Hyperspectral image noise detection and evaluation method and system

Abstract

A method and system for detecting and evaluating noise in a hyperspectral image relates to the technical field of image processing. The method comprises: first, performing superpixel segmentation on an original hyperspectral image using an improved linear iterative clustering method to obtain a plurality of superpixel sub-blocks; then, calculating a detrended cross-correlation coefficient and a detrended partial cross-correlation coefficient for the spectral reflectance sequence of each pairwise superpixel sub-block using a detrended partial cross-correlation method; simultaneously, obtaining a noise sequence eigenvalue range based on a randomly generated Gaussian white noise sequence, which is used as a reference for the noise range of a hyperspectral signal; finally, performing spectral decomposition on a correlation matrix constructed by the detrended partial cross-correlation coefficients of the pairwise superpixel sub-blocks to obtain eigenvalues ​​of the partial correlation matrix at each preset scale, and comparing the eigenvalue range of the noise sequence to obtain a noise ratio value and a detrended partial cross-correlation coefficient matrix after denoising; the application can optimize the accuracy of noise detection and evaluation of hyperspectral images.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method and system for detecting and evaluating hyperspectral image noise. Background Art

[0002] Hyperspectral images, with their hundreds of continuous narrow bands covering visible light to short-wave infrared, can accurately analyze the physical and chemical properties of target objects through material-specific spectral reflectance characteristics. However, their multi-band imaging characteristics make the data prone to serious mixed pixels and the frequent existence of the "different objects with the same spectrum" phenomenon. In addition, noise is inevitable in the process of hyperspectral signal acquisition and transmission. The noise detection and evaluation accuracy of hyperspectral images in existing technologies is poor, and it will trigger chain reactions such as spectral curve distortion and spatial texture blur, which seriously restrict the accuracy of ground object target recognition and ground object parameter inversion. Summary of the Invention

[0003] The problem solved by the present invention is how to optimize the accuracy of noise detection evaluation of hyperspectral images.

[0004] To solve the above problems, in a first aspect, the present invention provides a method for detecting and evaluating noise in a hyperspectral image, the method comprising:

[0005] The improved linear iterative clustering method is used to perform super-pixel segmentation on the initial hyperspectral image to obtain multiple super-pixel sub-blocks;

[0006] Calculate the detrended cross-correlation coefficient of the spectral reflectance sequences of any two superpixel sub-blocks to obtain a detrended cross-correlation coefficient matrix; the detrended cross-correlation coefficient matrix contains the detrended cross-correlation coefficients at multiple different preset scales, and the detrended cross-correlation coefficient is used to measure the correlation strength of the two spectral reflectance sequences at the same preset scale;

[0007] Based on the inverse matrix of the detrended cross-correlation coefficient matrix, first detrended partial cross-correlation coefficients of the spectral reflectances of any two superpixel sub-blocks are calculated to obtain a first detrended partial cross-correlation coefficient matrix; the first detrended partial cross-correlation coefficient matrix includes first detrended partial cross-correlation coefficients at multiple different preset scales, and the first detrended partial cross-correlation coefficients are used to evaluate the true correlation between the two spectral reflectances at the same preset scale;

[0008] Based on multiple randomly generated Gaussian white noise sequences, the second detrended partial cross-correlation coefficient of any two Gaussian white noise sequences is calculated to obtain a second detrended partial cross-correlation coefficient matrix; the second detrended partial cross-correlation coefficient is used to evaluate the true correlation between the two Gaussian white noise sequences at the same preset scale; the second detrended partial cross-correlation coefficient matrix is ​​subjected to spectral decomposition processing to obtain the eigenvalue ranges of the noise sequences at multiple different preset scales;

[0009] The first detrended partial cross-correlation coefficient matrix is ​​subjected to spectral decomposition, and the eigenvalue range of the noise sequence is used as a reference value to calculate the denoised target detrended partial cross-correlation coefficient matrix and the noise ratio of each superpixel sub-block for subsequent hyperspectral image analysis.

[0010] Optionally, the method of performing superpixel segmentation processing on the initial hyperspectral image using an improved linear iterative clustering method to obtain a plurality of superpixel sub-blocks includes:

[0011] Based on the original hyperspectral image and the first preset distance, a plurality of first cluster center points are obtained;

[0012] Taking the first cluster centers as the center and based on the second preset distance, searching for a neighborhood range of each first cluster center point;

[0013] Calculate the cluster distance between each neighboring pixel point in each neighborhood and the center point of the first cluster to which it belongs;

[0014] According to the cluster distance and minimum distance principle, the first cluster center point to which each adjacent pixel belongs is determined;

[0015] Based on the mean of all pixels belonging to the same first cluster center, the second cluster center is recalculated;

[0016] The second cluster center is used as the new first cluster center, and the target cluster center is obtained through multiple iterations.

[0017] Optionally, calculating the cluster distance between each adjacent pixel point in each neighborhood and the first cluster center point to which it belongs includes:

[0018] Based on the spectral vector of the first cluster center, the spectral vectors of the neighboring pixels, the spectral vector correlation coefficient and the Euclidean distance, the spectral distance between each neighboring pixel point in each neighborhood and the first cluster center point to which it belongs is obtained;

[0019] Based on the coordinates of the first cluster center and the coordinates of the neighboring pixels, the spatial distance between each neighboring pixel and the first cluster center to which it belongs is obtained;

[0020] The spectral distance and spatial distance are weighted to obtain the cluster distance between each neighboring pixel point in each neighborhood and the first cluster center point to which it belongs.

