A hyperspectral anomaly target fast detection method based on compression domain

Abstract

The application discloses a hyperspectral anomaly target fast detection method based on a compression domain, and relates to the field of anomaly target detection. The patent aims to solve the practical application bottleneck problem that it is difficult to simultaneously consider detection accuracy and timeliness in the process of searching for suspected targets under a large field of view by using a hyperspectral image, and designs an anomaly target detection method which can be embedded in a high-dimensional hyperspectral data compression process. The intermediate result of the compression process is used to highlight and detect the anomaly target in the equal dimension of space-spectrum by using a guided filter with a double window structure, and the method has the characteristics of being fast and accurate, and can improve the detection speed while improving the anomaly target detection accuracy.

Description

Technical Field

[0001] This invention belongs to the field of hyperspectral anomalous target detection, and particularly relates to a rapid detection method for hyperspectral anomalous targets based on compressed domain. Background Technology

[0002] Hyperspectral images are three-dimensional data that simultaneously contain two-dimensional spatial information and one-dimensional spectral information. The spatial information describes the location of the target, while the spectral information reflects the target's unique physical and chemical characteristics. The "image-spectrum integration" feature of hyperspectral images gives them a unique advantage in distinguishing different ground features, and they are widely used in military and civilian fields such as battlefield target detection and mineral exploration.

[0003] In recent years, researchers have proposed many methods for detecting anomalous targets in hyperspectral images. Existing methods can be mainly divided into two categories. The first is statistical methods, which establish different spectral distribution models based on statistics to describe the high-dimensional distribution characteristics of hyperspectral data, thereby separating anomalous spectral information that is inconsistent with the distribution model as anomalous targets. The most commonly used method is the Reed-Xiaoli (RX) detector, which assumes that the spectral information in hyperspectral images follows a multivariate Gaussian distribution. It describes the degree of anomalousness by calculating the Mahalanobis distance between the target spectrum and the background spectrum. This algorithm has low computational complexity and has become a widely used benchmark algorithm in this field. Based on this, researchers have proposed extended algorithms such as Local-RX and Kernel-RX. However, due to the influence of complex and diverse land cover in real-world scenes, the background spectral distribution is difficult to satisfy the Gaussian distribution assumption, affecting the detection performance of such algorithms. The second approach utilizes the low-rank or sparse characteristics of anomalous spectra, improving the performance of anomalous target detection through iterative optimization. However, this approach has high computational complexity and time consumption. For example, anomaly detection methods based on low-rank sparse representation overcome the influence of chaotic background spectral distribution by constructing local sparse constraints.

[0004] Among the aforementioned methods, detection algorithms based on statistical models are fast but have low accuracy, while those based on the low-rank or sparse characteristics of anomalous target spectra have high accuracy but are computationally time-consuming. Both face the practical application bottleneck of simultaneously achieving both detection accuracy and timeliness in large-field-of-view search tasks using hyperspectral images. Therefore, designing an efficient anomalous target detection method that balances detection capability and speed for high-dimensional, high-volume hyperspectral images is of great significance. To this end, this patent proposes a fast hyperspectral anomalous target detection algorithm framework based on compressed domains, utilizing the intermediate results of the multidimensional hyperspectral data compression process to achieve efficient and rapid detection of anomalous targets in hyperspectral images. Summary of the Invention

[0005] The purpose of this invention is to address the problem that existing hyperspectral image anomaly detection methods cannot simultaneously achieve both timeliness and accuracy. It proposes a fast hyperspectral anomaly detection algorithm framework based on the compressed domain, which can be embedded into the high-dimensional hyperspectral data compression process. A dual-window guided filter is designed to process the high- and low-frequency intermediate results of hyperspectral data compression, highlighting anomalies in both spatial and spectral dimensions. Furthermore, a novel anomaly degree description method is designed to quickly calculate the anomaly degree of each pixel, thereby locating the anomaly.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A rapid detection method for hyperspectral anomalous targets based on compressed domain includes the following steps:

[0008] Step 1: Multidimensional hyperspectral data compression based on Haar wavelet transform;

[0009] Given hyperspectral image data x∈R (M×N×L) The image is input band by band into the compression module, where the image of the i-th band can be represented as f. i Given (x,y), calculate the low-frequency subband LL(X,Y) with low-pass filtering in both horizontal and vertical directions, the horizontal detail subband LH(X,Y) with low-pass filtering in the horizontal direction and high-pass filtering in the vertical direction, the vertical detail subband HL(X,Y) with high-pass filtering in the horizontal direction and low-pass filtering in the vertical direction, and the diagonal detail subband HH(X,Y) with high-pass filtering in both horizontal and vertical directions. The specific calculation process can be expressed by the following formula:

[0010] LL(X,Y)=∑ k,l h(k)·h(l)·f i (x-2k,y-2l) (1)

[0011] LH(X,Y)=∑ k,l h(k)·g(l)·f i (x-2k,y-2l) (2)

[0012] HL(X,Y)=∑ k,l g(k)·h(l)·f i (x-2k,y-2l) (3)

[0013] HH(X,Y)=∑ k,l g(k)·g(l)·f i (x-2k,y-2l) (4)

[0014] Where h(k) and h(l) represent the coefficients of the low-pass filter in the horizontal and vertical directions, and g(k) and g(l) represent the coefficients of the high-pass filter in the horizontal and vertical directions.

