A fast anomaly detection method and device for hyperspectral images
By integrating tensors, structural sparse norms and convolutional neural network denoisers in hyperspectral image anomaly detection, the problems of insufficient information retention and noise interference in the existing technology are solved, and more accurate and complete anomalies are achieved. Test results.
Patent Information
- Application Number
- CN202410688201.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-30
- Publication Date
- 2025-01-21
- Estimated Expiration
- 2044-05-30
AI Technical Summary
The existing hyperspectral image abnormality detection technology is difficult to effectively retain spatial information and spectral information in the image, resulting in the distribution of abnormal regions that are not complete and authentic, and are susceptible to noise interference.
An integrated method based on tensors, structural sparse norm and convolutional neural network denoiser is used to design a hyperspectral image anomaly detection model. The spatial and spectral information is retained through tensor decomposition, structural sparse norm is introduced for feature extraction, and the use of The convolutional neural network denoiser ensures the preservation of key features.
It effectively retains the spatial and spectral information in hyperspectral images, improves the accuracy and completeness of anomaly detection, reduces noise interference, and retains the original details and clarity of the image to the maximum extent.
Smart Images
Figure CN118552802B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of hyperspectral image processing, and in particular, relates to a method and device for rapid anomaly detection of hyperspectral images. Background Art
[0002] With the rapid advancement of remote sensing technology, hyperspectral images have attracted much attention because they can provide rich spatial and spectral information at the same time. Hyperspectral anomaly detection, as an important research direction in this field, aims to use algorithms to identify pixels that are significantly different from the surrounding pixels in the absence of prior information about the target, and mark them as abnormal targets. This process not only divides the hyperspectral image into two parts: anomaly and background, but also helps to deeply analyze the characteristics of the ground objects, thus showing a wide range of application value in many fields such as mineral exploration, agricultural monitoring, and environmental protection.
[0003] However, it is worth noting that hyperspectral anomaly target detection is essentially an unsupervised learning process, which means that it is impossible to obtain information such as the target's spectral characteristics, color distribution, or texture structure before detection. This "blind detection" state makes the detection process more complex and difficult than traditional target detection. Therefore, the development of efficient and accurate anomaly detection algorithms is of great significance to improving the ability of hyperspectral image analysis.
[0004] Existing technologies have circumvented the defects of insufficient depth of spectral information mining in various matrix-based hyperspectral anomaly detection methods in history, and adopted tensor principal component analysis or tensor low-rank representation to perform anomaly detection in hyperspectral images. For example, Li et al. introduced TPCA into hyperspectral anomaly detection and applied Tucker decomposition to extract core tensors. Li et al. proposed a priority tensor approximation based on truncated nuclear norm while considering total variation regularization to control the piecewise smoothness of the embedded space and improve detection efficiency. Wang et al. used principal component analysis for data preprocessing and tensor low-rank sparse representation for anomaly detection. Sun et al. proposed combining low average rank with TV regularization to improve background representation. Feng et al. tried to use tensor ring decomposition. Although it effectively improved the accuracy and comprehensiveness of spectral information extraction, it still had limitations, which are as follows: 1. Existing technologies for target feature extraction of hyperspectral images often ignore their interrelated structural information, resulting in the incompleteness and authenticity of the distribution of abnormal areas in hyperspectral images.
[0005] 2. The accuracy of target feature extraction in hyperspectral images by existing technologies is easily affected by noise and lacks the ability to effectively process image noise. Summary of the invention
[0006] In view of this, in order to solve the problems raised in the above background technology, a fast image generation method and device based on a spatial sparse diffusion model are proposed.
[0007] The technical solution adopted by the present invention to solve its technical problem is: in the first aspect, the present invention provides a fast anomaly detection method for hyperspectral images, including: S1. Model function design: designing a core function by integrating tensors, structured sparse norms and convolutional neural network denoisers to construct a hyperspectral image anomaly detection model.
[0008] S2. Detection image input: Input the image to be detected into the hyperspectral image anomaly detection model.
[0009] S3. Model function optimization: The core function of the hyperspectral image anomaly detection model is optimized by introducing intermediate variables and constructing their corresponding Lagrangian functions.
[0010] S4. Model convergence detection: Adopt an iterative optimization mechanism to continuously analyze the optimization progress of each sub-variable in the hyperspectral image anomaly detection model, and use the built-in convergence detection mechanism to determine whether the Lagrangian function reaches the preset convergence standard. If it is, execute S5, otherwise repeat S3 and S4.
[0011] S5. Model result output: When the Lagrangian function reaches the preset convergence standard, the hyperspectral image anomaly detection model outputs the spectral anomaly detection result of the image to be detected and displays it visually.