[0021] Optionally, defining the spectral reflectance sequences of any two superpixel sub-blocks as a first spectral reflectance sequence and a second spectral reflectance sequence;

[0022] The step of calculating the detrended mutual correlation coefficient of the spectral reflectance sequences of any two superpixel sub-blocks to obtain a detrended mutual correlation coefficient matrix includes:

[0023] Based on a plurality of preset scales, a first local trend sequence of the first spectral reflectance sequence and a second local trend sequence of the second spectral reflectance sequence are obtained;

[0024] Based on multiple preset scales, the first local trend sequence and the second local trend sequence, a first variance fluctuation function related to the first spectral reflectance sequence, a second variance fluctuation function related to the second spectral reflectance sequence, and a third covariance fluctuation function related to the first spectral reflectance sequence and the second spectral reflectance sequence are obtained;

[0025] Multiplying the first variance fluctuation function and the second variance fluctuation function and taking the square root of the result to obtain an intermediate operation result; calculating the ratio of the third covariance fluctuation function and the intermediate operation result to obtain the detrended mutual correlation coefficient under multiple different preset scales;

[0026] A correlation coefficient matrix is ​​constructed for the obtained detrended mutual correlation coefficients under multiple different preset scales to obtain a detrended mutual correlation coefficient matrix.

[0027] Optionally, the plurality of preset scales include a first preset scale;

[0028] The method of obtaining a first local trend sequence of the first spectral reflectance sequence and a second local trend sequence of the second spectral reflectance sequence based on multiple preset scales includes:

[0029] Calculating a first cumulative deviation sequence of the first spectral reflectance sequence and a second cumulative deviation sequence of the second spectral reflectance sequence;

[0030] Dividing the first cumulative deviation sequence and the second cumulative deviation sequence into a plurality of overlapping subintervals of a first preset scale length;

[0031] In each subinterval, the least square method is used to fit the data to obtain a local trend sequence, where the local trend sequence includes a first local trend sequence of the first spectral reflectance sequence and a second local trend sequence of the second spectral reflectance sequence.

[0032] Optionally, obtaining a first variance fluctuation function related to the first spectral reflectance sequence, a second variance fluctuation function related to the second spectral reflectance sequence, and a third covariance fluctuation function related to the first spectral reflectance sequence and the second spectral reflectance sequence based on multiple preset scales, the first local trend sequence, and the second local trend sequence includes:

[0033] Based on the first preset scale, the first cumulative deviation sequence and the first local trend sequence, a first variance fluctuation function related to the first spectral reflectance sequence is calculated;

[0034] Based on the first preset scale, the second cumulative deviation sequence and the second local trend sequence, a second variance fluctuation function related to the second spectral reflectance sequence is calculated;

[0035] Based on the first preset scale, the first cumulative deviation sequence, the first local trend sequence, the second cumulative deviation sequence and the second local trend sequence, a third covariance fluctuation function related to the first spectral reflectance sequence and the second spectral reflectance sequence is calculated.

[0036] Optionally, defining a spectral reflectivity sequence of any two superpixel sub-blocks as a first spectral reflectivity and a second spectral reflectivity;

[0037] The step of calculating the first detrended partial cross-correlation coefficient of the spectral reflectance of any two superpixel sub-blocks based on the inverse matrix of the detrended cross-correlation coefficient matrix to obtain the first detrended partial cross-correlation coefficient matrix includes:

[0038] Obtaining, from the inverse matrix of the detrended mutual correlation coefficient matrix, a first inverse matrix element related to the first spectral reflectance, a second inverse matrix element related to the second spectral reflectance, and a third inverse matrix element related to the first spectral reflectance and the second spectral reflectance at the same preset scale;

[0039] Based on the first inverse matrix element, the second inverse matrix element and the third inverse matrix element, a first detrended partial cross-correlation coefficient of the spectral reflectance of any two superpixel sub-blocks at a plurality of different preset scales is calculated;

[0040] A correlation coefficient matrix is ​​constructed for the obtained first detrended partial cross-correlation coefficients under multiple different preset scales to obtain a first detrended partial cross-correlation coefficient matrix.

[0041] Optionally, performing spectral decomposition on the first detrended partial cross-correlation coefficient matrix and calculating the denoised target detrended partial cross-correlation coefficient matrix and the noise ratio of each superpixel sub-block using the eigenvalue range of the noise sequence as a reference value includes:

[0042] Perform spectral decomposition on the first detrended partial cross-correlation coefficient matrix at different preset scales to obtain the eigenvalue of the partial correlation matrix at each preset scale;

[0043] Taking the eigenvalue range of the noise sequence as the reference value, the eigenvalues ​​and corresponding eigenvectors of the partial correlation matrix within the eigenvalue range of the noise sequence are removed to obtain the target detrended partial correlation coefficient matrix;

[0044] The ratio of the partial correlation matrix eigenvalues ​​that fall within the eigenvalue range of the noise sequence to all eigenvalues ​​is calculated to obtain the noise ratio of each superpixel sub-block.

[0045] In a second aspect, the present invention provides a hyperspectral image noise detection and evaluation system, the system comprising:

[0046] A superpixel segmentation module is used to perform superpixel segmentation on the initial hyperspectral image using an improved linear iterative clustering method to obtain multiple superpixel sub-blocks;

[0047] A detrended cross-correlation operation module is used to calculate the detrended cross-correlation coefficient of the spectral reflectance sequences of any two superpixel sub-blocks to obtain a detrended cross-correlation coefficient matrix; the detrended cross-correlation coefficient matrix contains detrended cross-correlation coefficients at multiple different preset scales, and the detrended cross-correlation coefficient is used to measure the correlation strength of two spectral reflectance sequences at the same preset scale;

[0048] a detrended partial cross-correlation operation module, configured to calculate a first detrended partial cross-correlation coefficient of the spectral reflectance of any two superpixel sub-blocks based on an inverse matrix of the detrended partial cross-correlation coefficient matrix, thereby obtaining a first detrended partial cross-correlation coefficient matrix; the first detrended partial cross-correlation coefficient matrix comprising a plurality of first detrended partial cross-correlation coefficients at different preset scales, the first detrended partial cross-correlation coefficient being used to evaluate the true correlation between the two spectral reflectances at the same preset scale;

