Hyperspectral anomaly detection method based on Taylor decomposition and differential multi-core initial network

Abstract

The invention discloses a hyperspectral anomaly detection method based on Taylor decomposition and a differential multi-core initial network, and the method comprises the steps: firstly, loading an original hyperspectral image, and obtaining an average spectral curve Xmean; determining the average value of the first 5% maximum values in the spectral angular distance between the ith test pixel and the Xmean, determining the average value of the remaining values, recording the position of the first 5% maximum value as s, and recording the positions except s as t; obtaining an optimal preprocessing order v according to the difference between the average spectrum curve of the nth-order decomposed hyperspectral image at the position s and the average spectrum curve of the original hyperspectral image and the difference between the average spectrum curve of the nth-order decomposed hyperspectral image at the position t and the average spectrum curve of the original hyperspectral image; therefore, an optimal preprocessed hyperspectral image Hv is obtained; obtaining a difference feature map Y and a weakening diagonal feature map P through a multi-kernel difference initial network; adding Y and P, and repeating for four times to obtain a pure background image Z; and subtracting the background image Z from the original hyperspectral image to obtain a final detection result G.

Description

Technical Field

[0001] The present invention belongs to the field of image processing, and in particular relates to a hyperspectral anomaly detection method based on Taylor decomposition and differential multi-kernel initial network. Background Art

[0002] A hyperspectral image can be viewed as a 3D cube, and each band of the hyperspectral image can be understood as a two-dimensional picture to express the spatial information of the image. When observing each pixel, each pixel is composed of a spectral curve, which expresses the unique spectral information of the object from different spectral bands. Therefore, HSI has a unique advantage in scene understanding.

[0003] In recent years, deep learning-based methods have been widely used in hyperspectral anomaly detection. Li et al. first expanded the sample capacity, comparing pixels generated by a CNN network with pseudo-labeled data and using similarity measurements to distinguish between background and target. However, without labels, the ability of CNN to extract deep features decreases. Therefore, Xiang et al. proposed a method based on a guided autoencoder network (GAED). This method uses a guidance image generated from spectral features to act on the autoencoder network to generate a clean background. The error from the original image is the abnormal target. Wang et al. proposed a hyperspectral anomaly detection method based on an adaptive weighted loss function (Auto-AD). During the training process, the weight of potential abnormal pixels is reduced and the detection results are obtained by reconstructing the error. Jiang et al. proposed a low-rank embedded network (LREN). Using low-order optimization ideas, a dictionary with greater diversity is constructed to construct the background, and the residual is used to obtain the detection results. Fu et al. proposed a plug-and-play denoising regularized anomaly detection method (DeCNNAD). This method uses a plug-and-play prior to construct a dictionary and employs clustering to eliminate anomalous pixels as much as possible. Then, based on the CNN, an effective convolutional neural network denoiser is used to reconstruct the low-rank representation, resulting in a clean background after denoising. However, deep learning-based methods typically reconstruct the HSI globally, somewhat ignoring the local information surrounding the anomalous pixels. Using only global information inevitably increases the risk of missed detections. Summary of the Invention

[0004] In view of this, the main purpose of the present invention is to provide a hyperspectral anomaly detection method based on Taylor decomposition and differential multi-kernel initial network.

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

[0006] A hyperspectral anomaly detection method based on Taylor decomposition and differential multi-kernel initial network is proposed. The method is as follows:

[0007] Step 1: Load the original hyperspectral image H to obtain the average spectral curve X of the original hyperspectral image mean , according to the average spectrum curve X mean Determine the i-th test pixel With X mean Spectral angular distance Sure The average of the top 5% largest values Sure The average of the remaining values The position where the top 5% maximum value is obtained is recorded as s, and the position other than s is recorded as t;

[0008] Step 2: Decompose the average spectrum curve X according to the order n of Taylor decomposition. mean Decompose;

[0009] Step 3: Determine the difference between the average spectral curve of the original hyperspectral image at position s after the nth order decomposition and the average spectral curve of the original hyperspectral image Determine the difference between the average spectral curve of the hyperspectral image at position t after the nth order decomposition and the average spectral curve of the original hyperspectral image

[0010] Step 4: Utilization and The maximum ratio is obtained, and the optimal preprocessing order v is obtained. According to the preprocessing order v, the original hyperspectral image H is optimally preprocessed to obtain the hyperspectral image H of the vth order Taylor decomposition. v ;

[0011] Step 5: H v Send it to the multi-core differential initial network, use the differential multi-core initial module to obtain the differential feature map Y, and use the weakened diagonal module to obtain the weakened diagonal feature map P;

[0012] Step 6: Add Y and P, repeat four times, and get the pure background image Z;

[0013] Step 7: Subtract the background image Z from the original hyperspectral image to obtain the final detection result G.

