, and A i For the area attribute difference profile of the i-th spectral band, the threshold parameter for the area attribute is selected based on the spatial resolution ρ and the image size M×N, a max and a min These represent the upper and lower bounds of the area attribute threshold, respectively. Here, a... min =2 and a max <3×min(M,N) / ρ∩a max <20, δ A Set the step size to 1.

[0163] The difference formula and threshold selection method for the diagonal length attribute filter are as follows:

[0164] B i =|φ B (f i )-f i |+|f i -γ B (f i )|=φ B (f i )-γ B (f i (3),

[0165] λ B ={b min ,b min +δ B ,b min +2δ B ,...,b max} (4),

[0166] Among them, B i For the diagonal length attribute difference profile of the i-th spectral band, the threshold parameter for the diagonal length attribute is selected based on the spatial resolution ρ and the image size M×N, b max and b min These represent the upper and lower bounds of the diagonal length attribute threshold, where b is... min =3 and δ B Set the step size to 2.

[0167] The difference formula and threshold selection method for the standard deviation attribute filter are as follows:

[0168] S i =||φ s (f i )-f i |-|f i -γ s (f i )||(5),

[0169] λ S ={S min ,S min +δ S ,S min +2δ S ,...,S max}(6),

[0170] Among them, S i Let S be the standard deviation attribute difference profile of the i-th spectral band. The threshold parameter of the standard deviation is selected based on the mean of the spectral band. max and S min These represent the upper and lower bounds of the standard deviation attribute threshold, respectively 25% and 5% of the mean, δ. s Set the step size to 2.5% of the mean.

[0171] The difference formula and threshold selection method for the moment of inertia property filter are as follows:

[0172] I i =||φ I (f i )-f i |-|f i -γ I (f i )||(7),

[0173] λ I ={I min ,I min +δ I ,I min +2δ I ,...,I max}(8),

[0174] Among them, I i Let I be the difference profile of the moment of inertia property for the i-th spectral band. max and I min These are the upper and lower bounds of the inertia moment attribute threshold, respectively. Here, I is... max =0.9 and I min =0.1, δ I Set the step size to 0.1.

[0175] S23. By normalizing the feature profiles of different scales and attributes obtained by the difference operation, it is easier to analyze and compare the differences between different feature profiles.

[0176] S3. Use a genetic algorithm to classify each feature profile and intelligently select the most representative set of spatial spectral feature profiles.

[0177] S31. Number the normalized spatial spectral feature profiles and use them as the encoding values ​​of each chromosome in the genetic algorithm. Chromosomes are encoded using real numbers, ranging from 1 to the total number of spatial spectral feature profiles q. The number of chromosomes in an individual is the expected number of representative feature profiles n.

[0178] S32. The ratio of the difference coefficients between within and between classes when this profile combination is used as the optimal feature profile is expressed by the fitness function, and the formula is as follows:

[0179]

[0180] Among them, f ci and f cj For different optimal feature profiles, f h For a member of class ci, C ci This represents the number of members belonging to class ci. D NMI The coefficient of difference represents the difference between two feature profiles. The smaller the value, the greater the difference between the two feature profiles. The formula for calculating it is as follows:

[0181]

[0182] Among them, G i and G j f i and f j The set of gray values, g x and g y These are the individual grayscale values, p(g) x g y Then ) is the joint probability mass function.

[0183] S33. By performing selection, crossover, and mutation operations, the fitness function of each individual in the population is repeatedly calculated, and the feature profile combination with the smallest intra-class difference and the largest inter-class difference is intelligently selected, which is the optimal feature profile set representing the spatial-spectral characteristics of hyperspectral data.

[0184] S4. The most representative spatial spectral feature profiles selected by intelligent optimization are subjected to threshold weighted summation to filter out background information and highlight the target location, and jointly generate anomaly detection results.

[0185] S41. The most representative spatial spectral feature profiles selected by intelligent optimization are thresholded. The Otsu algorithm with maximum inter-class variance is used to calculate the feature profile threshold. The calculation formula is as follows:

[0186] B i =Thresh(f i ,θ)(11),

[0187] Where Thresh is the threshold calculation symbol, and θ is the feature profile threshold.

