Hyperspectral image noise label detection method and system, and medium

Abstract

The invention relates to the technical field of hyperspectral remote sensing image processing, in particular to a hyperspectral image noise label detection method, which comprises the following steps of: 1, preprocessing a source hyperspectral image; 2, constructing an isolated forest for the preprocessed hyperspectral image by using a dual-channel isolated forest algorithm to obtain a pixel abnormal probability of each pixel; 3, optimizing the pixel abnormal probability of each pixel to obtain an optimized noise label training sample set; and 4, removing the optimized noise label training sample set from the training sample set, and inputting the noise label training sample set into the trained classifier to obtain a final identification result. According to the invention, through a dual-channel isolated forest model, global and local spatial spectrum information is fully utilized, and an abnormal sample set can be accurately detected; in addition, the abnormal sample set is optimized through a clustering algorithm and the image block distance in sequence, a more accurate noise label sample set is obtained, and a guarantee is provided for a high-precision classification task.

Description

Technical Field

[0001] The present invention relates to the technical field of hyperspectral remote sensing image processing, and in particular to a method, system and medium for detecting hyperspectral image noise labels. Background Art

[0002] With the rapid development of remote sensing imaging technology, hyperspectral images have achieved the simultaneous acquisition of spectral reflectance and spatial information of ground objects, thus significantly improving the quality of earth observation. Therefore, hyperspectral images play a vital role in many fields such as agriculture, environmental monitoring and urban planning. Hyperspectral image classification technology aims to assign a unique label to each pixel in the image. This technology has received widespread attention in recent decades, further strengthening its application and research in various fields.

[0003] Although many state-of-the-art hyperspectral image classification algorithms have achieved satisfactory results, their success is largely attributed to the supervision information provided by a large number of labels. These labels are usually considered to be completely accurate, thus providing reliable training data for the algorithm. However, when acquiring images in various practical applications, they are often affected by factors such as atmospheric interference and manual labeling errors, and some mislabeled samples are inevitable. This means that some of the labels used in the training phase are not correctly labeled, and these mislabeled samples are called noisy labels. In recent years, researchers have proposed a series of solutions to solve and optimize noisy labels in the field of traditional natural images. However, due to the high dimensionality and spectral confusion problems of hyperspectral image data, the above methods are difficult to directly transplant to the field of hyperspectral image noisy label detection. Summary of the invention

[0004] The present invention provides a method, system and medium for detecting noise labels in a hyperspectral image to solve the technical problems mentioned in the background technology.

[0005] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0006] The present invention provides a method for detecting noise labels in hyperspectral images, comprising the following steps:

[0007] S1, preprocessing the source hyperspectral image to obtain a preprocessed hyperspectral image;

[0008] S2, construct an isolation forest for the preprocessed hyperspectral image using a dual-channel isolation forest algorithm to obtain the pixel anomaly probability of each pixel;

[0009] S3, optimizing the pixel anomaly probability of each pixel to obtain an optimized noise label training sample set;

[0010] S4. Remove the optimized noise label training sample set from the training sample set and input it into the trained classifier to obtain the final recognition result.

[0011] Furthermore, the S1 specifically includes the following steps:

[0012] S11. Source hyperspectral image for input The dimension reduction is performed by using the average segmentation and spectral fusion method to obtain the hyperspectral image after dimension reduction. in represents a three-dimensional real space, H×W is the spatial resolution of the image, B is the number of bands of the source hyperspectral image, and O is the number of bands after dimensionality reduction;

[0013] S12. Deblurring the hyperspectral image after dimensionality reduction using a relative total variation algorithm to obtain a preprocessed hyperspectral image R.

[0014] Furthermore, the relative total variation algorithm in S12 realizes image denoising and de-texturing operations by expanding the differences in the structure and texture details of the hyperspectral image after dimensionality reduction, which can be expressed as:

[0015] R=RTV(I,λ,σ)

[0016] Among them, RTV represents the deblurring operation of the hyperspectral image after dimensionality reduction using the relative total variation algorithm, λ is the weight parameter of the regularization term, and σ is the scale parameter.