[0012] In a second aspect, the present invention provides a rapid anomaly detection device for hyperspectral images, comprising: a model function design module, a detection image input module, a model function optimization module, a model convergence detection module, a model result output module and a cloud database.
[0013] The model function design module is connected to the detection image input module, the detection image input module is connected to the model function optimization module, the model function optimization module is connected to the model convergence detection module, the model convergence detection module is respectively connected to the model function optimization module and the model result output module, and the cloud database is connected to the model function optimization module.
[0014] The model function design module is used to design the core function by integrating tensors, structured sparse norms, and convolutional neural network denoisers to build a hyperspectral image anomaly detection model.
[0015] The detection image input module is used to input the image to be detected into the hyperspectral image anomaly detection model.
[0016] The model function optimization module is used to optimize the core function of the hyperspectral image anomaly detection model by introducing intermediate variables and constructing their corresponding Lagrangian functions.
[0017] The model convergence detection module is used to adopt an iterative optimization mechanism to continuously analyze the optimization progress of each sub-variable in the hyperspectral image anomaly detection model, and use the built-in convergence detection mechanism to determine whether the Lagrangian function reaches the preset convergence standard. If it is, the model result output module is executed, otherwise the model function optimization module and the model convergence detection module are repeatedly executed.
[0018] The model result output module, when the Lagrangian function reaches the preset convergence standard, the hyperspectral image anomaly detection model outputs the spectral anomaly detection result of the image to be detected and displays it visually.
[0019] Cloud database, used to store various test data sets.
[0020] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) The present invention adopts a tensor decomposition model to perform hyperspectral anomaly detection, effectively retaining the spatial information and spectral information in the image, thereby avoiding the loss of details during the anomaly detection process, and further improving the information mining capability and anomaly detection capability of hyperspectral images.
[0021] (2) The present invention introduces a structured sparse norm to extract features and optimize the solution of hyperspectral images, so that the intrinsic correlation and structural information in the hyperspectral image data can be taken into account during the feature extraction process. It can not only effectively reduce redundant features, but also retain important spatial-spectral features, thereby simulating and verifying the true distribution of simulated abnormal information.
[0022] (3) The present invention introduces a convolutional neural network denoiser to ensure that key local features in the hyperspectral image, such as edges, textures and microstructures, will not be smoothed or lost during the denoising process. This not only enables high-quality denoising, but also maximizes the retention of the original details and clarity of the image. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.
[0024] Figure 1 It is a schematic diagram of the method flow of the present invention.
[0025] Figure 2 This is a connection diagram of the device modules of the present invention.
[0026] Figure 3 Schematic diagram of the structure of the hyperspectral feature anomaly detection model.
[0027] Figure 4 Schematic diagram of the cross-validation results of each test data set.
[0028] Figure 5 It is a schematic diagram for comparing the test results of the method of the present invention.
[0029] Figure 6 Schematic diagram of ROC curve comparison of the method of the present invention.
[0030] Figure 7 It is a schematic diagram of box-line comparison of the method of the present invention.
[0031] Figure 8 It is a schematic diagram of the relative error convergence curve of the present invention. DETAILED DESCRIPTION
[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0033] Reference Figure 1 and Figure 3 As shown, the first aspect of the present invention provides a fast anomaly detection method for hyperspectral images, including: S1. Model function design: designing a core function by integrating tensors, structured sparse norms and convolutional neural network denoisers to construct a hyperspectral image anomaly detection model.
[0034] Specifically, the core function expression of the hyperspectral image anomaly detection model is: Where ε is the abnormal information of the image to be detected, which is expressed as a tensor, w is the corresponding coefficient tensor related to the abnormality, A is the dictionary representing the background part, x is the image to be detected, Ω(ε) is the structured sparse induced norm, ||·|| WTNN is the weighted tensor nuclear norm, φ(ε) is the convolutional neural network denoiser, and λ and μ are the preset weight coefficients respectively.
[0035] Specifically, the structured sparsity induced norm expression is: where e j is the jth column of the tensor ε, j = 1, 2, ..., n, G is the set of clusters of the image hyperspectral data to be detected, g is the unit group in the G set, ||·|| ∞ For L ∞ Norm.
[0036] It should be noted that the above stx=A*w+ε means: the background part of the image to be detected is reconstructed using the dictionary A and the coefficient tensor w, and the remaining part is used as the abnormal part of the image to be detected.
[0037] The embodiment of the present invention adopts a tensor decomposition mode to perform hyperspectral anomaly detection, effectively retaining the spatial information and spectral information in the image, thereby avoiding the loss of details during the anomaly detection process, and further improving the information mining capability and anomaly detection capability of the hyperspectral image.