[0188] S42. Calculate the number of in-class members C of each selected feature profile in the genetic algorithm classification. ci The ratio of the number of spatial spectral feature profiles to the total number of feature profiles q is used as a weight. The optimal feature profiles after thresholding are then summed using a weighted average to obtain the final detection result. x and y are the horizontal and vertical coordinates of the image, respectively. The calculation formula for the detection result is as follows:

[0189]

[0190] An example provided by an embodiment of the present invention is described as follows: An image of a US town is acquired using a HYDICE sensor. The image includes targets of various materials, shapes, and sizes, such as rooftops and vehicles, and the background includes roads, farmland, etc., exhibiting a certain degree of complexity. The result is as follows... Figures 8a-8c As shown, where, Figure 8a An RGB pseudo-color image of the HYDICE dataset provided according to an embodiment of the present invention is shown. Figure 8b A truth graph of the HYDICE dataset provided according to an embodiment of the present invention is shown. Figure 8c The detection performance of the HYDICE dataset provided according to an embodiment of the present invention is shown.

[0191] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0192] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

[0193] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A hyperspectral anomaly detection system based on multi-feature intelligent optimization, characterized in that, The hyperspectral anomaly detection system based on multi-feature intelligent optimization includes a spectral band selection module, a spatial spectral feature extraction module, a spatial spectral feature profile optimization module, and an anomaly detection result generation module, wherein... The spectral band selection module is used to select spectral bands from the original hyperspectral data through cross-subspace joint processing to form an optimal spectral band set. The spatial spectral feature extraction module performs multi-scale, multi-type attribute filtering and differential operations on each spectral band in the optimal spectral band set to extract the spatial spectral feature information of the abnormal target, while suppressing background information and enhancing the difference between the target and the background. The spatial spectral feature profile optimization module is used to classify the feature profiles obtained with different attributes and different threshold parameters, and automatically and intelligently select the most representative set of spatial spectral feature profiles. The spatial spectral feature profile optimization module includes a spatial spectral feature profile numbering unit, a difference coefficient calculation unit, and an optimal feature profile selection unit, wherein... The spatial spectrum feature profile numbering unit is used to number the normalized spatial spectrum feature profile as the encoding value of each chromosome in the genetic algorithm. The chromosome adopts a real number encoding method, ranging from 1 to the total number q of the spatial spectrum feature profiles. The number of chromosomes contained in an individual is the expected number of representative feature profiles n. When the combination of profiles represented by the fitness function is used as the optimal feature profile, the formula for calculating the ratio of intra-class to inter-class difference coefficients in the difference coefficient calculation unit is as follows: (9)； Among them, f ci and f cj For different optimal feature profiles, f h For a member of class ci, C ci Then it is the number of members belonging to class ci, D NMI The coefficient of variation represents the difference between two feature profiles. A smaller coefficient indicates a greater difference between the two profiles. NMI The calculation formula is as follows: (10)； Among them, G i and G j f i and f j The set of gray values, g x and g y These are the individual grayscale values, p(g) x g y Then it is the joint probability mass function; The optimal feature profile screening unit performs selection, crossover, and mutation operations, repeatedly calculates the fitness function of each individual in the population, and intelligently selects the feature profile combination with the smallest intra-class difference and the largest inter-class difference, which is the optimal feature profile set representing the spatial-spectral characteristics of the hyperspectral data. The anomaly detection result generation module is used to perform threshold-weighted summation on the most representative spatial spectral feature profiles selected by intelligent optimization, so that the target information is continuously superimposed and supplemented, background information is filtered out and the target position is highlighted, and anomaly detection results are jointly generated.

2. The hyperspectral anomaly detection system based on multi-feature intelligent optimization according to claim 1, characterized in that, The spectral band selection module includes a correlation coefficient calculation unit, a subspace partitioning unit, an optimal index factor calculation unit, and an optimal spectral band set extraction unit. The correlation coefficient calculation unit uses the correlation coefficient matrix to calculate the correlation between each spectral band; The subspace partitioning unit is used to divide the original hyperspectral data into multiple subspaces; The optimal index factor calculation unit is used to calculate the optimal index factor between each spectral segment in the same subspace and the two adjacent spectral segments, and find the spectral segment with the largest information content and the smallest correlation to represent the subspace. The optimal spectral band set extraction module is used to extract the spectral bands representing each subspace to obtain the optimal spectral band set representing the original hyperspectral data.