[0049] A noise eigenvalue calculation module is used to calculate the second detrended partial cross-correlation coefficient of any two Gaussian white noise sequences based on multiple randomly generated Gaussian white noise sequences to obtain a second detrended partial cross-correlation coefficient matrix; the second detrended partial cross-correlation coefficient is used to evaluate the true correlation between the two Gaussian white noise sequences at the same preset scale; and the second detrended partial cross-correlation coefficient matrix is ​​subjected to spectral decomposition processing to obtain the eigenvalue range of the noise sequence at multiple different preset scales;

[0050] The comparison and denoising module is used to perform spectral decomposition on the first detrended partial cross-correlation coefficient matrix, and use the eigenvalue range of the noise sequence as a reference value to calculate the denoised target detrended partial cross-correlation coefficient matrix and the noise ratio of each superpixel sub-block for subsequent hyperspectral image analysis.

[0051] In a third aspect, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the hyperspectral image noise detection and evaluation method as described above is implemented.

[0052] The beneficial effects of the present invention are:

[0053] Compared with the prior art, the hyperspectral image noise detection and evaluation method and system provided by the present invention first perform dimensionality reduction processing on the original hyperspectral image, and use the improved linear iterative clustering method to perform superpixel segmentation on the original hyperspectral image to obtain multiple superpixel sub-blocks; then, considering the non-stationary characteristics of the spectral reflectance sequence, the spectral reflectance sequence of each pair of superpixel sub-blocks is calculated using the detrended partial cross-correlation method to calculate the detrended cross-correlation coefficient and the detrended partial cross-correlation coefficient, eliminating the influence of the spectral reflectance of other superpixel sub-blocks on the target superpixel sub-block, and obtaining the correlation coefficients under different preset scales to truly reflect the correlation between the two superpixel sub-blocks, so that the correlation between the two superpixel sub-blocks is more accurate, which can avoid the problem that the multivariate linear regression cannot truly fit the correlation between the superpixel sub-blocks, resulting in noise distortion in the detection, thereby improving the noise detection and evaluation. Accuracy estimation; at the same time, according to the randomly generated Gaussian white noise sequence with the same length as the spectral reflectance sequence, the detrended partial cross-correlation method is used to calculate the detrended partial cross-correlation coefficient between the two Gaussian white noise sequences, and a detrended partial cross-correlation coefficient matrix is ​​constructed. It is further spectrally decomposed to obtain the eigenvalue range of the noise sequence, which is used as a reference for the noise range of the hyperspectral signal; finally, the correlation matrix constructed by the detrended partial cross-correlation coefficient of the two superpixel sub-blocks is spectrally decomposed to obtain the eigenvalue of the partial correlation matrix under each preset scale. Compared with the eigenvalue range of the noise sequence, the part within the range is regarded as noise, and the part outside the range is regarded as information. Then, the noise ratio value and the detrended partial cross-correlation coefficient matrix after denoising are calculated, which are used for subsequent superpixel sub-block analysis of the hyperspectral image to improve the accuracy of subsequent hyperspectral image target recognition and inversion. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 FIG2 is a flow chart of a method for detecting and evaluating hyperspectral image noise according to an embodiment of the present invention;

[0055] Figure 2 FIG2 shows a schematic diagram of the structure of a hyperspectral image noise detection and evaluation system according to an embodiment of the present application. DETAILED DESCRIPTION

[0056] The present invention is described in detail below with reference to specific embodiments.

[0057] like Figure 1 As shown, the present invention provides a method for detecting and evaluating noise in a hyperspectral image, the method comprising:

[0058] Step S100 : performing superpixel segmentation processing on the initial hyperspectral image using an improved linear iterative clustering method to obtain a plurality of superpixel sub-blocks.

[0059] Specifically, an original hyperspectral image I is obtained in advance. Assuming that the original hyperspectral image I is imaging hyperspectral data with spatial pixels of A×B and the number of bands N, an improved linear iterative clustering method (Simple Linear Iterative Clustering, SLIC) is used to initialize the cluster center, calculate the cluster distance between each pixel and the cluster center, and update the cluster center. Then, the original hyperspectral image I is superpixel segmented to obtain K superpixel sub-blocks, where K is a positive integer. Furthermore, in this embodiment, the improved linear iterative clustering method described in step S100 includes the following steps S110 to S160:

[0060] Step S110: Based on the original hyperspectral image and the first preset distance, a plurality of first cluster center points are obtained. Specifically, assuming that the step size is The distance (i.e., the first preset distance) is used to divide superpixels, and the K Initial cluster centers C i = [ v i , x i , y i ] (i.e., the first cluster center), where 1≤ i ≤ K, A·B is the number of pixels.

[0061] Step S120: With the first cluster centers as the centers, based on the second preset distance, a neighborhood range of each first cluster center is searched. For example, for each cluster center, a 2S×2S neighborhood range of the cluster center is searched.

[0062] Step S130 calculates the cluster distance between each neighboring pixel in each neighborhood and the center of the first cluster to which it belongs. Because hyperspectral images differ from ordinary natural images, cluster distances are considered from both the spectral and spatial dimensions of hyperspectral images. After calculating the spectral and spatial distances, these two distances are weighted to obtain the final cluster distance.

[0063] In this embodiment, the method for calculating spectral distance includes step S131, which obtains the spectral distance between each neighboring pixel point in each neighborhood and the first cluster center point to which it belongs based on the spectral vector of the first cluster center point, the spectral vectors of the neighboring pixels, the spectral vector correlation coefficient, and the Euclidean distance. Specifically, the spectral distance can be defined as:

[0064]

[0065] in, vi is the spectral vector of the cluster center pixel, v j is the spectral vector of the neighboring pixels, yes v i and v j The correlation coefficient of is the Euclidean distance between these two vectors.