[0017] Furthermore, the S2 specifically includes the following steps:

[0018] S21, for the preprocessed hyperspectral image R, randomly select 3% of the total number of pixels as a sampling subset;

[0019] S22, performing threshold cutting on the sampling subset, dividing the pixels in the sampling subset into two first child nodes;

[0020] S23, repeat the operation of dividing the child nodes in S22 until the tree height reaches the height limit or the number of pixels in each first child node is only 1, and the construction of a first isolated tree is completed;

[0021] S24, looping S22 to S23, continuously repeating the construction process of the first isolation tree T times, completing the construction of the first global isolation forest model, and obtaining the global pixel anomaly probability Pg through the first global isolation forest model test;

[0022] S25, extracting the local pixel abnormality probability P1 for the preprocessed hyperspectral image R;

[0023] S26, based on the global pixel abnormal probability P gThe pixel anomaly probability Pf is obtained by pixel-level fusion based on the local pixel anomaly probability P1. The pixel anomaly probability Pf is expressed by the formula as follows:

[0024] P f =P g P1

[0025] Here, · is the dot product symbol.

[0026] Furthermore, the S25 specifically includes the following steps:

[0027] S251, performing principal component analysis on the preprocessed hyperspectral image R, extracting a first principal component graph, and inputting the first principal component graph into an entropy rate superpixel segmentation algorithm to obtain n superpixel blocks, and using the superpixel blocks as indexes, segmenting the preprocessed hyperspectral image R into n local homogeneous regions;

[0028] S252, randomly selecting pixels in the mth homogeneous area, and continuously dividing it into multiple second child nodes by a first set threshold value, until the tree height reaches a height limit or the number of pixels in each second child node is only 1, thereby completing the construction of a second isolated tree;

[0029] S253, repeat the tree construction process T times to complete the construction of the second isolation forest model of the mth local homogeneous region, take m=1, 2, ..., n, and complete the construction of the second isolation forest model of all homogeneous regions;

[0030] S254, by testing the second isolation forest model of all homogeneous regions, the pixel anomaly probability P based on the local region is obtained. l .

[0031] Furthermore, the S3 specifically includes the following steps:

[0032] S31, pixel abnormality probability P f The K-means clustering algorithm is used for clustering. The training sample set is divided into two categories through the K-means clustering algorithm, namely the real label training sample set S c And the noise label training sample set S n ;

[0033] S32, training sample set S for noise labels n The pixel distance of the image block is calculated to complete the optimization, and the optimized noise label training sample set is obtained.

[0034] Furthermore, the S32 specifically includes the following steps:

[0035] S321, first calculate the undirected similarity measure d representing pixels a and b u (al , b l ):

[0036] S322, calculate observation pixel X ij and X st The image patch distance d IPD (X ij , X st );

[0037] S323, calculate pixel X p For the image block distance of the training sample set, that is, pixel X p The average distance between the image blocks of all pixels in the training sample set

[0038] S324, training sample set S for noise label n Calculate the distance between all pixels in the training sample set and the image block, and use the following criteria to determine whether it is a noise label sample:

[0039]

[0040] Wherein, β is the second set threshold;

[0041] S325: Integrate all noise label samples in S324 to obtain an optimized noise label training sample set

[0042] Furthermore, the undirected similarity measure d between pixels a and b in S321 is u (a l , b l ), and the formula is as follows:

[0043]

[0044] Among them, Ω represents the spatial neighborhood set; a l and b l Respectively represent the spatial neighborhood set Ω(X ij ) and Ω(X st )’s lth pixel, that is w is a positive odd number; d(a l , b) represents pixel a l and pixel b spectral similarity function; d(b l , a) represents pixel b l and the spectral similarity function of pixel a.

[0045] Furthermore, the observed pixel X in S322 ij and X st The image patch distance d IPD (X ij, X st ), and the formula is as follows:

[0046]

[0047] Furthermore, the pixel X in S323 a The average distance between the image blocks of all pixels in the training sample set The formula is as follows:

[0048]

[0049] Where N represents the total number of pixels in the training sample set, X q Indicates that the training sample set is excluding pixel X p For other pixels other than , q represents the number of pixels in the training sample set.

[0050] Furthermore, the S4 specifically includes the following steps:

[0051] S41, the optimized noise label training sample set Remove from the training sample set to obtain an optimized training sample set with true labels;

[0052] S42, using the optimized real label training sample set as input, training the SVM classification model, using the trained classifier to test the optimized real label training sample set to obtain the final recognition result.

[0053] The second aspect of the present invention further provides a hyperspectral image noise label detection system, comprising a microprocessor and a memory connected to each other, wherein the microprocessor is programmed or configured to execute the hyperspectral image noise label detection method described above.

[0054] The third aspect of the present invention further provides a computer-readable storage medium, in which a computer program is stored. The computer program is used to be programmed or configured by a microprocessor to execute the above-mentioned hyperspectral image noise label detection method.