[0014] Preferably, the step 1 comprises the following steps:

[0015] (101) Select the i-th test pixel According to the following formula, according to X mean and Determine spectral angular distance for

[0016]

[0017] (102) Confirm The average of the top 5% largest values Sure The average of the remaining values The position where the top 5% maximum value is obtained is recorded as s, and the position other than s is recorded as t.

[0018] Preferably, the step 2 comprises the following steps:

[0019] (201) According to the following formula, Taylor decomposition is used to calculate the average spectrum curve X mean Decompose into

[0020]

[0021] Among them, f'[a] is the first-order difference of the spectral curve at point a, f"[a] is the second-order difference of the spectral curve at point a, and so on. The original hyperspectral image is recorded as H according to the order of Taylor decomposition. n , where n represents the hyperspectral image after decomposition at the order of n.

[0022] Preferably, the step three comprises the following steps:

[0023] (301) According to the following formula, the difference between the average spectral curve of the original hyperspectral image at position s after the n-th order decomposition and the average spectral curve of the original hyperspectral image is determined: for

[0024]

[0025] in, represents the average value of the spectral curve of the hyperspectral image at position s after the nth order decomposition;

[0026] (302) According to the following formula, the difference between the average spectral curve of the original hyperspectral image at position t after the n-th order decomposition and the average spectral curve of the original hyperspectral image is determined: for

[0027]

[0028] in, It represents the average value of the spectral curve of the original hyperspectral image at position t after the n-th order decomposition.

[0029] Preferably, the step 4 comprises the following steps:

[0030] (401) According to the following formula, using and The maximum ratio finds the optimal preprocessing order v as

[0031]

[0032] in, Indicates the value of n when the maximum value is achieved during n transformation;

[0033] (402) The original hyperspectral image H is optimally preprocessed according to the preprocessing order v to obtain the hyperspectral image H of the vth order Taylor decomposition. v .

[0034] Preferably, the step five comprises the following steps:

[0035] (501) As follows, using 3×3 depth convolution, we obtain the depth convolution feature map H′ N 'for

[0036] H″ N =DWConv 3×3 (H v ) (6)

[0037] Among them, DWConv 3×3 (·) indicates a 3×3 depthwise convolution operation;

[0038] (502) According to the following formula, multiple depth convolution kernels of different sizes (5×5-3×3, 7×7-3×3 and 9×9-3×3) are used for multi-scale feature extraction to obtain the differential depth feature map F DW1 , F DW2 and F DW3 for

[0039]

[0040] Among them, DWConv 5×5-3×3 (·) indicates that the depth convolution operation is performed using the convolution kernel of size 5×5-3×3, DWConv 7×7-5×5 (·) indicates that the depth convolution operation is performed using the convolution kernel of size 7×7-5×5, DWConv 9×9-7×7 (·) indicates depthwise convolution using a 9×9-7×7 kernel.

[0041] (503) According to the following formula, the three differential depth feature maps are integrated using 1×1 convolution to obtain the final multi-core differential feature map Y:

[0042] Y = Conv 1×1 (F DW1 +F DW2 +F DW3 ) (8)

[0043] Among them, Conv 1×1 Indicates a 1×1 convolution operation;

[0044] (504) According to the following formula, using 7×7 weakened diagonal convolution, the weakened diagonal feature map P is obtained as

[0045] P = WDDWConv 7×7 (H″ N ) (9)

[0046] Among them, WDDWConv 7×7 (·) indicates that the diagonal operation is weakened using a 7×7 convolution kernel with a minimum diagonal value.

[0047] Preferably, the step six comprises the following steps:

[0048] (601) According to the following formula, the result of integrating Y and P using 1×1 convolution is repeated four times to obtain the pure background image Z:

[0049] Z=Conv 1×1 (Y+P) (10).

[0050] Preferably, the step seven comprises the following steps:

[0051] According to the following formula, the final detection result G is

[0052] G=HZ (11).

