Hyperspectral air-surface mask variational autoencoder anomaly detection method

Abstract

The application belongs to the technical field of computer vision. The application provides a hyperspectral space-spectrum mask variational auto-encoding anomaly detection method for on-site security. The space-spectrum joint mask technology designed in the embodiment of the disclosure fully utilizes the multi-dimensional information of space and spectrum, improves the difficulty of effectively reconstructing the abnormal area in the reconstruction process of the network, makes the network pay more attention to the normal spatial and spectral features when learning the image data distribution, and significantly improves the detection ability of complex abnormal scenes. The VAE reconstruction network and the mask combination mechanism designed fully utilize the advantages of probability modeling of VAE and the advantages of attention mechanism of mask, improve the accuracy and generalization ability of anomaly detection, and strengthen the explainability of the model. And the residual image processing method designed through the domain transformation recursive filter effectively relieves the noise problem caused by the complex background, and realizes higher detection precision and robustness.

Description

TECHNICAL FIELD

[0001] Embodiments of the present disclosure relate to the technical field of computer vision, and particularly relate to a hyperspectral air-spectrum mask variational auto-encoding anomaly detection method for vicinagearth security. BACKGROUND

[0002] The current global security situation is becoming increasingly complex, and national and regional security is facing new challenges. The concept of vicinagearth security (VS) has emerged as the times require. Vicinagearth security is a set of technical systems that integrates multi-element, cross-domain, three-dimensional, collaborative, and intelligent technologies to meet the diverse needs of defense, protection, production, safety, and rescue in vicinagearth space. Its application scenarios include low-altitude security, underwater security, and cross-domain security. The vicinagearth space range covers water, land, and airspace from 1000 meters below sea level to 10000 meters above sea level. In the vicinagearth security system, hyperspectral images are an important input data. Compared with natural images, hyperspectral images have higher spectral resolution and can provide more spectral feature information in the fine identification of various ground objects. Hyperspectral image anomaly detection technology analyzes the spectral characteristics of each pixel in the hyperspectral image to identify abnormal regions that are significantly different from the normal background. It not only has important significance in precision agriculture, urban management, and public security, but also has broad application prospects in environmental monitoring, search and rescue, and military reconnaissance. Since the accuracy of anomaly detection directly affects the reliability of the analysis results of hyperspectral images, it is crucial to develop efficient hyperspectral image anomaly detection methods. Compared with natural image anomaly detection, hyperspectral image anomaly detection has unique task characteristics: (1) Hyperspectral images have high spectral resolution and complex spectral information, with high correlation and redundancy. (2) Abnormal regions are usually sparse and have weak spectral performance, and are easily disturbed by complex backgrounds and noise, so the detection algorithm needs to have high-precision spectral resolution and anti-interference ability. (3) Due to the complexity of the data collection process, the difficulty of sample labeling, and the high cost of manual labor, high-quality hyperspectral anomaly detection datasets are scarce. These characteristics make the hyperspectral image anomaly detection task highly complex, so it is of great significance to develop efficient and accurate anomaly detection methods. Currently, mainstream hyperspectral image anomaly detection techniques mainly rely on advanced spectral analysis and machine learning techniques. According to different anomaly detection strategies, related research work can be divided into four categories: statistical-based, data representation-based, data decomposition-based, and deep learning-based:

[0003] The first category is statistical-based methods. However, statistical-based methods are usually simple and fast, but they require the assumption that data follows a Gaussian distribution, and they are not robust to background noise in data, making it difficult to handle complex scenarios.

[0004] The second category is data-representation-based methods. However, these methods introduce nonlinear concepts and local background calculations, which can reduce interference from complex backgrounds. But they involve significant computation and are highly sensitive to background noise, resulting in a higher false alarm rate.

[0005] The third category is based on data decomposition methods. While data decomposition methods can extract background and anomaly information more precisely and reduce the impact of background noise, they may overlook some details in the background, leading to less accurate decomposition results and higher computational complexity.

[0006] The fourth category is deep learning-based methods. However, while deep learning-based methods can automatically extract effective features and effectively handle complex scenes, most current methods only focus on the suppression effect of one-dimensional masks on anomaly reconstruction, ignoring the interrelationship between spatial and background mask structures. This may lead to the loss of image details and reduce detection efficiency.