[0066] The method for calculating the spatial distance includes step S132, which obtains the spatial distance between each neighboring pixel point and the first cluster center point to which it belongs, based on the coordinates of the first cluster center point and the coordinates of the neighboring pixel points. Specifically, the spatial distance can be defined as:

[0067]

[0068] in, I i =( x i ，y i ) is the coordinate of the cluster center pixel, I j =( x j ，y j ) are the coordinates of its neighboring pixels.

[0069] After obtaining the spectral distance and spatial distance in steps S131 and S132, the spectral distance and spatial distance are weighted in step S133 to obtain the cluster distance between each neighboring pixel point in each neighborhood and the first cluster center point to which it belongs. The cluster distance can be calculated as follows:

[0070] ;

[0071] in, α is the spatial distance weight factor, d spec is the spectral distance, d spat is the spatial distance, .

[0072] In step S140, the first cluster center to which each neighboring pixel point belongs is determined based on the cluster distance and the minimum distance principle; that is, after the cluster distance D from each pixel point to its cluster center is calculated within the neighborhood of each cluster center, the cluster center to which it should belong is determined based on the minimum distance principle.

[0073] In step S150, the second cluster center is recalculated based on the mean of all pixels belonging to the same first cluster center. The positions and spectra of all pixels in the neighborhood of the same cluster center are averaged after step S140 to recalculate the cluster center.

[0074] In step S160, the second cluster center point is used as the new first cluster center point, and the target cluster center point is obtained through multiple iterations; that is, the second cluster center point is used as the new first cluster center point, and steps S120 to S150 are repeated until convergence (such as the change is less than the threshold or the cluster center no longer moves).

[0075] It should be noted that in this embodiment, the linear iterative clustering method is improved to fully consider the spectral dimension and spatial dimension information. The original hyperspectral image I is segmented into K superpixel sub-blocks through the improved linear iterative clustering algorithm, which is beneficial to improving the efficiency and accuracy of subsequent analysis.

[0076] Step S200, calculate the detrended cross-correlation coefficient of the spectral reflectance sequence of any two superpixel sub-blocks to obtain a detrended cross-correlation coefficient matrix; the detrended cross-correlation coefficient matrix includes detrended cross-correlation coefficients at multiple different preset scales, and the detrended cross-correlation coefficient is used to measure the strength of the correlation between two spectral reflectance sequences at the same preset scale.

[0077] Specifically, based on the K superpixel sub-blocks obtained in step S100, any two superpixel sub-blocks are selected, and the spectral reflectance sequence of the any two superpixel sub-blocks is defined as the first spectral reflectance sequence { x ( i )} and the second spectral reflectance sequence { y ( i )}, multiple different preset scales include a first preset scale s， Can be 6≤ s ≤ N / 4. In the following steps, in addition to considering the correlation between the center pixel and the area pixels, the correlation between the area pixels is also considered, which can make the expressed correlation more accurate. Furthermore, in this embodiment, step S200 includes the following steps S210 to S240:

[0078] Step S210: Based on multiple preset scales, a first local trend sequence of the first spectral reflectance sequence and a second local trend sequence of the second spectral reflectance sequence are obtained. x ( i )}, the second spectral reflectance sequence { y ( i )} and the first preset scale s as an example, first, the first spectral reflectance sequence { x (i )} the first cumulative deviation sequence X k and the second spectral reflectance sequence { y ( i )} the second cumulative deviation sequence Y k : After that, the first cumulative deviation sequence X k and the second cumulative deviation sequence Y k Divide into equal parts Ns The length of the overlapping (multiple) equal length is s +1 (first preset scale length); then, in each subinterval, the least squares method is used to fit the data to obtain the local trend sequence and , that is, the first local trend sequence of the first spectral reflectance sequence is obtained by fitting the data using the linear least squares method and the second local trend series of the second spectral reflectance series .

[0079] Step S220: Based on the multiple preset scales, the first local trend sequence, and the second local trend sequence, a first variance fluctuation function associated with the first spectral reflectance sequence, a second variance fluctuation function associated with the second spectral reflectance sequence, and a third covariance fluctuation function associated with the first spectral reflectance sequence and the second spectral reflectance sequence are obtained. Specifically, after step S210, the covariance of each subinterval after trend removal is calculated: ; and after averaging each window, we get the covariance fluctuation function: ;when x=y When the average cross-covariance fluctuation function Degenerates into a single 0 sequence variance function. That is, based on the first preset scale s, the first cumulative deviation sequence X k and the first local trend series , the first variance fluctuation function related to the first spectral reflectance sequence is calculated as: Based on the first preset scale s, the second cumulative deviation sequence Y k and the second local trend series , the second variance fluctuation function related to the second spectral reflectance sequence is calculated as: Based on the first preset scale s, the first cumulative deviation sequence X k , the first local trend sequence , the second cumulative deviation sequence Yk and the second local trend series , the third covariance fluctuation function related to the first spectral reflectance sequence and the second spectral reflectance sequence is calculated as: .

[0080] In step S230, the first variance fluctuation function and the second variance fluctuation function are multiplied and the square root of the result is taken to obtain an intermediate calculation result. The ratio of the third covariance fluctuation function and the intermediate calculation result is calculated to obtain the detrended cross-correlation coefficient under multiple different preset scales. In other words, the detrended cross-correlation coefficient defined in this embodiment is: .

[0081] Step S240 , constructing a correlation coefficient matrix for the obtained detrended mutual correlation coefficients at multiple different preset scales to obtain a detrended mutual correlation coefficient matrix.