[0015] Subsequently, the LL subband was used as the low-frequency subband, and the LH, HL and HH subbands were weighted and fused to become the high-frequency subband.

[0016] Step 2: Band selection of high and low frequency subbands after compression.

[0017] Principal component analysis was used to process the high-frequency and low-frequency subbands separately and select characteristic bands. This process can be described as follows: the compressed high-frequency and low-frequency subbands were reconstructed into matrix I respectively. h ∈R M×N×L with I l ∈R M×N×L The average value of the i-th band can be calculated as follows:

[0018]

[0019] Then copy the average value to the original image size. The average matrix can be further expanded as follows:

[0020] I mean (i,x)=I mean (i,1),x=1,…M·N (6) Then the covariance matrix C of I can be calculated as:

[0021]

[0022] The eigenvalues ​​of the covariance matrix C are E = [e1,…,e2]. L ] and the eigenvector V = [v1, ..., v L The following relationship exists:

[0023] C = VEV T (8)

[0024] Where T represents the transpose operation. We sort the eigenvectors in descending order of their eigenvalues, and select the b eigenvectors corresponding to the first b eigenvalues ​​to construct a new matrix V. b =[v1,…,v b Finally, the original hyperspectral image is projected into the new feature space, as follows:

[0025]

[0026] Finally, the reduction results I for the high-frequency and low-frequency subbands were obtained respectively. hpca ∈R M×N×k with I lpca ∈R M×N×k .

[0027] Step 3: Hyperspectral image anomalies are highlighted based on dual-window guided filters;

[0028] The compressed high- and low-frequency subbands are input band by band into the guidance filter as guidance images p. Then, the mean μ of guidance images p is calculated in two local windows of different sizes. p and variance The formulas for calculating the mean and variance are as follows:

[0029] μ p =mean window (p) (10)

[0030]

[0031] Then, input the same compressed spectral image as the image to be processed, and calculate the covariance σ between the guiding image p and the image to be processed I. ip The calculation formula is as follows:

[0032] σ ip =cov window (I,p) (12)

[0033] The specific parameters a and b used for filtering are determined by the above variables, and the specific formula is as follows:

[0034]

[0035] b=μI-a·μ p (14)

[0036] ε is a regularization parameter to prevent division by zero. According to the output rule, the output image is q = a·p + b.

[0037] Finally, the final output image I is obtained by subtraction. final =q in -q out .

[0038] Step 4: Rapid Anomaly Calculation

[0039] The single-band image processed by the guiding filter is reconstructed according to bands and input into the variable ANO, which is quickly calculated using a diagonal matrix, as designed in this patent. i As I final Recombined pixel vector I ri The abnormal situation is calculated using the following formula:

[0040]

[0041] The resulting detection map has pixel values ​​equal to the anomaly score of each pixel.

[0042] Step 5: Weighted fusion of high and low frequency anomalies

[0043] After statistically analyzing the anomalies in the high-frequency and low-frequency subbands, a weighted sum is calculated. Under different threshold conditions, the detection rate and false alarm rate are calculated. Based on the specified false alarm rate α, the corresponding global threshold T is adaptively determined, and the final detection results are as follows:

[0044]

[0045] Result(x) = 1 indicates that the cell is an abnormal cell, while Result(x) = 0 indicates that the cell is not abnormal.

[0046] Beneficial effects:

[0047] This paper presents a fast hyperspectral anomaly detection architecture in the compressed domain, achieving a hyperspectral anomaly detection algorithm that balances detection performance and speed. The architecture utilizes intermediate results from high-dimensional hyperspectral image compression, namely the low-frequency subband (LL) and the weighted fused high-frequency subbands (LH, HL, and HH), to fully leverage the high-frequency and low-frequency features of the hyperspectral image. The high- and low-frequency subbands are sequentially passed through a dual-window guided filter to more effectively highlight potential anomalies. An efficient method for calculating the anomaly level based on diagonal matrix operations is designed to quantify the anomaly level corresponding to each pixel. Finally, an anomaly detector that weightedly fuses the high- and low-frequency anomaly levels is designed. Experiments show that the proposed compressed-domain-based fast anomaly detection architecture outperforms commonly used methods in terms of detection accuracy and time, meeting the practical application requirements for fast and accurate search of suspected targets in a large field of view. Attached Figure Description

[0048] Figure 1 Here is the flowchart for the baseline RX algorithm;

[0049] Figure 2 This is a flowchart of the data processing method of the present invention;

[0050] Figure 3 This is a schematic diagram of the dual-window guided filter of the present invention;

[0051] Figure 4 This is the ground truth map of the hyperspectral image to be detected;

[0052] Figure 5a The image shows the processing results using the benchmark RX algorithm.