3. The hyperspectral anomaly detection system based on multi-feature intelligent optimization according to claim 1, characterized in that, The spatial spectrum feature extraction module includes a spatial structure feature representation unit, a threshold selection and difference operation unit for attribute filters, and a normalization processing unit for spatial spectrum feature profiles. The spatial structure feature characterization unit is used to select four attribute types: area, diagonal length, standard deviation, and moment of inertia, to construct multiple spatial features and fully characterize the spatial structure features of hyperspectral data. The threshold selection and differential operation unit of the attribute filter is used for threshold selection of different attribute filters and acquisition of corresponding differential profiles. Multiple threshold parameters are used to describe the spatial information of anomalous targets at different scales. The difference formula and threshold selection method for area attribute filters are as follows: (1)； (2)； in, and These represent the coarsening and thinning operations, respectively. i For the area attribute difference profile of the i-th spectral band, the threshold parameter of the area attribute is selected in relation to the spatial resolution. Related to the image size M×N, a max and a min These represent the upper and lower bounds of the area attribute threshold, respectively. Here, a... min =2 and δ A Set the step size to 1; The difference formula and threshold selection method for the diagonal length attribute filter are as follows: (3)； (4)； Among them, B i The diagonal length attribute difference profile is the i-th spectral band, and the threshold parameter of the diagonal length attribute is selected in relation to the spatial resolution. Related to the image size M×N, b max and b min These are the upper and lower bounds of the diagonal length attribute threshold, which are here... and δ B Set the step size to 2; The difference formula and threshold selection method for the standard deviation attribute filter are as follows: (5)； (6)； Among them, S i Let S be the standard deviation attribute difference profile of the i-th spectral band, where the threshold parameter of the standard deviation is selected based on the mean of the spectral band. max and S min These represent the upper and lower bounds of the standard deviation attribute threshold, respectively 25% and 5% of the mean, δ. s Set the step size to 2.5% of the mean. The difference formula and threshold selection method for the moment of inertia property filter are as follows: (7)； (8)； Among them, I i Let I be the difference profile of the moment of inertia for the i-th spectral band. max and I min These are the upper and lower bounds of the threshold value of the moment of inertia, respectively. Here, I is... max =0.9 and I min =0.1, δ I Set the step size to 0.1; The normalization processing unit of the spatial spectrum feature profile normalizes the feature profiles of different scales and attributes obtained by the difference operation, which facilitates the analysis and comparison of differences between different feature profiles.

4. The hyperspectral anomaly detection system based on multi-feature intelligent optimization according to claim 1, characterized in that, The anomaly detection result generation module includes a feature profile threshold calculation unit and a detection result acquisition unit, wherein... The feature profile threshold calculation unit is used to perform thresholding on the most representative spatial spectral feature profiles selected by intelligent optimization. The feature profile threshold is calculated using the Otsu algorithm with maximum inter-class variance, and the calculation formula is as follows: (11)； Where Thresh is the threshold calculation symbol, and θ is the feature profile threshold; The detection result acquisition unit is used to determine the number C of each selected feature profile within its class in the genetic algorithm classification. ci The ratio of the number of spatial spectral feature profiles to the total number of q is used as a weight to perform a weighted summation on each optimal feature profile after the threshold, obtaining the final detection result. x and y are the horizontal and vertical coordinates of the image, respectively. The calculation formula for the detection result is as follows: (12)。 5. A hyperspectral anomaly detection method based on multi-feature intelligent optimization, implemented using the hyperspectral anomaly detection system based on multi-feature intelligent optimization as described in claim 1, characterized in that, The hyperspectral anomaly detection method based on multi-feature intelligent optimization specifically includes the following steps: S1. Perform spectral band optimization on the original hyperspectral data and extract the optimal spectral band set from the original spectral data through cross-subspace joint extraction; S2. Perform multi-scale and multi-type attribute filtering and difference operations on each spectral band in the optimal spectral band set to extract the spatial spectral feature information of the abnormal target, suppress background information, and enhance the difference between the target and the background. S3. Use a genetic algorithm to classify each feature profile and intelligently select the most representative set of spatial spectral feature profiles. Step S3 specifically includes the following steps: S31. The normalized spatial spectral feature profiles are numbered and used as the encoding values ​​of each chromosome in the genetic algorithm. The chromosomes are encoded using real numbers, ranging from 1 to the total number q of the spatial spectral feature profiles. The number of chromosomes in an individual is the expected number of representative feature profiles n. S32. The ratio of the intra-class to inter-class dissimilarity coefficients when the combination of these profiles is used as the optimal feature profile is expressed by the fitness function, and its formula is as follows: (9)； Among them, f ci and f cj For different optimal feature profiles, f h For a member of class ci, C ci Then it is the number of members belonging to class ci, D NMI The coefficient of difference represents the difference between two feature profiles. The smaller the value, the greater the difference between the two feature profiles. The formula for calculating it is as follows: (10)； Among them, G i and G j f i and f j The set of gray values, g x and g y These are the individual grayscale values, p(g) x g y Then it is the joint probability mass function; S33. By performing selection, crossover, and mutation operations, the fitness function of each individual in the population is repeatedly calculated, and the feature profile combination with the smallest intra-class difference and the largest inter-class difference is intelligently selected, which is the optimal feature profile set representing the spatial-spectral characteristics of the hyperspectral data. S4. The most representative spatial spectral feature profiles selected by intelligent optimization are subjected to threshold weighted summation to filter out background information and highlight the target position, and anomaly detection results are generated jointly.