[0007] Therefore, it is necessary to improve one or more of the problems existing in the above-mentioned related technical solutions.

[0008] It should be noted that this section is intended to provide background or context for the technical solutions of this disclosure as set forth in the claims. The description herein does not constitute an admission that it is prior art simply because it is included in this section. Summary of the Invention

[0009] The purpose of this disclosure is to provide a hyperspectral spatial-spectral mask variational autoencoder anomaly detection method for near-site security, thereby overcoming at least to some extent one or more problems caused by the limitations and defects of related technologies.

[0010] According to embodiments of this disclosure, a hyperspectral spatial-spectral mask variational autoencoder anomaly detection method for near-site security is provided, the method comprising:

[0011] The hyperspectral image is spatially masked and filled with pixel values ​​that follow a Gaussian distribution to obtain a hyperspectral image with the spatial mask applied.

[0012] The hyperspectral image with spatial masking is reduced in dimension and segmented to obtain several image patches of equal size;

[0013] For each target band in each image patch, its adjacent bands are obtained as input, and a spectral mask is applied to the target band to obtain the spatial-spectral mask band.

[0014] The VAE reconstruction network comprising two layers of encoders and two layers of decoders is used to reconstruct the spatial-spectral mask band in combination with adjacent bands to obtain a reconstructed image; wherein the first layer of the encoder and the decoder of the VAE reconstruction network adopts an exponential linear unit activation function, and the reconstruction loss is calculated by the mean square error and the Kullback-Leibler divergence.

[0015] A preliminary residual image is generated according to the comparison between the reconstructed image and the hyperspectral image, and a domain transformation recursive filtering process is performed on the residual image to obtain a filtered residual image.

[0016] The LRX method is used to process the filtered residual image to obtain an abnormal pixel distribution map.

[0017] Further, in the step of performing spatial mask processing on the hyperspectral image and filling in pixel values subject to a Gaussian distribution to obtain a hyperspectral image to which a spatial mask is applied, the step comprises:

[0018] The hyperspectral image is divided into a plurality of equal square blocks, a part of the blocks is randomly selected, and a mask area is determined in the selected blocks by a four-way random iterative expansion method until the mask area reaches a preset range.

[0019] The pixels in the mask area are replaced with random pixel values subject to a Gaussian distribution to obtain a hyperspectral image to which a spatial mask is applied.

[0020] Further, the preset spatial mask area is 30 to 50 groups, and the value range of the number of mask pixels in each area is 15 to 32.

[0021] The mean of the Gaussian distribution is 0, the variance is 1, and the upper and lower limits are the median and minimum values of the original image pixels.

[0022] Further, in the step of performing dimensionality reduction and segmentation on the hyperspectral image to which a spatial mask is applied to obtain a plurality of equal image blocks, the step comprises:

[0023] The hyperspectral image to which a spatial mask is applied is subjected to K-means dimensionality reduction, and the reduced image is cropped into a plurality of equal image blocks using a sliding window.

[0024] Further, the expression of the total loss of the VAE reconstruction network is:

[0025] (1)

[0026] wherein, p N is the number of square blocks, MSE MSE is the mean square error of the original pixel and the reconstructed pixel, KL KL is the Kullback-Leibler divergence of the original pixel and the reconstructed pixel. KL ​

[0027] Further, in the step of generating a preliminary residual image according to the comparison between the reconstructed image and the hyperspectral image, and performing domain transformation recursive filtering on the residual image to obtain a filtered residual image, the step comprises:

[0028] The preliminary residual image is obtained by comparing the hyperspectral image and the reconstructed image, and the preliminary residual image is converted to a transformed domain by domain transformation to obtain a transformed residual image;

[0029] A recursive filter is applied to the transformed residual image in the transformed domain to obtain a result of recursive filtering;

[0030] The result of recursive filtering is converted back to the original space by inverse transformation to obtain the filtered residual image.