[0055] Beneficial effects of the present invention:

[0056] The present invention uses a dual-channel isolation forest model (i.e., a first isolation forest model and a second isolation forest model) to fully utilize global and local spatial spectrum information and accurately detect abnormal sample sets;

[0057] At the same time, the present invention can effectively detect and remove noise labels in hyperspectral images, solving the problem of a sharp drop in classification performance due to the presence of noise labels in the training set, thereby achieving efficient and accurate classification of objects in the presence of noise labels;

[0058] In addition, the present invention optimizes the abnormal sample set by clustering algorithm and image block distance in turn to obtain a more accurate noise label sample set, thus providing a guarantee for high-precision classification tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 is a flow chart of the present invention;

[0060] Figure 2 It is a schematic diagram of an example of the present invention;

[0061] Figure 3 Figure 1 is the recognition result of the present invention on Pavia University data, where (a) is the ground truth label, (b) is the SVM recognition result without noise labels, (c) is the SVM recognition result with noise labels, (d)-(g) are the results obtained by several other existing hyperspectral image noise label detection methods, and Figure (h) is the recognition result obtained by the present invention. DETAILED DESCRIPTION

[0062] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. Preferred embodiments of the present invention are provided in the drawings. However, the present invention can be implemented in many other different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0063] The terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0064] Reference Figure 1 The embodiment of the present invention provides a method for detecting noise labels in a hyperspectral image, comprising the following steps:

[0065] S1, preprocessing the source hyperspectral image to obtain a preprocessed hyperspectral image;

[0066] S2, construct an isolation forest for the preprocessed hyperspectral image using a dual-channel isolation forest algorithm to obtain the pixel anomaly probability of each pixel;

[0067] S3, optimizing the pixel anomaly probability of each pixel to obtain an optimized noise label training sample set;

[0068] S4. Remove the optimized noise label training sample set from the training sample set and input it into the trained classifier to obtain the final recognition result.

[0069] In some embodiments, the S1 specifically includes the following steps:

[0070] S11. Source hyperspectral image for input The dimension reduction is performed by using the average segmentation and spectral fusion method to obtain the hyperspectral image after dimension reduction. in represents a three-dimensional real space, H×W is the spatial resolution of the image, B is the number of bands of the source hyperspectral image, and O is the number of bands after dimensionality reduction.

[0071] In this embodiment, the number of bands O after dimensionality reduction is set to 20. The method based on average segmentation and spectral band fusion is a simple and effective hyperspectral data dimensionality reduction technology, the principle of which is to evenly segment the image into multiple subsets, and then fuse the data in each subset to obtain reduced dimensionality data, the main advantage of which is that the physical meaning of the data can be retained at the same time.

[0072] S12, performing deblurring processing on the hyperspectral image after dimensionality reduction using a relative total variation algorithm to obtain a preprocessed hyperspectral image R.

[0073] In some embodiments, the relative total variation algorithm in S12 implements image denoising and de-texturing operations by expanding the differences in structure and texture details of the hyperspectral image after dimensionality reduction, which can be expressed as:

[0074] R=RTV(I,λ,σ)

[0075] Among them, RTV represents the deblurring operation of the hyperspectral image after dimensionality reduction using the relative total variation algorithm, λ is the weight parameter of the regularization term, and σ is the scale parameter.

[0076] In some embodiments, S2 specifically includes the following steps:

[0077] S21, for the preprocessed hyperspectral image R, randomly select 3% of the total number of pixels as a sampling subset;

[0078] S22, performing threshold cutting on the sampling subset, dividing the pixels in the sampling subset into two first child nodes, namely, a left node and a right node;

[0079] S23, repeat the operation of dividing the child nodes in S22 until the tree height reaches the height limit or the number of pixels in each first child node is only 1, and the construction of a first isolated tree is completed;

[0080] S24, loop S22 to S23, repeat the construction process of the first isolation tree T times, complete the construction of the first global isolation forest model, and obtain the global pixel anomaly probability P through the first global isolation forest model test g ;

[0081] S25: extract the local pixel abnormality probability P for the preprocessed hyperspectral image R l ;

[0082] S26, based on the global pixel abnormal probability P g and the local pixel abnormality probability P l Perform pixel-level fusion to obtain the final pixel anomaly probability P f ; Pixel abnormality probability P f The formula is as follows:

[0083] P f =P g ·P l

[0084] Here, · is the dot product symbol.