[0053] Compared to existing technologies, this method fully addresses spectral mixing effects by proposing a hyperspectral anomaly detection method based on adaptive Taylor decomposition and a local multi-scale multi-kernel inception network. Adaptive Taylor decomposition is introduced to preprocess the original image, and a differential multi-kernel inception module is used to enhance the network's ability to resolve multi-scale issues. A weakened diagonal module is also used to reduce the impact of anomalies on background reconstruction. The final anomaly detection result is obtained by subtracting the original hyperspectral image from the purified background image generated by the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is a flow chart of the present invention;

[0055] Figure 2 The original hyperspectral image inputted by the present invention;

[0056] Figure 3 This is a diagram of the hyperspectral anomaly detection result in the present invention; Specific implementation plan

[0057] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific implementations described herein are only used to solve the present invention and are not intended to limit the present invention.

[0058] The embodiment of the present invention provides a hyperspectral anomaly detection method based on Taylor decomposition and differential multi-core initial network, such as Figure 1 As shown, the method is:

[0059] Step 1: Load the original hyperspectral image H and obtain the average spectral curve X of the entire hyperspectral image mean , determine the i-th test pixel With X mean Spectral angular distance Sure The average of the top 5% largest values Sure The average of the remaining values The position where the top 5% maximum value is obtained is recorded as s, and the position other than s is recorded as t;

[0060] In the above solution, step 1 is specifically implemented by the following steps:

[0061] (101) Load the original hyperspectral image to obtain the average spectral curve X of the original hyperspectral image mean , select the test pixel According to the following formula, use X mean and Obtain spectral angular distance for

[0062]

[0063] Specifically, the original hyperspectral image loaded is Champagne Bay, with a resolution of 100×100 pixels, a spatial resolution of 4.4 meters, and 188 bands. The background includes the ocean, forest, and a road, and the abnormal target is an airplane. The average spectral curve X of the original hyperspectral image loaded is mean is {765.1946,892.2119,964.6791,1236.3748,1820.3739,1949.8862,1884.7938,1808.9521,2033.5284,2163.1286,......}, the value range of i is 1 to 10000, {0.3027,0.2743,0.2700,0.2808,0.3068,0.3003,0.2807,0.2912,0.2875,0.2965,......};

[0064] (102) Confirm The average of the top 5% largest values The position where the top 5% maximum value is obtained is recorded as s, and the The average of the remaining values The positions other than s are recorded as t.

[0065] Specifically, is 0.9587, The value of s is 0.2937, and the values ​​of s are (35,19), (40,22), (34,16), (32,30), and (30,19). The values ​​of the other positions are all t.

[0066] Step 2: Use Taylor decomposition to decompose the spectral curve of the original hyperspectral image. The original hyperspectral image is recorded as H according to the order of Taylor decomposition. n , where n represents the hyperspectral image after decomposition at the order of n;

[0067] The step 2 is specifically implemented by the following steps:

[0068] (201) According to the following formula, Taylor decomposition is used to decompose the spectral curve of the original hyperspectral image into

[0069]

[0070] Among them, f'[a] is the first-order difference of the spectral curve at point a, f"[a] is the second-order difference of the spectral curve at point a, and so on. The original hyperspectral image is recorded as H according to the order of Taylor decomposition. n , where n represents the hyperspectral image after decomposition at the order of n.

[0071] Specifically, n is an integer between 1 and 5.

[0072] Step 3: Determine the difference between the average spectral curve of the hyperspectral image at position s after the nth order decomposition and the average spectral curve of the original hyperspectral image Determine the difference between the average spectral curve of the hyperspectral image at position t after the nth order decomposition and the average spectral curve of the original hyperspectral image

[0073] The step three is specifically implemented by the following steps:

[0074] (301) According to the following formula, the difference between the average spectral curve of the hyperspectral image position s after the n-th order decomposition and the average spectral curve of the original hyperspectral image is determined: for

[0075]

[0076] in, represents the average value of the spectral curve of the hyperspectral image at position s after the nth order decomposition;

[0077] Specifically, The size is (0.9432, 0.9108, 0.8643, 0.8186, 0.7359).

[0078] (302) According to the following formula, the difference between the average spectral curve at position t of the hyperspectral image after the n-th order decomposition and the average spectral curve of the original hyperspectral image is determined: for

[0079]

[0080] in, It represents the average value of the spectral curve of the hyperspectral image at position t after the n-th order decomposition.

[0081] Specifically, The size is (0.2891, 0.2637, 0.2114, 0.1989, 0.1793).