[0031] Further, the process of converting the preliminary residual image to the transformed domain by domain transformation to obtain the transformed residual image comprises:

[0032] (2)

[0033] wherein, is the transformed residual image, is a domain transformation operation;

[0034] The recursive filtering process comprises:

[0035] (3)

[0036] wherein, is a result of the i-th iteration, is a recursive filter, denotes a convolution operation, is a result of the (i-1)-th iteration, and the initial condition is ;

[0037] The process of converting the result of recursive filtering back to the original space by inverse transformation comprises:

[0038] (4)

[0039] wherein, is the filtered residual image, is an inverse transformation operation.

[0040] Further, the number of iterations of the recursive filter is 3.

[0041] The technical scheme provided by the embodiments of the present disclosure can include the following beneficial effects:

[0042] ​​In the embodiments of the present disclosure, by using the hyperspectral space-spectrum mask variational auto-encoding anomaly detection method for ground-based security, on the one hand, the designed space-spectrum joint mask technology makes full use of the multi-dimensional information of space and spectrum, improves the difficulty of effectively reconstructing the abnormal area in the network reconstruction process, makes the network pay more attention to the normal spatial and spectral features when learning the image data distribution, and significantly improves the detection ability of complex abnormal scenes. On the other hand, the designed VAE reconstruction network and the mask combination mechanism make full use of the advantages of VAE probability modeling and the advantages of mask attention mechanism, improve the accuracy and generalization ability of anomaly detection, and strengthen the explainability of the model. And the designed residual image processing method uses the domain transformation recursive filter to effectively alleviate the noise problem caused by the complex background, and realizes higher detection accuracy and robustness. BRIEF DESCRIPTION OF DRAWINGS

[0043] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and serve to explain the principles of the present disclosure. It is readily apparent to one of ordinary skill in the art that the accompanying drawings in the following description merely represent several embodiments of the present disclosure and that additional embodiments of the present disclosure can be obtained from these drawings without creative efforts.

[0044] Figure 1 A step diagram of the hyperspectral space-spectrum mask variational auto-encoding anomaly detection method for ground-based security in an exemplary embodiment of the present disclosure is shown.

[0045] Figure 2 A specific flowchart of the hyperspectral space-spectrum mask variational auto-encoding anomaly detection method for ground-based security in an exemplary embodiment of the present disclosure is shown.

[0046] Figure 3 A whole framework diagram of the hyperspectral space-spectrum mask variational auto-encoding anomaly detection method for ground-based security in an exemplary embodiment of the present disclosure is shown.

[0047] Figure 4 A ROC curve comparison diagram of the method proposed in the present application and other methods on all data sets in an exemplary embodiment of the present disclosure is shown.

[0048] Figure 5 A visualization comparison diagram of the present application on all data sets in an exemplary embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0049] Example implementations are now described with reference to the drawings. Example implementations can, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these implementations are provided so that this disclosure will be thorough and complete, and will fully convey the concept of example implementations to those skilled in the art. The described features, structures, or characteristics can be combined in one or more implementations in any suitable manner.

[0050] In addition, the drawings are merely schematic and are not drawn to scale. Identical or similar elements are denoted by the same reference numerals throughout the various figures, where like reference numerals indicate like elements. Some of the blocks in the drawings are function blocks that represent functions implemented by one or more circuits, software, or combinations thereof.

[0051] A hyperspectral spatial-spectral mask variational auto-encoding anomaly detection method for ground proximity security is provided in the present example implementation. Referring to FIG. 1, the hyperspectral spatial-spectral mask variational auto-encoding anomaly detection method for ground proximity security can include: Figure 1

[0052] Step S101: Perform spatial mask processing on the hyperspectral image, and fill in pixel values conforming to a Gaussian distribution to obtain a hyperspectral image to which a spatial mask is applied;

[0053] Step S102: Perform dimension reduction and segmentation on the hyperspectral image to which the spatial mask is applied to obtain a plurality of equal-sized image blocks;

[0054] Step S103: According to each target band in each image block, obtain its adjacent bands as input, and apply a spectral mask to the target band to obtain a spatial-spectral mask band;

[0055] Step S104: Use a VAE reconstruction network containing two layers of encoders and two layers of decoders to reconstruct the spatial-spectral mask band in combination with the adjacent bands to obtain a reconstructed image; wherein the first layer of the encoder and the decoder of the VAE reconstruction network adopts an exponential linear unit activation function, and the reconstruction loss is jointly calculated by a mean square error and a Kullback-Leibler divergence;

[0056] Step S105: Generate a preliminary residual image according to a comparison between the reconstructed image and the hyperspectral image, and perform domain transformation recursive filtering processing on the residual image to obtain a filtered residual image;

[0057] Step S106: Use the LRX method to process the filtered residual image to obtain an anomaly pixel distribution map.