[0038] S2. Detection image input: Input the image to be detected into the hyperspectral image anomaly detection model.
[0039] S3. Model function optimization: The core function of the hyperspectral image anomaly detection model is optimized by introducing intermediate variables and constructing their corresponding Lagrangian functions.
[0040] Specifically, the optimization of the core function of the hyperspectral image anomaly detection model includes: introducing intermediate variables z and y to obtain an equivalent expression of the core function of the hyperspectral image anomaly detection model: Construct the corresponding Lagrangian function expression L β (z,ε,y,w,Q1,Q2,Q3) to further optimize the core function, L β (z,ε,y,w,Q1,Q2,Q3)=R1+R2+R3+R4, R1=||z|| WTNN +λΩ(ε)+μφ(y), Q1, Q2, and Q3 are Lagrangian operators, and β is a preset coefficient.
[0041] S4. Model convergence detection: Adopt an iterative optimization mechanism to continuously analyze the optimization progress of each sub-variable in the hyperspectral image anomaly detection model, and use the built-in convergence detection mechanism to determine whether the Lagrangian function reaches the preset convergence standard. If it is, execute S5, otherwise repeat S3 and S4.
[0042] The embodiment of the present invention introduces a structured sparse norm to perform feature extraction and optimization on hyperspectral images, so that the intrinsic correlation and structural information in the hyperspectral image data can be taken into account in the feature extraction process, which can not only effectively reduce redundant features, but also retain important spatial-spectral features, thereby simulating and verifying the real distribution of simulated abnormal information.
[0043] Specifically, the iterative optimization mechanism includes: using the local minimum explicit solution of a single variable to replace the minimum solution of the current optimization core function, substituting each variable back into the Lagrangian function in a single iteration process to check whether it converges. The iterative process is as follows: Where k is the number of iterations, further defined by the formula
[0044] Get the solution value of each sub-variable respectively, where Represented as the matrix E folded into a tensor along the i-th dimension, Denoiser is a plug-and-play convolutional neural network denoiser.
[0045] The embodiment of the present invention introduces a convolutional neural network denoiser to ensure that key local features in the hyperspectral image, such as edges, textures and tiny structures, will not be smoothed or lost during the denoising process. It can not only achieve high-quality denoising but also retain the original details and clarity of the image to the maximum extent.
[0046] Specifically, the built-in convergence detection mechanism includes: within the rated number of iterations, if the relative error of the core function value after a single iteration is less than or equal to a preset reasonable error threshold, or if the error change of the core function value for more than a preset number of iterations is less than or equal to a preset reasonable error change threshold, then it is judged that the Lagrangian function has reached the preset convergence standard; otherwise, it is judged that the Lagrangian function has not reached the preset convergence standard.
[0047] Specifically, the process of determining the values of λ, μ, and β in the core function includes: setting a preset number of groups, each group containing different λ, μ, and β values, substituting them into each test data set stored in the WEB cloud and conducting a ten-fold cross-validation experiment, obtaining the experimental results of each test data set and calculating its AUC value under different λ, μ, and β values in the group, and selecting a group of corresponding λ, μ, and β values with the best AUC value performance as the values of λ, μ, and β in the core function.
[0048] It should be noted that the above-mentioned test data sets specifically refer to the ABU-Urabn3 data set, the ABU-Urabn4 data set and the ABU-Urabn5 data set.
[0049] It should also be noted that the values of λ, μ, and β are set to 10 -6 -10 2 , a total of 10 groups.
[0050] It should be further explained that the above-mentioned ten-fold cross-validation experiment is a statistical method for evaluating the performance of machine learning models. The basic idea is to divide the original data into a set number of parts, use one of them as a validation set each time, and use the remaining parts as training sets, and perform a set number of training and validation. Finally, the average error of the set number of experiments can be calculated to estimate the performance of the model.
[0051] Reference Figure 4As shown, (a) and (d) are the results of the ten-fold cross-validation experiment on the ABU-Urban3 dataset, (b) and (e) are the results of the ten-fold cross-validation experiment on the ABU-Urban4 dataset, and (c) and (f) are the results of the ten-fold cross-validation experiment on the ABU-Urban5 dataset. It can be seen from the figure that the values of λ, μ, and β in the core function are 0.01, 0.01, and 0.1, respectively.
[0052] Specifically, the calculation formula of the AUC value includes: That M and N are the number of positive samples and negative samples in the test data set, respectively. M , P N are the proportion of positive samples and the proportion of negative samples in the test data set, respectively.