[0082] It should be noted that the detrended correlation coefficient of this embodiment is r DCCA ( s ) has two advantages: (1) due to the detrending step in each interval, it can characterize the correlation between two non-stationary series, fully considering the non-stationary characteristics of the spectral reflectance series; (2) it can describe the correlation between the two series at different preset scales; that is, the detrended mutual correlation coefficient r DCCA ( s ) can be measured on a preset scale of s The spectral reflectance sequence { x ( i )}and{ y ( i )}, satisfying -1≤ r ( s )≤1.

[0083] Step S300, based on the inverse matrix of the detrended mutual correlation coefficient matrix, calculate the first detrended partial mutual correlation coefficient of the spectral reflectance of any two superpixel sub-blocks to obtain a first detrended partial mutual correlation coefficient matrix; the first detrended partial mutual correlation coefficient matrix includes a plurality of first detrended partial mutual correlation coefficients at different preset scales, and the first detrended partial mutual correlation coefficient is used to evaluate the true correlation between the two spectral reflectances at the same preset scale.

[0084] It should be noted that the detrended correlation coefficient in step S200 r DCCA ( s ) can characterize the correlation between superpixel sub-blocks, but it does not consider the situation where the correlation between local sub-blocks and other sub-blocks is exaggerated due to the collinearity phenomenon between pixels. r DCCA ( s ) is used to define the detrended partial mutual correlation coefficient, taking into full account the true correlation between the spatial dimension and the spectral dimension of each uniform sub-block, and excluding the average spectral reflectance sequence of the target sub-block from other sub-blocks { x ( i )}and{ y ( i )}Interference of correlation.

[0085] In this embodiment, the partial correlation technique is used to r DCCA ( s ) is defined as follows:

[0086] ;

[0087] in, C ( s ) is r DCCA ( s ) is the element in the inverse matrix of the correlation coefficient matrix, that is: ;

[0088] above, r DPCCA ( s ) can be used to evaluate the spectral reflectance of the two sub-blocks z u ( i )and z v ( t ) in the preset scale of s It can be understood that in this embodiment, the spectral reflectivity sequence of the superpixel sub-block is a set of reflectivity values ​​of all pixels in the superpixel sub-block in multiple bands, and the spectral reflectivity of the superpixel sub-block is the reflectivity value corresponding to all pixels in the superpixel sub-block in a certain band. In other words, the spectral reflectivity of the superpixel sub-block is the average of the spectral reflectivities of all pixels in the sub-block. The spectral reflectivity of the superpixel sub-block is associated with the spectral reflectivity sequence of the superpixel sub-block. The method for generating the spectral reflectivity sequence and spectral reflectivity of the superpixel sub-block can be found in the prior art and will not be elaborated here.

[0089] According to the expression of the detrended partial cross-correlation coefficient, step S300 can be divided into the following steps: first, from the inverse matrix of the detrended cross-correlation coefficient matrix, obtain the first inverse matrix element related to the first spectral reflectance, the second inverse matrix element related to the second spectral reflectance, and the third inverse matrix element related to the first spectral reflectance and the second spectral reflectance at the same preset scale; then, based on the first inverse matrix element, the second inverse matrix element and the third inverse matrix element, calculate the first detrended partial cross-correlation coefficient of the spectral reflectance of any two super-pixel sub-blocks at multiple different preset scales; finally, the obtained first detrended partial cross-correlation coefficients at multiple different preset scales are r DPCCA ( s ) Construct the correlation coefficient matrix and obtain the first detrended partial correlation coefficient matrix R ( s ).

[0090] Step S400: Calculate the second detrended partial cross-correlation coefficient of any two Gaussian white noise sequences based on a plurality of randomly generated Gaussian white noise sequences to obtain a second detrended partial cross-correlation coefficient matrix; the second detrended partial cross-correlation coefficient is used to evaluate the true correlation between the two Gaussian white noise sequences at the same preset scale; perform spectral decomposition on the second detrended partial cross-correlation coefficient matrix to obtain the eigenvalue ranges of the noise sequences at a plurality of different preset scales.

[0091] For example: According to Random Matrix Theory (RMT), computer generated K The length is N Gaussian white noise sequence n k ( i ), k =1,2,…, K ; i =1,2,…, N ; Calculate the detrended partial cross-correlation coefficient of each pair of Gaussian white noise series r DPCCA ( n k , n v , s ). At each preset scale s Next, add the above r DPCCA ( n k , n v , s ) is constructed as K × KThe partial correlation coefficient matrix (i.e. the second detrended partial mutual correlation coefficient matrix) is then spectrally decomposed to obtain its minimum and maximum eigenvalues, respectively. and , the remaining eigenvalues ​​satisfy ; This range serves as a reference for the eigenvalue range obtained in the spectral decomposition of the partial correlation coefficient matrix of the superpixel sub-block of the hyperspectral image. The part falling into this range is considered to be noise, while the part outside this range is considered to be information.

[0092] Step S500, performs spectral decomposition processing on the first detrended partial cross-correlation coefficient matrix, and uses the eigenvalue range of the noise sequence as a reference value to calculate the target detrended partial cross-correlation coefficient matrix after denoising and the noise ratio of each super-pixel sub-block, so as to subsequently perform hyperspectral image analysis. Specifically, step S500 in this embodiment includes: performing spectral decomposition processing on the first detrended partial cross-correlation coefficient matrix at different preset scales to obtain the eigenvalues ​​of the partial correlation matrix at each preset scale; using the eigenvalue range of the noise sequence as a reference value, removing the eigenvalues ​​and corresponding eigenvectors of the partial correlation matrix within the eigenvalue range of the noise sequence to obtain the target detrended partial cross-correlation coefficient matrix; calculating the ratio of the eigenvalues ​​of the partial correlation matrix that fall within the eigenvalue range of the noise sequence to all eigenvalues ​​to obtain the noise ratio of each super-pixel sub-block.