[0058] ​By the above hyperspectral spatial-spectral mask variational auto-encoding anomaly detection method for ground-based security, on the one hand, the designed spatial-spectral joint mask technology fully utilizes the multi-dimensional information of space and spectrum, improves the difficulty of effectively reconstructing the abnormal area in the network during the reconstruction process, makes the network pay more attention to the normal spatial and spectral features when learning the image data distribution, and significantly improves the detection ability of complex abnormal scenes. On the other hand, the designed VAE reconstruction network and the mask combination mechanism fully utilize the advantages of VAE probability modeling and mask attention mechanism, improve the accuracy and generalization ability of anomaly detection, and strengthen the explainability of the model. And the designed residual image processing method uses the domain transformation recursive filter to effectively alleviate the noise problem caused by complex background, and realizes higher detection accuracy and robustness.

[0059] In the following, reference will be made to Figures 1 to 5 The above hyperspectral spatial-spectral mask variational auto-encoding anomaly detection method for ground-based security in the present example embodiment will be described in more detail.

[0060] In step S101, the hyperspectral image is subjected to spatial mask processing, and the pixel values conforming to the Gaussian distribution are filled to obtain the hyperspectral image subjected to spatial mask.

[0061] Specifically, the hyperspectral image is divided into several equal square blocks, a part of the blocks are randomly selected, and the mask area is determined in the selected blocks by a four-way random iterative expansion method until the mask area reaches a preset range.

[0062] The pixels in the mask area are replaced with random pixel values conforming to the Gaussian distribution to obtain the hyperspectral image subjected to spatial mask.

[0063] Further, the preset spatial mask area is 30 to 50 groups, and the value range of the number of mask pixels in each area is 15 to 32.

[0064] The mean of the Gaussian distribution is 0, the variance is 1, and the upper and lower limits are the median and minimum value of the original image pixels.

[0065] More specifically, when spatial mask is performed, the image is divided into equal square regions, and of them are randomly selected to apply spatial mask. Then, a starting pixel point is randomly selected from each selected region, and the mask pixels are determined by a four-way random iterative expansion method until the mask area in the region reaches , The random selection is from the integer interval . In order to prevent the model from identifying the spatial mask area as a new abnormal area, the mask area needs to be filled with random pixel values conforming to the Gaussian distribution .

[0066] In step S102, the hyperspectral image to which the spatial mask is applied is reduced in dimension and segmented to obtain a plurality of equal-sized image blocks.

[0067] Specifically, the hyperspectral image to which the spatial mask is applied is reduced in dimension by K-means, and the reduced image is cropped into a plurality of equal-sized image blocks by a sliding window.

[0068] More specifically, when performing spectral masking, first, the spatial mask image is reduced in dimension by K-means, and after obtaining the reduced image, the image is segmented into a plurality of patches by a sliding window. Then, each band of each patch is reconstructed in turn. In the reconstruction process, the adjacent bands of the band to be reconstructed are taken as the network input, and the band to be reconstructed itself is masked and does not participate in the reconstruction process.

[0069] In steps S103 and S104, according to each target band in each image block, the adjacent bands thereof are taken as input, and the target band is subjected to spectral masking to obtain a spatial-spectral mask band; the spatial-spectral mask band is reconstructed by a VAE reconstruction network comprising two layers of encoders and two layers of decoders, in combination with the adjacent bands, to obtain a reconstructed image; wherein the first layers of the encoders and the decoders of the VAE reconstruction network all adopt exponential linear unit activation functions, and the reconstruction loss is jointly calculated by mean square error and Kullback-Leibler divergence.

[0070] Specifically, the network comprises two layers of encoders and two layers of decoders to more deeply capture and extract feature information of data. The first layers of the encoders and the decoders all adopt exponential linear unit activation functions. The reconstruction loss of the VAE is calculated using mean square error and Kullback-Leibler divergence, and the total loss formula is:

[0071] (1)

[0072] In steps S105 and S106, a preliminary residual image is generated by comparing the reconstructed image and the hyperspectral image, and the residual image is subjected to domain transformation recursive filtering to obtain a filtered residual image; the filtered residual image is subjected to LRX method to obtain an abnormal pixel distribution map.