[0082] Step 4: Utilization and The maximum ratio is obtained, and the optimal preprocessing order v is obtained, thereby obtaining the optimal preprocessing hyperspectral image H v ;

[0083] The step 4 is specifically implemented by the following steps:

[0084] (401) According to the following formula, using and The maximum ratio finds the optimal preprocessing order v as

[0085]

[0086] in, Indicates the value of n when the maximum value is achieved during n transformation;

[0087] Specifically, v=n=4.

[0088] (402) Obtain the optimal preprocessed hyperspectral image H according to the v-th order Taylor decomposition v .

[0089] Specifically, the optimal preprocessed hyperspectral image H4 is obtained according to the 4th-order Taylor decomposition.

[0090] Step 5: H v It is sent into the multi-core differential initial network, and the differential feature map Y is obtained by using the differential multi-core initial module, and the weakened diagonal feature map P is obtained by using the weakened diagonal module.

[0091] The step 5 is specifically implemented by the following steps:

[0092] (501) As follows, using 3×3 depth convolution, we obtain the depth convolution feature map H′ N 'for

[0093] H″ N =DWConv 3×3 (H v ) (6)

[0094] Among them, DWConv 3×3 (·) indicates a 3×3 depthwise convolution operation;

[0095] (502) According to the following formula, multiple depth convolution kernels of different sizes (5×5-3×3, 7×7-3×3 and 9×9-3×3) are used for multi-scale feature extraction to obtain the differential depth feature map F DW1 , F DW2 and F DW3 for

[0096]

[0097] Among them, DWConv 5×5-3×3 (·) indicates that the depth convolution operation is performed using the convolution kernel of size 5×5-3×3, DWConv 7×7-5×5 (·) indicates that the depth convolution operation is performed using the convolution kernel of size 7×7-5×5, DWConv 9×9-7×7 (·) indicates depthwise convolution using a 9×9-7×7 kernel.

[0098] (503) According to the following formula, the three differential depth feature maps are integrated using 1×1 convolution to obtain the final multi-core differential feature map Y:

[0099] Y = Conv 1×1 (F DW1 +F DW2 +F DW3 ) (8)

[0100] Among them, Conv 1×1 Indicates a 1×1 convolution operation;

[0101] (504) According to the following formula, using 7×7 weakened diagonal convolution, the weakened diagonal feature map P is obtained as

[0102] P = WDDWConv 7×7 (H″ N ) (9)

[0103] Among them, WDDWConv 7×7 (·) indicates that the diagonal operation is weakened using a 7×7 convolution kernel with a minimum diagonal value.

[0104] Step 6. Add Y and P, repeat four times, and get the pure background image Z.

[0105] The step six is ​​specifically implemented by the following steps:

[0106] (601) According to the following formula, the pure background image Z is obtained by integrating Y and P using 1×1 convolution:

[0107] Z=Conv 1×1 (Y+P) (10)

[0108] In the above solution, step seven is specifically implemented by the following steps:

[0109] Step 7: Subtract the background image Z from the original hyperspectral image H to obtain the final detection result G.

[0110] (701) According to the following formula, the final detection result G is

[0111] G=HZ (11)

[0112] Specifically, based on Figure 2 The original hyperspectral image inputted by the present invention is Figure 3 This is the final detection result diagram, and the red position in the diagram is the abnormal target.

Claims

1. A hyperspectral anomaly detection method based on Taylor decomposition and differential multi-kernel initial network, characterized in that: The method is: Step 1: Load the original hyperspectral image H to obtain the average spectral curve X of the original hyperspectral image mean , according to the average spectrum curve X mean Determine the i-th test pixel With X mean Spectral angular distance Sure The average of the top 5% largest values Sure The average of the remaining values The position where the top 5% maximum value is obtained is recorded as s, and the position other than s is recorded as t; Step 2: Decompose the average spectrum curve X according to the order n of Taylor decomposition. mean Decompose; Step 3: Determine the difference between the average spectral curve of the original hyperspectral image at position s after the nth order decomposition and the average spectral curve of the original hyperspectral image Determine the difference between the average spectral curve of the hyperspectral image at position t after the nth order decomposition and the average spectral curve of the original hyperspectral image Step 4: Utilization and The maximum ratio is obtained, and the optimal preprocessing order v is obtained. According to the preprocessing order v, the original hyperspectral image H is optimally preprocessed to obtain the hyperspectral image H of the vth order Taylor decomposition. v ; Step 5: H v Send it to the multi-core differential initial network, use the differential multi-core initial module to obtain the differential feature map Y, and use the weakened diagonal module to obtain the weakened diagonal feature map P; Step 6: Add Y and P, repeat four times, and get the pure background image Z; Step 7: Subtract the background image Z from the original hyperspectral image to obtain the final detection result G.