[0093] The following example illustrates the partial correlation coefficient of the detrended r DPCCA ( s ) is the detrended partial cross-correlation matrix R ( s ) at each preset scale s Perform spectral decomposition on the preset scale s The relevant eigenvalues ​​are l 1( s )≤ l 2( s )≤…≤ l K ( s ).then ,in, P ( s ) is the scale of s The characteristic vector of Λ( s ) is the scale s The diagonal matrix of , whose diagonal elements are l 1( s ), l 2( s ),…, l M ( s Then, replace Λ( s) part of the elements, namely:

[0094] ;

[0095] in, and , here and They represent the detrended partial cross-correlation matrix calculated based on the Gaussian white noise sequence R ( s )'s minimum and maximum eigenvalues.

[0096] Detrended partial cross-correlation matrix R ( s ) denoising, the target detrended partial correlation coefficient matrix after denoising is obtained as: It is understandable that only the eigenvectors corresponding to the eigenvalues ​​that are not replaced by 0 are retained in P ( s ) * In order to ensure that the trace of the partial correlation coefficient matrix does not change before and after the transformation, adjust R ( s ) * The diagonal elements of are 1, R ( s ) * It is the correlation coefficient matrix for subsequent hyperspectral image superpixel sub-block analysis.

[0097] According to the above s ) * , we can quantitatively evaluate the noise of the hyperspectral sub-block image and obtain the noise ratio of each super-pixel sub-block. Noi is the noise ratio of the sub-block image, Inf is the information ratio of the sub-block image, then we get

[0098]

[0099] in Indicates that for all scales l ( s )Eigenvalues ​​are averaged;#{ T} represents a collection T Number of elements Inf =1- Noi .

[0100] As described above, the hyperspectral image noise detection and evaluation method provided by the embodiment of the present application first performs dimensionality reduction processing on the original hyperspectral image, and uses the improved linear iterative clustering method to perform superpixel segmentation on the original hyperspectral image to obtain multiple superpixel sub-blocks; then, considering the non-stationary characteristics of the spectral reflectance sequence, the detrended partial cross-correlation method is used to calculate the detrended cross-correlation coefficient and the detrended partial cross-correlation coefficient for the spectral reflectance sequence of each pair of superpixel sub-blocks, eliminating the influence of the spectral reflectance of other superpixel sub-blocks on the target superpixel sub-block, and obtaining the correlation coefficients under different preset scales to truly reflect the correlation between the two superpixel sub-blocks, so that the correlation between the two superpixel sub-blocks is more accurate, which can avoid the problem of noise distortion caused by the inability to truly fit the correlation between superpixel sub-blocks by using multiple linear regression, thereby improving the accuracy of noise detection evaluation. At the same time, according to the randomly generated Gaussian white noise sequence with the same length as the spectral reflectance sequence, the detrended partial cross-correlation method is used to calculate the detrended partial cross-correlation coefficient between the two Gaussian white noise sequences, and a detrended partial cross-correlation coefficient matrix is ​​constructed. The spectral decomposition is further performed to obtain the eigenvalue range of the noise sequence, which is used as a reference for the noise range of the hyperspectral signal. Finally, the correlation matrix constructed by the detrended partial cross-correlation coefficient of the two super-pixel sub-blocks is spectrally decomposed to obtain the eigenvalue of the partial correlation matrix under each preset scale. Compared with the eigenvalue range of the noise sequence, the part within the range is regarded as noise, and the part outside the range is regarded as information. Then, the noise ratio value and the detrended partial cross-correlation coefficient matrix after denoising are calculated, which are used for subsequent super-pixel sub-block analysis of the hyperspectral image to improve the accuracy of subsequent hyperspectral image target recognition and inversion.

[0101] Second, as Figure 2 As shown, the embodiment of the present application also provides a hyperspectral image noise detection and evaluation system, which includes a superpixel segmentation module, a detrending cross-correlation operation module, a detrending partial cross-correlation operation module, a noise eigenvalue operation module and a comparison and denoising module, wherein:

[0102] A superpixel segmentation module is used to perform superpixel segmentation on the initial hyperspectral image using an improved linear iterative clustering method to obtain multiple superpixel sub-blocks;

[0103] A detrended cross-correlation operation module is used to calculate the detrended cross-correlation coefficient of the spectral reflectance sequences of any two superpixel sub-blocks to obtain a detrended cross-correlation coefficient matrix; the detrended cross-correlation coefficient matrix contains detrended cross-correlation coefficients at multiple different preset scales, and the detrended cross-correlation coefficient is used to measure the correlation strength of two spectral reflectance sequences at the same preset scale;

[0104] a detrended partial cross-correlation operation module, configured to calculate a first detrended partial cross-correlation coefficient of the spectral reflectance of any two superpixel sub-blocks based on an inverse matrix of the detrended partial cross-correlation coefficient matrix, thereby obtaining a first detrended partial cross-correlation coefficient matrix; the first detrended partial cross-correlation coefficient matrix comprising a plurality of first detrended partial cross-correlation coefficients at different preset scales, the first detrended partial cross-correlation coefficient being used to evaluate the true correlation between the two spectral reflectances at the same preset scale;

[0105] A noise eigenvalue calculation module is used to calculate the second detrended partial cross-correlation coefficient of any two Gaussian white noise sequences based on multiple randomly generated Gaussian white noise sequences to obtain a second detrended partial cross-correlation coefficient matrix; the second detrended partial cross-correlation coefficient is used to evaluate the true correlation between the two Gaussian white noise sequences at the same preset scale; and the second detrended partial cross-correlation coefficient matrix is ​​subjected to spectral decomposition processing to obtain the eigenvalue range of the noise sequence at multiple different preset scales;

[0106] The comparison and denoising module is used to perform spectral decomposition on the first detrended partial cross-correlation coefficient matrix, and use the eigenvalue range of the noise sequence as a reference value to calculate the denoised target detrended partial cross-correlation coefficient matrix and the noise ratio of each superpixel sub-block for subsequent hyperspectral image analysis.