[0073] Specifically, the residual image processing method: by comparing the original image and the reconstructed image, a preliminary residual image is obtained and is converted to a transformed domain by domain transformation:

[0074] (2)

[0075] wherein, for the transformed residual image, is the domain transform operation. Then a recursive filter is applied in the transformed domain, and the recursive filtering process is:

[0076] (3)

[0077] where, is the result after the th iteration, is the filter, denotes the convolution operation, and the initial condition is The result of the recursive filtering is converted back to the original space by inverse transform:

[0078] (4)

[0079] where, is the filtered residual image, is the inverse transform operation. Finally, the LRX algorithm is used to detect the filtered residual image to generate the abnormal pixel distribution map.

[0080] In one specific embodiment, as shown in Figure 2 , is the specific flow chart of the hyperspectral spatial-spectral mask variational auto-encoding anomaly detection method for ground proximity security; as shown in Figure 3 , is the overall framework diagram of the hyperspectral spatial-spectral mask variational auto-encoding anomaly detection method for ground proximity security. The implementation steps of the present application are as follows:

[0081] 1. Prepare the data set

[0082] Establish the data set required for training and testing:

[0083] (1) Airport-Beach-Urban (ABU) data set: collected by the airborne visible / infrared imaging spectrometer (AVIRIS) and the reflective optical system imaging spectrometer (ROSIS-03) sensors. Six representative images were used in the experiment, with sizes of 100 × 100 × 205, 150 × 150 × 188, 100 × 100 × 193, 100 × 100 × 188, 150 × 150 × 102, and 100 × 100 × 191, respectively.

[0084] (2) Salinas data set: collected by the AVIRIS sensor. The spatial resolution is 3.7 meters, with a total of 204 spectral bands and a size of 180 × 180 pixels.

[0085] (3) Lakeshore dataset: collected by SAMSON sensor. The spatial resolution is 10 meters, with 156 spectral bands, and the size is 200 x 200 pixels.

[0086] The model is trained in an unsupervised manner, and all datasets are used as both training and testing datasets.

[0087] 2. Design of VAE reconstruction network model based on spatial-spectral joint mask

[0088] Design of spatial-spectral joint mask, VAE reconstruction network, and residual image processing method to form a VAE network model based on spatial-spectral joint mask.

[0089] Spatial-spectral joint mask: when performing spatial mask, the image is divided into equal square regions, and regions are randomly selected to apply spatial mask. Then, a starting pixel point is randomly selected from each selected region, and the mask pixels are determined by four-way random iterative expansion until the mask area in the region reaches , Randomly selected from the integer interval . To prevent the model from identifying the spatial mask region as a new abnormal region, the mask region needs to be filled with random pixel values following a Gaussian distribution ; when performing spectral mask, first perform K-means dimensionality reduction on the spatial mask image, and then divide the image into several patches by sliding window after obtaining the reduced image. Then, each band of each patch is reconstructed in turn. In the reconstruction process, the adjacent bands of the band to be reconstructed are used as the network input, while the band to be reconstructed itself is masked and does not participate in the reconstruction process.

[0090] VAE reconstruction network: this network contains two layers of encoder and two layers of decoder to capture and extract feature information of data more deeply. The first layer of the encoder and the decoder uses exponential linear unit activation function. The reconstruction loss of VAE is calculated using mean square error and Kullback-Leibler divergence, and the total loss formula is:

[0091] (1)

[0092] Residual image processing method: by comparing the original image and the reconstructed image, the preliminary residual image is obtained, and it is converted to the transform domain by domain transformation:

[0093] (2)

[0094] where for the transformed residual image, is a domain transform operation. Then a recursive filter is applied in the transformed domain, and the recursive filtering process is:

[0095] (3)

[0096] wherein, is the result after the th iteration, is the filter, denotes the convolution operation, and the initial condition is The result after the recursive filtering is converted back to the original space by inverse transform:

[0097] (4)

[0098] wherein, is the filtered residual image, is an inverse transform operation. Finally, the LRX algorithm is used to detect the filtered residual image to generate the abnormal pixel distribution map.