[0085] In some embodiments, the S25 specifically includes the following steps:

[0086] S251, performing principal component analysis on the preprocessed hyperspectral image R, extracting a first principal component graph, and inputting the first principal component graph into an entropy rate superpixel segmentation algorithm (ERS) to obtain n superpixel blocks, and using them as indexes to segment the preprocessed hyperspectral image R into n local homogeneous regions;

[0087] S252, randomly selecting pixels in the mth homogeneous area, and continuously dividing it into multiple second child nodes by a first set threshold value, until the tree height reaches a height limit or the number of pixels in each second child node is only 1, thereby completing the construction of a second isolated tree;

[0088] S253, repeat the tree construction process T times to complete the construction of the second isolation forest model of the mth local homogeneous region, take m=1, 2, ..., n, and complete the construction of the second isolation forest model of all homogeneous regions;

[0089] S254, by testing the second isolation forest model of all homogeneous regions, the pixel anomaly probability P based on the local region is obtained. l The second isolation forest model is a local model.

[0090] In some embodiments, S3 specifically includes the following steps:

[0091] S31, pixel abnormality probability P fThe K-means clustering algorithm is used for clustering. The number of cluster centers is set to 2. The training sample set is divided into two categories by the K-means clustering algorithm, namely the real label training sample set S c And the noise label training sample set S n ;

[0092] S32, training sample set S for noise labels n The pixel distance of the image block is calculated to complete the optimization, and the optimized noise label training sample set is obtained.

[0093] In some embodiments, the S32 specifically includes the following steps:

[0094] S321, first calculate the undirected similarity measure d representing pixels a and b u (a l , b l ):

[0095] S322, calculate observation pixel X ij and X st The image patch distance d IPD (X ij , X st );

[0096] S323, calculate pixel X p For the image block distance of the training sample set, that is, pixel X p The average distance between the image blocks of all pixels in the training sample set

[0097] S324, training sample set S for noise label n Calculate the distance between all pixels in the training sample set and the image block, and use the following criteria to determine whether it is a noise label sample:

[0098]

[0099] Wherein, β is the second set threshold;

[0100] S325: Integrate all noise label samples in S324 to obtain an optimized noise label training sample set

[0101] In some embodiments, the undirected similarity measure d of pixels a and b in S321 is u (a l , b l ), and the formula is as follows:

[0102]

[0103] Among them, Ω represents the spatial neighborhood set; a l and b l Respectively represent the spatial neighborhood set Ω(X ij ) and Ω(X st )’s lth pixel, that is w is a positive odd number; d(a l , b) represents pixel a l and pixel b spectral similarity function; d(b l , a) represents pixel b l and the spectral similarity function of pixel a.

[0104] In some embodiments, the observation pixel X in S322 ij and X st The image patch distance d IPD (X ij , X s t), which is expressed by the following formula:

[0105]

[0106] In some embodiments, the pixel X in S323 a The average distance between the image blocks of all pixels in the training sample set The formula is as follows:

[0107]

[0108] Where N represents the total number of pixels in the training sample set, X q Indicates that the training sample set is excluding pixel X p For other pixels other than , q represents the number of pixels in the training sample set.

[0109] In some embodiments, the S4 specifically includes the following steps:

[0110] S41, the optimized noise label training sample set Remove from the training sample set to obtain an optimized training sample set with true labels;

[0111] S42, using the optimized real label training sample set as input, training the SVM classification model, using the trained classifier to test the optimized real label training sample set to obtain the final recognition result.

[0112] In order to verify a hyperspectral image noise label detection method in this embodiment, the widely used hyperspectral dataset Pavia University is used to test the present invention. The public dataset Pavia University was acquired using the ROSIS-3 imager in the University of Pavia, Italy. The data space size is 610×340 and the spatial resolution is 1.3 meters. After removing 12 noise-interfered bands, the remaining 103 spectral bands were used for subsequent classification experiments. In addition, the dataset contains a total of 9 ground object categories. The test was conducted on the Pavia University dataset, and the performance of different algorithms was evaluated using three typical objective evaluation indicators: OA, AA, and Kappa coefficient. The compared algorithms included a method based on spatial density peak clustering (SDP), a method based on hierarchical constrained energy minimization (HCEM), a method based on superpixel-to-pixel weighting distance (SPWD), and a method based on multilayer spectral-spatial graphs (MSSG).