[0107] It should be noted that the hyperspectral image noise detection and evaluation system of this embodiment can execute the hyperspectral image noise detection and evaluation method provided in this embodiment, and its implementation principle is similar; the actions performed by each module and unit in the hyperspectral image noise detection and evaluation system of this embodiment correspond to the steps in the hyperspectral image noise detection and evaluation method of this embodiment. For the detailed functional description of each module of the hyperspectral image noise detection and evaluation system, please refer to the description in the corresponding hyperspectral image noise detection and evaluation method shown in the previous text, which will not be repeated here.

[0108] In a third aspect, an embodiment of the present application further provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the hyperspectral image noise detection and evaluation method as described above is implemented. The specific steps are not repeated here. Please refer to the description of the above embodiment for details.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A method for detecting and evaluating noise in hyperspectral images, characterized in that: The method comprises: The improved linear iterative clustering method is used to perform super-pixel segmentation on the initial hyperspectral image to obtain multiple super-pixel sub-blocks; Calculate the detrended cross-correlation coefficient of the spectral reflectance sequences of any two superpixel sub-blocks to obtain a detrended cross-correlation coefficient matrix; the detrended cross-correlation coefficient matrix contains the detrended cross-correlation coefficients at multiple different preset scales, and the detrended cross-correlation coefficient is used to measure the correlation strength of the two spectral reflectance sequences at the same preset scale; Based on the inverse matrix of the detrended cross-correlation coefficient matrix, first detrended partial cross-correlation coefficients of the spectral reflectances of any two superpixel sub-blocks are calculated to obtain a first detrended partial cross-correlation coefficient matrix; the first detrended partial cross-correlation coefficient matrix includes first detrended partial cross-correlation coefficients at multiple different preset scales, and the first detrended partial cross-correlation coefficients are used to evaluate the true correlation between the two spectral reflectances at the same preset scale; Based on multiple randomly generated Gaussian white noise sequences, the second detrended partial cross-correlation coefficient of any two Gaussian white noise sequences is calculated to obtain a second detrended partial cross-correlation coefficient matrix; the second detrended partial cross-correlation coefficient is used to evaluate the true correlation between the two Gaussian white noise sequences at the same preset scale; the second detrended partial cross-correlation coefficient matrix is ​​subjected to spectral decomposition processing to obtain the eigenvalue ranges of the noise sequences at multiple different preset scales; The first detrended partial cross-correlation coefficient matrix is ​​subjected to spectral decomposition, and the eigenvalue range of the noise sequence is used as a reference value to calculate the denoised target detrended partial cross-correlation coefficient matrix and the noise ratio of each superpixel sub-block for subsequent hyperspectral image analysis.

2. The method for detecting and evaluating hyperspectral image noise according to claim 1, wherein: The improved linear iterative clustering method is used to perform super-pixel segmentation processing on the initial hyperspectral image to obtain a linear iterative clustering method in multiple super-pixel sub-blocks, including: Based on the original hyperspectral image and the first preset distance, a plurality of first cluster center points are obtained; Taking the first cluster centers as the center and based on the second preset distance, searching for a neighborhood range of each first cluster center point; Calculate the cluster distance between each neighboring pixel point in each neighborhood and the center point of the first cluster to which it belongs; According to the cluster distance and minimum distance principle, the first cluster center point to which each adjacent pixel belongs is determined; Based on the mean of all pixels belonging to the same first cluster center, the second cluster center is recalculated; The second cluster center is used as the new first cluster center, and the target cluster center is obtained through multiple iterations.

3. The method for detecting and evaluating hyperspectral image noise according to claim 2, wherein: The calculating of the cluster distance between each adjacent pixel point in each neighborhood and the first cluster center point to which it belongs includes: Based on the spectral vector of the first cluster center, the spectral vectors of the neighboring pixels, the spectral vector correlation coefficient and the Euclidean distance, the spectral distance between each neighboring pixel point in each neighborhood and the first cluster center point to which it belongs is obtained; Based on the coordinates of the first cluster center and the coordinates of the neighboring pixels, the spatial distance between each neighboring pixel and the first cluster center to which it belongs is obtained; The spectral distance and spatial distance are weighted to obtain the cluster distance between each neighboring pixel point in each neighborhood and the first cluster center point to which it belongs.

4. The method for detecting and evaluating hyperspectral image noise according to claim 1, wherein: defining the spectral reflectance sequences of any two superpixel sub-blocks as a first spectral reflectance sequence and a second spectral reflectance sequence; The step of calculating the detrended mutual correlation coefficient of the spectral reflectance sequences of any two superpixel sub-blocks to obtain a detrended mutual correlation coefficient matrix includes: Based on a plurality of preset scales, a first local trend sequence of the first spectral reflectance sequence and a second local trend sequence of the second spectral reflectance sequence are obtained; Based on multiple preset scales, the first local trend sequence and the second local trend sequence, a first variance fluctuation function related to the first spectral reflectance sequence, a second variance fluctuation function related to the second spectral reflectance sequence, and a third covariance fluctuation function related to the first spectral reflectance sequence and the second spectral reflectance sequence are obtained; Multiplying the first variance fluctuation function and the second variance fluctuation function and taking the square root of the result to obtain an intermediate operation result; calculating the ratio of the third covariance fluctuation function and the intermediate operation result to obtain the detrended mutual correlation coefficient under multiple different preset scales; A correlation coefficient matrix is ​​constructed for the obtained detrended mutual correlation coefficients under multiple different preset scales to obtain a detrended mutual correlation coefficient matrix.