[0053] S5. Model result output: When the Lagrangian function reaches the preset convergence standard, the hyperspectral image anomaly detection model outputs the spectral anomaly detection result of the image to be detected and displays it visually.
[0054] Reference Figure 5 As shown, the hyperspectral anomaly detection method proposed in the present invention is compared with nine existing methods, namely RX, RPCA, LSMAD, LRASR, TPCA, GTVLRR, DeCNN-AD, PTA and PCA-TLRSR, and a visualization result is obtained. The detection results and AUC values obtained by the hyperspectral anomaly detection method of the present invention and other methods on the San Diego, ABU-Urabn2, and ABU-Urban3 datasets are shown. (a) Input image, (b) True value, (c) RX, (d) RPCA, (e) LSMAD, (f) LRASR, (g) TPCA, (h) GTVLRR, (i) DeCNN-AD, (j) PTA, (k) PCA-TLRSR, (l) This work, it can be seen that the present invention exhibits stronger background suppression ability, and the detection results are consistent with the ground truth. Its excellent performance can be attributed to the introduction of structural sparsity and plug-and-play convolutional neural network denoising.
[0055] Reference Figure 6As shown, the ROC curve represents the change of detection probability with the change of false alarm rate. If the ROC curve related to the method is near the upper left corner, it indicates that the performance is good, characterized by a high detection rate and a low FAR. (a), (b), and (c) respectively represent the ROC curves of the proposed hyperspectral anomaly detection method and the existing nine methods on the SanDiego, ABU-Beach2, and ABU-Urban3 datasets. It can be seen that with the increase of FAR, the PD value of the method of the present invention exceeds that of other methods, proving the effectiveness of the hyperspectral anomaly detection method of the present invention.
[0056] Reference Figure 7 As shown in the figure, the upper box represents the abnormal distribution, and the lower box represents the background distribution. The distance between the two boxes can intuitively reflect the ability to separate abnormalities. The larger the distance between the two boxes, the better the abnormality detection. (a), (b), and (c) respectively represent the box plots of the hyperspectral anomaly detection method proposed in the present invention and the nine existing methods on the San Diego, ABU-Urban2, and ABU-Urban3 datasets, respectively. It can be seen that the hyperspectral anomaly detection method of the present invention has the best separation ability.
[0057] Reference Figure 8 As shown, (a), (b), and (c) respectively represent the convergence of the invention on the San Diego dataset, the HYDICE-Urban dataset, and the ABU dataset. It can be seen from the figure that in the first 40 iterations, the relative error fluctuates, and in the subsequent iterations, the convergence curve gradually presents a smoother trend. In addition, it can be observed that the relative errors of all datasets within 100 iterations are close to zero, which further verifies the convergence of the invention.
[0058] Reference Figure 2 As shown, the second aspect of the present invention provides a rapid anomaly detection device for hyperspectral images, including: a model function design module, a detection image input module, a model function optimization module, a model convergence detection module, a model result output module and a cloud database.
[0059] The model function design module is connected to the detection image input module, the detection image input module is connected to the model function optimization module, the model function optimization module is connected to the model convergence detection module, the model convergence detection module is respectively connected to the model function optimization module and the model result output module, and the cloud database is connected to the model function optimization module.
[0060] The model function design module is used to design the core function by integrating tensors, structured sparse norms, and convolutional neural network denoisers to build a hyperspectral image anomaly detection model.
[0061] The detection image input module is used to input the image to be detected into the hyperspectral image anomaly detection model.
[0062] The model function optimization module is used to optimize the core function of the hyperspectral image anomaly detection model by introducing intermediate variables and constructing their corresponding Lagrangian functions.
[0063] The model convergence detection module is used to adopt an iterative optimization mechanism to continuously analyze the optimization progress of each sub-variable in the hyperspectral image anomaly detection model, and use the built-in convergence detection mechanism to determine whether the Lagrangian function reaches the preset convergence standard. If it is, the model result output module is executed, otherwise the model function optimization module and the model convergence detection module are repeatedly executed.
[0064] The model result output module, when the Lagrangian function reaches the preset convergence standard, the hyperspectral image anomaly detection model outputs the spectral anomaly detection result of the image to be detected and displays it visually.
[0065] Cloud database, used to store various test data sets.
[0066] It should be noted that the test data sets stored in the above cloud database are obtained by external input.
[0067] The above contents are merely examples and explanations of the structure of the present invention. The technicians in this technical field may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the structure of the invention or exceed the scope defined by the present invention, they should all fall within the protection scope of the present invention.