[0113] In the Pavia University dataset, the number of real training samples and the number of noise samples for each type of ground object are 50 and 8 respectively. Four popular noise label detection methods (SDP, HCEM, SPWD, MSSG) and the objective evaluation indicators of the present invention on the Pavia University dataset are shown in Table 1. In addition, the classification results of the SVM classifier for noise-free data and noisy data are also listed in Table 1. As can be seen from Table 1, the objective evaluation indicators of the present invention are better than other methods, and are improved by about 1 to 5 percentage points compared with other comparison methods, indicating that the present invention can effectively detect and remove noise label samples in hyperspectral images. In addition, Figure 3 The classification results of four typical noise label detection methods (SDP, HCEM, SPWD, MSSG) and the method of this embodiment are shown, from which it can be seen that the use of the present invention can effectively remove noise labels, ensure high-precision classification of smooth areas, and significantly reduce misclassification of areas with dense texture structures.

[0114] Table 1: Objective performance indicators of the present invention and various popular noisy label detection methods on the Pavia University dataset.

[0115]

[0116] In summary, the method for detecting noise labels of hyperspectral images based on dual-channel isolation forests in this embodiment includes: preprocessing the source hyperspectral image, and obtaining the preprocessed hyperspectral image by dimensionality reduction and defuzzification in sequence; constructing an isolation forest for the preprocessed hyperspectral image using the dual-channel isolation forest algorithm, and obtaining the probability that each pixel is an abnormal sample by fusing the global and local pixel abnormality probabilities; optimizing the pixel abnormality probability to obtain an optimized training sample set; using the optimized training sample set as input, and using the trained classifier to obtain the recognition result. The present invention fully utilizes the global and local spatial spectrum information through the dual-channel isolation forest model to accurately detect the abnormal sample set. In addition, the present invention also optimizes the abnormal sample set through clustering algorithm and image block distance in sequence to obtain a more accurate noise label sample set, which provides a guarantee for high-precision classification tasks.

[0117] The present invention makes full use of global and local spatial spectrum information through a dual-channel isolation forest model (i.e., the first isolation forest model and the second isolation forest model) to accurately detect abnormal sample sets. At the same time, the present invention can effectively detect and remove noise labels in hyperspectral images, solving the problem of a sharp drop in classification performance due to the presence of noise labels in the training set, thereby achieving efficient and accurate classification of objects in the presence of noise labels; in addition, the present invention also optimizes the abnormal sample set through clustering algorithms and image block distances in turn to obtain a more accurate noise label sample set, providing a guarantee for high-precision classification tasks.

[0118] The second aspect of the present invention further provides a hyperspectral image noise label detection system, comprising a microprocessor and a memory connected to each other, wherein the microprocessor is programmed or configured to execute the hyperspectral image noise label detection method described above.

[0119] The third aspect of the present invention further provides a computer-readable storage medium, in which a computer program is stored. The computer program is used to be programmed or configured by a microprocessor to execute the above-mentioned hyperspectral image noise label detection method.

[0120] Those skilled in the art will appreciate that the present invention may be provided as a method, system, or computer program product. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, may be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that instructions executed by the processor of a computer or other programmable data processing device generate instructions for implementing the functions in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the functions specified in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0121] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered within the protection scope of the present invention. In addition, the technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the ability of ordinary technicians in the field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for detecting noise labels in hyperspectral images, characterized in that: The steps include: S1, preprocessing the source hyperspectral image to obtain a preprocessed hyperspectral image; S2, construct an isolation forest for the preprocessed hyperspectral image using a dual-channel isolation forest algorithm to obtain the pixel anomaly probability of each pixel; S3, optimizing the pixel anomaly probability of each pixel to obtain an optimized noise label training sample set; S4. Remove the optimized noise label training sample set from the training sample set and input it into the trained classifier to obtain the final recognition result.

2. The method for detecting hyperspectral image noise labels according to claim 1, characterized in that: The S1 specifically includes the following steps: S11. Source hyperspectral image for input The dimension reduction is performed by using the average segmentation and spectral fusion method to obtain the hyperspectral image after dimension reduction. in represents a three-dimensional real space, H×W is the spatial resolution of the image, B is the number of bands of the source hyperspectral image, and O is the number of bands after dimensionality reduction; S12, performing deblurring processing on the hyperspectral image after dimensionality reduction using a relative total variation algorithm to obtain a preprocessed hyperspectral image R; The relative total variation algorithm in S12 realizes image denoising and de-texturing operations by expanding the differences in the structure and texture details of the hyperspectral image after dimensionality reduction, which can be expressed as: R=RTV(I,λ,σ) Among them, RTV represents the deblurring operation of the hyperspectral image after dimensionality reduction using the relative total variation algorithm, λ is the weight parameter of the regularization term, and σ is the scale parameter.