[0099] 3. Experimental conditions

[0100] All experiments of the present application are completed on a server configured with Intel(R) Xeon(R) Silver 4210R CPU and 1 piece of 24GB NVIDIA Geforce RTX 4090 GPU, the system environment is Ubuntu 22.04.2, the algorithm is implemented by using Python language and relies on Pytorch library, wherein the used Python version is 3.9.19, and the Pytorch version is 1.13.1.

[0101] The training and test data used in the experiment are from public datasets Airport-Beach-Urban, Salinas and Lakeshore, and the specific content of the dataset has been introduced in the foregoing.

[0102] 4. Experimental content

[0103] According to the steps given in the specific embodiment, the results of the network model are shown in the following table.

[0104] Table 1 AUC comparison results of the present application on all datasets

[0105]

[0106] Table 1 is the AUC comparison results of the method proposed in the present application and other methods on all datasets, and the bold font represents the best result. Figure 4is a ROC curve comparison chart of the method proposed in this application and other methods on all data sets. Figure 5 is a visualization comparison chart of this application on all data sets. As shown in the chart, the detection performance of the hyperspectral spatial-spectral mask variational auto-encoding anomaly detection method for on-site security-oriented in this application has reached an advanced level.

[0107] Through the above-mentioned hyperspectral spatial-spectral mask variational auto-encoding anomaly detection method for on-site security-oriented, on the one hand, the designed spatial-spectral joint mask technology makes full use of the multi-dimensional information of space and spectrum, improves the difficulty of effectively reconstructing the abnormal area in the network reconstruction process, makes the network pay more attention to the normal spatial and spectral features when learning the image data distribution, and significantly improves the detection ability of complex abnormal scenes. On the other hand, the designed VAE reconstruction network and mask combination mechanism make full use of the advantages of VAE probability modeling and mask attention mechanism, improve the accuracy and generalization ability of anomaly detection, and strengthen the explainability of the model. And the designed residual image processing method effectively alleviates the noise problem caused by complex background through the domain transformation recursive filter, realizes higher detection accuracy and robustness.

[0108] It should be understood that the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise" and the like in the above description indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the embodiments of the present disclosure and simplifying the description, and therefore cannot be understood as indicating or implying that the indicated device or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the embodiments of the present disclosure.

[0109] In addition, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the embodiments of the present disclosure, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0110] In the embodiments of the present disclosure, unless specifically defined and limited otherwise, the terms "mount", "connect", "connection", "fixed", and the like should be understood broadly, for example, can be fixed connection, can also be detachable connection, or integrated; can be mechanical connection, can also be electrical connection; can be direct connection, can also be indirect connection through an intermediate medium, can be internal communication of two elements or interaction relationship between two elements. For those skilled in the art, the specific meanings of the above terms in the present disclosure can be understood according to the specific circumstances.

[0111] In the embodiments of the present disclosure, unless specifically defined and limited otherwise, the first feature "on" or "under" the second feature can include that the first and second features are in direct contact, or that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, the first feature "on", "above" and "over" the second feature includes that the first feature is directly above and obliquely above the second feature, or only indicates that the horizontal height of the first feature is higher than that of the second feature. The first feature "under", "below" and "under" the second feature includes that the first feature is directly below and obliquely below the second feature, or only indicates that the horizontal height of the first feature is less than that of the second feature.

[0112] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present disclosure. In the specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in the specification.

[0113] Other embodiments of the present disclosure will be apparent to those skilled in the art upon consideration of the specification and practice of the applications disclosed. The present application is intended to cover any variations, uses or adaptive changes of the present disclosure that follow the general principles of the present disclosure and include known or customary technical means in the art not disclosed by the present disclosure. The specification and examples are only considered as exemplary, and the true scope and spirit of the present disclosure are indicated by the appended claims.