2. The hyperspectral anomaly detection method based on Taylor decomposition and differential multi-core initial network according to claim 1 is characterized in that: The specific steps of step 1 are as follows: (101) Select the i-th test pixel According to the following formula, according to X mean and Determine spectral angular distance for (102) Confirm The average of the top 5% largest values Sure The average of the remaining values The position where the top 5% maximum value is obtained is recorded as s, and the position other than s is recorded as t.

3. The hyperspectral anomaly detection method based on Taylor decomposition and differential multi-kernel initial network according to claim 1 or 2, characterized in that: The specific steps of step 2 are as follows: (201) According to the following formula, Taylor decomposition is used to calculate the average spectrum curve X mean Decompose into Among them, f'[a] is the first-order difference of the spectral curve at point a, f"[a] is the second-order difference of the spectral curve at point a, and so on. The original hyperspectral image is recorded as H according to the order of Taylor decomposition. n , where n represents the hyperspectral image after decomposition at the order of n.

4. The hyperspectral anomaly detection method based on Taylor decomposition and differential multi-core initial network according to claim 3 is characterized in that: The specific steps of step three are as follows: (301) According to the following formula, the difference between the average spectral curve of the original hyperspectral image at position s after the n-th order decomposition and the average spectral curve of the original hyperspectral image is determined: for in, represents the average value of the spectral curve of the hyperspectral image at position s after the nth order decomposition; (302) According to the following formula, the difference between the average spectral curve of the original hyperspectral image at position t after the n-th order decomposition and the average spectral curve of the original hyperspectral image is determined: for in, It represents the average value of the spectral curve of the original hyperspectral image at position t after the n-th order decomposition.

5. The hyperspectral anomaly detection method based on Taylor decomposition and differential multi-core initial network according to claim 4 is characterized in that: The specific steps of step 4 are as follows: (401) According to the following formula, using and The maximum ratio finds the optimal preprocessing order v as in, Indicates the value of n when the maximum value is achieved during n transformation; (402) The original hyperspectral image H is optimally preprocessed according to the preprocessing order v to obtain the hyperspectral image H of the vth order Taylor decomposition. v .

6. The hyperspectral anomaly detection method based on Taylor decomposition and differential multi-kernel initial network according to claim 5 is characterized in that: The specific steps of step five are as follows: (501) According to the following, use 3×3 depth convolution to obtain the depth convolution feature map H″ N for H″ N =DWConv 3×3 (H v ) (6) Among them, DWConv 3×3 (·) indicates a 3×3 depthwise convolution operation; (502) According to the following formula, multiple depth convolution kernels of different sizes (5×5-3×3, 7×7-3×3 and 9×9-3×3) are used for multi-scale feature extraction to obtain the differential depth feature map F DW1 , F DW2 and F DW3 for Among them, DWConv 5×5-3×3 () indicates that the depth convolution operation is performed using the convolution kernel of size 5×5-3×3, DWConv 7×7-5×5 () indicates that the depth convolution operation is performed using the convolution kernel of size 7×7-5×5, DWConv 9×9-7×7 (·) indicates depthwise convolution using a 9×9-7×7 kernel. (503) According to the following formula, the three differential depth feature maps are integrated using 1×1 convolution to obtain the final multi-core differential feature map Y: Y=Conv 1×1 (F DW1 +F DW2 +F DW3 ) (8) Among them, Conv 1×1 Indicates a 1×1 convolution operation; (504) According to the following formula, using 7×7 weakened diagonal convolution, the weakened diagonal feature map P is obtained as P=WDDWConv 7×7 (H″ N ) (9) Among them, WDDWConv 7×7 (·) indicates that the diagonal operation is weakened using a 7×7 convolution kernel with a minimum diagonal value.

7. The hyperspectral anomaly detection method based on Taylor decomposition and differential multi-kernel initial network according to claim 6 is characterized in that: The specific steps of step six are as follows: (601) According to the following formula, the result of integrating Y and P using 1×1 convolution is repeated four times to obtain the pure background image Z: Z=Conv 1×1 (Y+P) (10)。 8. The hyperspectral anomaly detection method based on Taylor decomposition and differential multi-kernel initial network according to claim 7 is characterized in that: The specific steps of step seven are as follows: According to the following formula, the final detection result G is G=HZ (11).

Images