5. The method for detecting and evaluating hyperspectral image noise according to claim 4, wherein: The plurality of preset scales include a first preset scale; The method of obtaining a first local trend sequence of the first spectral reflectance sequence and a second local trend sequence of the second spectral reflectance sequence based on multiple preset scales includes: Calculating a first cumulative deviation sequence of the first spectral reflectance sequence and a second cumulative deviation sequence of the second spectral reflectance sequence; Dividing the first cumulative deviation sequence and the second cumulative deviation sequence into a plurality of overlapping subintervals of a first preset scale length; In each subinterval, the least square method is used to fit the data to obtain a local trend sequence, where the local trend sequence includes a first local trend sequence of the first spectral reflectance sequence and a second local trend sequence of the second spectral reflectance sequence.

6. The method for detecting and evaluating hyperspectral image noise according to claim 5, wherein: The method of obtaining, based on multiple preset scales, the first local trend sequence and the second local trend sequence, a first variance fluctuation function related to the first spectral reflectance sequence, a second variance fluctuation function related to the second spectral reflectance sequence, and a third covariance fluctuation function related to the first spectral reflectance sequence and the second spectral reflectance sequence, comprises: Based on the first preset scale, the first cumulative deviation sequence and the first local trend sequence, a first variance fluctuation function related to the first spectral reflectance sequence is calculated; Based on the first preset scale, the second cumulative deviation sequence and the second local trend sequence, a second variance fluctuation function related to the second spectral reflectance sequence is calculated; Based on the first preset scale, the first cumulative deviation sequence, the first local trend sequence, the second cumulative deviation sequence and the second local trend sequence, a third covariance fluctuation function related to the first spectral reflectance sequence and the second spectral reflectance sequence is calculated.

7. The method for detecting and evaluating hyperspectral image noise according to claim 1, wherein: Define the spectral reflectance of any two superpixel sub-blocks as a first spectral reflectance and a second spectral reflectance; The step of calculating the first detrended partial cross-correlation coefficient of the spectral reflectance of any two superpixel sub-blocks based on the inverse matrix of the detrended cross-correlation coefficient matrix to obtain the first detrended partial cross-correlation coefficient matrix includes: Obtaining, from the inverse matrix of the detrended mutual correlation coefficient matrix, a first inverse matrix element related to the first spectral reflectance, a second inverse matrix element related to the second spectral reflectance, and a third inverse matrix element related to the first spectral reflectance and the second spectral reflectance at the same preset scale; Based on the first inverse matrix element, the second inverse matrix element and the third inverse matrix element, a first detrended partial cross-correlation coefficient of the spectral reflectance of any two superpixel sub-blocks at a plurality of different preset scales is calculated; A correlation coefficient matrix is ​​constructed for the obtained first detrended partial cross-correlation coefficients under multiple different preset scales to obtain a first detrended partial cross-correlation coefficient matrix.

8. The method for detecting and evaluating hyperspectral image noise according to claim 7, wherein: The spectral decomposition process is performed on the first detrended partial cross-correlation coefficient matrix, and the eigenvalue range of the noise sequence is used as a reference value to calculate the denoised target detrended partial cross-correlation coefficient matrix and the noise ratio of each superpixel sub-block, including: Perform spectral decomposition on the first detrended partial cross-correlation coefficient matrix at different preset scales to obtain the eigenvalue of the partial correlation matrix at each preset scale; Taking the eigenvalue range of the noise sequence as the reference value, the eigenvalues ​​and corresponding eigenvectors of the partial correlation matrix within the eigenvalue range of the noise sequence are removed to obtain the target detrended partial correlation coefficient matrix; The ratio of the partial correlation matrix eigenvalues ​​that fall within the eigenvalue range of the noise sequence to all eigenvalues ​​is calculated to obtain the noise ratio of each superpixel sub-block.

9. Hyperspectral image noise detection and evaluation system, characterized in that: The system comprises: A superpixel segmentation module is used to perform superpixel segmentation on the initial hyperspectral image using an improved linear iterative clustering method to obtain multiple superpixel sub-blocks; A detrended cross-correlation operation module is used to calculate the detrended cross-correlation coefficient of the spectral reflectance sequences of any two superpixel sub-blocks to obtain a detrended cross-correlation coefficient matrix; the detrended cross-correlation coefficient matrix contains detrended cross-correlation coefficients at multiple different preset scales, and the detrended cross-correlation coefficient is used to measure the correlation strength of two spectral reflectance sequences at the same preset scale; a detrended partial cross-correlation operation module, configured to calculate a first detrended partial cross-correlation coefficient of the spectral reflectance of any two superpixel sub-blocks based on an inverse matrix of the detrended partial cross-correlation coefficient matrix, thereby obtaining a first detrended partial cross-correlation coefficient matrix; the first detrended partial cross-correlation coefficient matrix comprising a plurality of first detrended partial cross-correlation coefficients at different preset scales, the first detrended partial cross-correlation coefficient being used to evaluate the true correlation between the two spectral reflectances at the same preset scale; A noise eigenvalue calculation module is used to calculate the second detrended partial cross-correlation coefficient of any two Gaussian white noise sequences based on multiple randomly generated Gaussian white noise sequences to obtain a second detrended partial cross-correlation coefficient matrix; the second detrended partial cross-correlation coefficient is used to evaluate the true correlation between the two Gaussian white noise sequences at the same preset scale; and the second detrended partial cross-correlation coefficient matrix is ​​subjected to spectral decomposition processing to obtain the eigenvalue range of the noise sequence at multiple different preset scales; The comparison and denoising module is used to perform spectral decomposition on the first detrended partial cross-correlation coefficient matrix, and use the eigenvalue range of the noise sequence as a reference value to calculate the denoised target detrended partial cross-correlation coefficient matrix and the noise ratio of each superpixel sub-block for subsequent hyperspectral image analysis.

10. An electronic device, characterized in that The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for detecting and evaluating hyperspectral image noise according to any one of claims 1 to 8 is implemented.

Images