Claims

1. A hyperspectral air spectrum mask variational auto-encoding anomaly detection method for ground-oriented security, characterized in that, The method comprises: spatial mask processing is performed on the hyperspectral image, and pixel values conforming to a Gaussian distribution are filled to obtain a hyperspectral image to which a spatial mask is applied; dimensionality reduction and segmentation are performed on the hyperspectral image to which the spatial mask is applied to obtain a plurality of equal-sized image blocks; according to each target waveband in each image block, an adjacent waveband thereof is obtained as input, and a spectral mask is applied to the target waveband to obtain a spatial-spectral mask waveband; the spatial-spectral mask waveband is reconstructed by using a VAE reconstruction network comprising two layers of encoders and two layers of decoders, in combination with the adjacent waveband, to obtain a reconstructed image; wherein the first layers of the encoders and the decoders of the VAE reconstruction network adopt exponential linear unit activation functions, and a reconstruction loss is jointly calculated by a mean square error and a Kullback-Leibler divergence; a preliminary residual image is generated by comparing the reconstructed image and the hyperspectral image, and domain transformation recursive filtering is performed on the residual image to obtain a filtered residual image; the filtered residual image is processed by using an LRX method to obtain an abnormal pixel distribution map.

2. The hyperspectral and spectral mask variational autoencoder anomaly detection method for ground-oriented security according to claim 1, characterized in that, In the step of performing spatial mask processing on the hyperspectral image and filling pixel values conforming to a Gaussian distribution to obtain a hyperspectral image to which a spatial mask is applied, the step comprises: the hyperspectral image is divided into a plurality of equal-sized square blocks, a part of the blocks are randomly selected, and a mask area is determined in the selected blocks by a four-way random iterative expansion method until the mask area reaches a preset range; pixels in the mask area are replaced with random pixel values conforming to a Gaussian distribution to obtain the hyperspectral image to which the spatial mask is applied.

3. The hyperspectral and spectral mask variational autoencoder anomaly detection method for ground-oriented security according to claim 2, characterized in that, The preset spatial mask area is 30 to 50 groups, and the value range of the number of mask pixels in each area is 15 to 32; the Gaussian distribution has a mean value of 0 and a variance of 1, and upper and lower limits of the original image pixel median value and the minimum value.

4. The hyperspectral and spectral mask variational autoencoder anomaly detection method for ground-oriented security according to claim 3, characterized in that, In the step of performing dimensionality reduction and segmentation on the hyperspectral image to which the spatial mask is applied to obtain a plurality of equal-sized image blocks, the step comprises: K-means dimensionality reduction is performed on the hyperspectral image to which the spatial mask is applied, and a sliding window is used to crop the dimensionally reduced image into a plurality of equal-sized image blocks.

5. The hyperspectral and spectral mask variational autoencoder anomaly detection method for ground-oriented security according to claim 4, characterized in that, The expression of the total loss of the VAE reconstruction network is: （1） wherein, p is the number of pixels within the square block, MSE is the mean square error of the original pixel and the reconstructed pixel, KL is the KL divergence of the original pixel and the reconstructed pixel.

6. The hyperspectral and spectral mask variational autoencoder anomaly detection method for ground-oriented security according to claim 5, characterized in that, In the step of generating a preliminary residual image by comparing the reconstructed image and the hyperspectral image, and performing domain transformation recursive filtering on the residual image to obtain a filtered residual image, the step comprises: a preliminary residual image is obtained by comparing the hyperspectral image and the reconstructed image, and the preliminary residual image is converted to a transformed domain by domain transformation to obtain a transformed residual image; a recursive filter is applied to the transformed residual image in the transformed domain to obtain a result of recursive filtering; the result of recursive filtering is converted back to the original space by inverse transformation to obtain the filtered residual image.

7. The hyperspectral aerial spectrum mask variational autoencoder anomaly detection method for ground-oriented security according to claim 6, characterized in that, The process of converting the preliminary residual image to the transformed domain by domain transformation to obtain the transformed residual image is: （2） wherein, is the transformed residual image, is a domain transform operation; The recursive filtering process is: （3） in, For the first The result after the iteration For recursive filters, This represents the convolution operation. For the first The result after -1 iterations, with the initial conditions as follows: ; The process of converting the result of recursive filtering back to the original space by inverse transformation is: （4） wherein, is the filtered residual image, is the inverse transform operation.

8. The hyperspectral and spectral mask variational autoencoder anomaly detection method for ground-oriented security according to claim 6, characterized in that, The number of iterations of the recursive filter is 3.

Images