6. The hyperspectral anomaly detection method based on multi-feature intelligent optimization according to claim 5, characterized in that, Step S1 specifically includes the following steps: S11. Calculate the correlation between each spectral band using the correlation coefficient matrix, and divide the hyperspectral data into multiple subspaces; S12. Calculate the optimal index factor between each spectral segment in the same subspace and the two adjacent spectral segments, and find the spectral segment with the maximum information content and the minimum correlation to represent the subspace. S13. Extract the spectral segments representing each subspace to obtain the optimal set of spectral segments that can represent the original hyperspectral data.

7. The hyperspectral anomaly detection method based on multi-feature intelligent optimization according to claim 5, characterized in that, Step S2 specifically includes the following steps: S21. Select four attribute types: area, diagonal length, standard deviation, and moment of inertia to construct multiple spatial features and fully characterize the spatial structure features of hyperspectral data. S22. For different attribute types, multiple threshold parameters are used to describe the spatial information of anomalous targets at different scales, among which, The difference formula and threshold selection method for area attribute filters are as follows: (1)； (2)； in, and These represent the coarsening and thinning operations, respectively. i For the area attribute difference profile of the i-th spectral band, the threshold parameter of the area attribute is selected in relation to the spatial resolution. Related to the image size M×N, a max and a min These represent the upper and lower bounds of the area attribute threshold, respectively. Here, a... min =2 and δ A Set the step size to 1; The difference formula and threshold selection method for the diagonal length attribute filter are as follows: (3)； (4)； Among them, B i The diagonal length attribute difference profile is the i-th spectral band, and the threshold parameter of the diagonal length attribute is selected in relation to the spatial resolution. Related to the image size M×N, b max and b min These are the upper and lower bounds of the diagonal length attribute threshold, which are here... and δ B Set the step size to 2; The difference formula and threshold selection method for the standard deviation attribute filter are as follows: (5)； (6)； Among them, S i Let S be the standard deviation attribute difference profile of the i-th spectral band, where the threshold parameter of the standard deviation is selected based on the mean of the spectral band. max and S min These represent the upper and lower bounds of the standard deviation attribute threshold, respectively 25% and 5% of the mean, δ. s Set the step size to 2.5% of the mean. The difference formula and threshold selection method for the moment of inertia property filter are as follows: (7)； (8)； Among them, I i Let I be the difference profile of the moment of inertia for the i-th spectral band. max and I min These are the upper and lower bounds of the threshold value of the moment of inertia, respectively. Here, I is... max =0.9 and I min =0.1, δ I Set the step size to 0.1; S23. By normalizing the feature profiles of different scales and attributes obtained by the difference operation, it is easier to analyze and compare the differences between different feature profiles.

8. The hyperspectral anomaly detection method based on multi-feature intelligent optimization according to claim 5, characterized in that, Step S4 specifically includes the following steps: S41. The most representative spatial spectral feature profiles selected by intelligent optimization are subjected to thresholding. The feature profile threshold is calculated using the Otsu algorithm with maximum inter-class variance. The calculation formula is as follows: (11)； Where Thresh is the threshold calculation symbol, and θ is the feature profile threshold; S42. The detection result acquisition unit will determine the number of in-class members C of each selected feature profile in the genetic algorithm classification. ci The ratio of the number of spatial spectral feature profiles to the total number of feature profiles q is used as a weight. The optimal feature profiles after thresholding are then summed using a weighted average to obtain the final detection result. x and y are the horizontal and vertical coordinates of the image, respectively. The calculation formula for the detection result is as follows: (12)。

Images