3. The method for detecting hyperspectral image noise labels according to claim 1, characterized in that: The S2 specifically includes the following steps: S21, for the preprocessed hyperspectral image R, randomly select 3% of the total number of pixels as a sampling subset; S22, performing threshold cutting on the sampling subset, dividing the pixels in the sampling subset into two first child nodes; S23, repeat the operation of dividing the child nodes in S22 until the tree height reaches the height limit or the number of pixels in each first child node is only 1, and the construction of a first isolated tree is completed; S24, loop S22 to S23, repeat the construction process of the first isolation tree T times, complete the construction of the first global isolation forest model, and obtain the global pixel anomaly probability P through the first global isolation forest model test g ; S25, extracting the local pixel abnormality probability P1 for the preprocessed hyperspectral image R; S26, based on the global pixel abnormal probability P g The pixel-level fusion based on the local pixel anomaly probability P1 is performed to obtain the final pixel anomaly probability P f ; Pixel abnormality probability P f The formula is as follows: P f =P g ·P l Here, · is the dot product symbol.

4. The method for detecting hyperspectral image noise labels according to claim 3, characterized in that: The S25 specifically includes the following steps: S251, performing principal component analysis on the preprocessed hyperspectral image R, extracting a first principal component graph, and inputting the first principal component graph into an entropy rate superpixel segmentation algorithm to obtain n superpixel blocks, and using the superpixel blocks as indexes, segmenting the preprocessed hyperspectral image R into n local homogeneous regions; S252, randomly selecting pixels in the mth homogeneous area, and continuously dividing it into multiple second child nodes by a first set threshold value, until the tree height reaches a height limit or the number of pixels in each second child node is only 1, thereby completing the construction of a second isolated tree; S253, repeat the tree construction process T times to complete the construction of the second isolation forest model of the mth local homogeneous region, take m=1, 2, ..., n, and complete the construction of the second isolation forest model of all homogeneous regions; S254, by testing the second isolation forest model of all homogeneous regions, the pixel anomaly probability P based on the local region is obtained. l .

5. The method for detecting hyperspectral image noise labels according to claim 3, characterized in that: The S3 specifically includes the following steps: S31, pixel abnormality probability P f The K-means clustering algorithm is used for clustering. The training sample set is divided into two categories through the K-means clustering algorithm, namely the real label training sample set S c And the noise label training sample set S n ; S32, noise label training sample set S n The pixel distance of the image block is calculated to complete the optimization, and the optimized noise label training sample set is obtained.

6. The method for detecting hyperspectral image noise labels according to claim 5, characterized in that: The S32 specifically includes the following steps: S321, first calculate the undirected similarity measure d of pixels a and b u (a l , b l ), and the formula is as follows: Among them, Ω represents the spatial neighborhood set; a l and b l Respectively represent the spatial neighborhood set Ω(X ij ) and Ω(X st )’s lth pixel, that is and w is a positive odd number; d(a l , b) represents pixel a l and pixel b spectral similarity function; d(b l , a) represents pixel b l and the spectral similarity function of pixel a; S322, calculate the observed pixel x ij and x st The image patch distance d IPD (X ij , X st ); S323, calculate pixel X p For the image block distance of the training sample set, that is, pixel X p The average distance between the image blocks of all pixels in the training sample set S324, training sample set S for noise label n Calculate the distance between all pixels in the training sample set and the image block, and use the following criteria to determine whether it is a noise label sample: Wherein, β is the second set threshold; S325: Integrate all noise label samples in S324 to obtain an optimized noise label training sample set 7. The method for detecting hyperspectral image noise labels according to claim 6, characterized in that: The observed pixel x in S322 ij and x st The image patch distance d IPD (X ij , X st ), and the formula is as follows:

8. The method for detecting hyperspectral image noise labels according to claim 7, characterized in that: The pixel X in S323 p The average distance between the image blocks of all pixels in the training sample set The formula is as follows: Where N represents the total number of pixels in the training sample set, X q Indicates that the training sample set is excluding pixel X p For other pixels other than , q represents the number of pixels in the training sample set.

9. A hyperspectral image noise label detection system, characterized in that: The invention comprises a microprocessor and a memory connected to each other, wherein the microprocessor is programmed or configured to execute the hyperspectral image noise label detection method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is used to be programmed or configured by a microprocessor to execute the hyperspectral image noise label detection method according to any one of claims 1 to 8.

Images