Claims
1. A fast anomaly detection method for hyperspectral images, characterized in that: include: S1. Model function design: Design core functions by integrating tensors, structured sparse norms, and convolutional neural network denoisers to build a hyperspectral image anomaly detection model; S2. Detection image input: input the image to be detected into the hyperspectral image anomaly detection model; S3. Model function optimization: By introducing intermediate variables and constructing their corresponding Lagrangian functions, the core function of the hyperspectral image anomaly detection model is optimized; S4. Model convergence detection: Adopting the iterative optimization mechanism, continuously analyzing the optimization progress of each sub-variable in the hyperspectral image anomaly detection model, and using the built-in convergence detection mechanism to determine whether the Lagrangian function reaches the preset convergence standard, if it is determined to be, then execute S5, otherwise repeat S3 and S4; S5. Model result output: When the Lagrangian function reaches the preset convergence standard, the hyperspectral image anomaly detection model outputs the spectral anomaly detection result of the image to be detected and displays it visually; The core function expression of the hyperspectral image anomaly detection model is: ,in is the abnormal information of the image to be detected, which is expressed as a tensor. is the corresponding coefficient tensor of the anomaly correlation, is a dictionary representing the background part, is the image to be detected, is the structured sparsity induced norm, is the weighted tensor nuclear norm, is the convolutional neural network denoiser, are preset weight coefficients respectively; The optimization of the core function of the hyperspectral image anomaly detection model includes: introducing intermediate variables and , get the equivalent expression of the core function of the hyperspectral image anomaly detection model: , construct the corresponding Lagrangian function expression To further optimize the core functions, , , , , , are the Lagrangian operators, is the preset coefficient.
2. The method for rapid anomaly detection in hyperspectral images according to claim 1, characterized in that: The structured sparsity induced norm expression is: ,in For tensors No. List, , is the cluster set of hyperspectral data of the image to be detected, for The group of units in the set, for Norm.
3. The method for rapid anomaly detection in hyperspectral images according to claim 1, characterized in that: The iterative optimization mechanism includes: using the local minimum explicit solution of a single variable to replace the minimum solution of the current optimization core function, substituting each variable back into the Lagrangian function in a single iteration process to check whether it converges. The iterative process is as follows: ,in is the number of iterations, further defined by the formula , , , Get the solution value of each sub-variable respectively, where Represented as a matrix Along the Dimensions are collapsed into tensors, A plug-and-play convolutional neural network denoiser.
4. The method for rapid anomaly detection in hyperspectral images according to claim 3, characterized in that: The built-in convergence detection mechanism includes: within the rated number of iterations, if the relative error of the core function value after a single iteration is less than or equal to a preset reasonable error threshold or the error change of the core function value for more than a preset number of iterations is less than or equal to a preset reasonable error change threshold, then it is judged that the Lagrangian function has reached the preset convergence standard; otherwise, it is judged that the Lagrangian function has not reached the preset convergence standard.
5. The method for rapid anomaly detection in hyperspectral images according to claim 1, characterized in that: In the core function The process of determining the value includes: setting the number of preset groups, each group contains different The values are substituted into each test data set stored in the WEB cloud and a ten-fold cross validation experiment is performed to obtain the experimental results of each test data set and calculate its The AUC value under the numerical value, select the corresponding group with the best AUC value Values as core functions The numerical value of .
6. The method for rapid anomaly detection in hyperspectral images according to claim 5, characterized in that: The calculation formula of the AUC value includes: ,in , are the number of positive samples and the number of negative samples in the test data set, respectively. are the proportion of positive samples and the proportion of negative samples in the test data set, respectively.
7. A fast anomaly detection device for hyperspectral images, used to execute the method according to any one of claims 1 to 6, characterized in that: include: The model function design module is used to design the core function by integrating tensors, structured sparse norms, and convolutional neural network denoisers to build a hyperspectral image anomaly detection model; A detection image input module is used to input the image to be detected into the hyperspectral image anomaly detection model; The model function optimization module is used to optimize the core function of the hyperspectral image anomaly detection model by introducing intermediate variables and constructing their corresponding Lagrangian functions; The model convergence detection module is used to adopt an iterative optimization mechanism to continuously analyze the optimization progress of each sub-variable in the hyperspectral image anomaly detection model, and use the built-in convergence detection mechanism to determine whether the Lagrangian function reaches the preset convergence standard. If it is, the model result output module is executed, otherwise the model function optimization module and the model convergence detection module are repeatedly executed; Model result output module: when the Lagrangian function reaches the preset convergence standard, the hyperspectral image anomaly detection model outputs the spectral anomaly detection result of the image to be detected and displays it visually; Cloud database, used to store various test data sets.