[0053] Figure 5b The image shows the processing result using the method of the present invention. Detailed Implementation

[0054] This invention is achieved through the following technical solution:

[0055] Step 1: Multidimensional hyperspectral data compression based on Haar wavelet transform;

[0056] Given hyperspectral image data x∈R (M×N×L) The image is input band by band into the compression module, where the image of the i-th band can be represented as f. i Given (x,y), calculate the low-frequency subband LL(X,Y) with low-pass filtering in both horizontal and vertical directions, the horizontal detail subband LH(X,Y) with low-pass filtering in the horizontal direction and high-pass filtering in the vertical direction, the vertical detail subband HL(X,Y) with high-pass filtering in the horizontal direction and low-pass filtering in the vertical direction, and the diagonal detail subband HH(X,Y) with high-pass filtering in both horizontal and vertical directions. The specific calculation process can be expressed by the following formula:

[0057] LL(X,Y)=∑ k,l h(k)·h(l)·f i (x-2k,y-2l) (1)

[0058] LH(X,Y)=∑ k,l h(k)·g(l)·f i (x-2k,y-2l) (2)

[0059] HL(X,Y)=∑ k,l g(k)·h(l)·f i (x-2k,y-2l) (3)

[0060] HH(X,Y)=∑ k,l g(k)·g(l)·f i (x-2k,y-2l) (4)

[0061] Where h(k) and h(l) represent the coefficients of the low-pass filter in the horizontal and vertical directions, and g(k) and g(l) represent the coefficients of the high-pass filter in the horizontal and vertical directions.

[0062] Subsequently, the LL subband was used as the low-frequency subband, and the LH, HL and HH subbands were weighted and fused to become the high-frequency subband.

[0063] Step 2: Band selection of high and low frequency subbands after compression.

[0064] Principal component analysis was used to process the high-frequency and low-frequency subbands separately and select characteristic bands. This process can be described as follows: the compressed high-frequency and low-frequency subbands were reconstructed into matrix I respectively. h ∈R M×N×L with I l ∈R M×N×L The average value of the i-th band can be calculated as follows:

[0065]

[0066] Then copy the average value to the original image size. The average matrix can be further expanded as follows:

[0067] I mean (i,x)=I mean (i,1),x=1,…,M·N (6) Then the covariance matrix C of I can be calculated as:

[0068]

[0069] The eigenvalues ​​of the covariance matrix C are E = [e1,…,e2]. L ] and the eigenvector V = [v1, ..., v L The following relationship exists:

[0070] C = VEV T (8)

[0071] Where T represents the transpose operation. We sort the eigenvectors in descending order of their eigenvalues, and select the b eigenvectors corresponding to the first b eigenvalues ​​to construct a new matrix V. b =[v1,…,v b Finally, the original hyperspectral image is projected into the new feature space, as follows:

[0072]

[0073] Finally, the reduction results I for the high-frequency and low-frequency subbands were obtained respectively. hpca ∈R M×N×k with I lpca ∈R M×N×k .

[0074] Step 3: Hyperspectral image anomalies are highlighted based on dual-window guided filters;

[0075] The compressed high- and low-frequency subbands are input band by band into the guidance filter as guidance images p. Then, the mean μ of guidance images p is calculated in two local windows of different sizes. p and variance The formulas for calculating the mean and variance are as follows:

[0076] μ p =mean window (p) (10)

[0077]

[0078] Then, input the same compressed spectral image as the image to be processed, and calculate the covariance σ between the guiding image p and the image to be processed I.ip The calculation formula is as follows:

[0079] σ ip =cov window (I,p) (12) Determine the specific parameters a and b used for filtering through the above variables. The specific formula is as follows:

[0080]

[0081] b=μI-a·μ p (14)

[0082] ε is a regularization parameter to prevent division by zero. According to the output rule, the output image is q = a·p + b.

[0083] Finally, the final output image I is obtained by subtraction. final =q in -q out .

[0084] Step 4: Rapid Anomaly Calculation

[0085] The single-band image processed by the guiding filter is reconstructed according to bands and input into the variable ANO, which is quickly calculated using a diagonal matrix, as designed in this patent. i As I final Recombined pixel vector I ri The abnormal situation is calculated using the following formula:

[0086]

[0087] The resulting detection map has pixel values ​​equal to the anomaly score of each pixel.

[0088] Step 5: Weighted fusion of high and low frequency anomalies

[0089] After statistically analyzing the anomalies in the high-frequency and low-frequency subbands, a weighted sum is calculated. Under different threshold conditions, the detection rate and false alarm rate are calculated. Based on the specified false alarm rate α, the corresponding global threshold T is adaptively determined, and the final detection results are as follows:

[0090]

[0091] Result(x) = 1 indicates that the cell is an abnormal cell, while Result(x) = 0 indicates that the cell is not abnormal.

[0092] Real hyperspectral image processing experiments:

[0093] The following comparative experiments demonstrate the effectiveness of the present invention by comparing the dual-window guided filtering hyperspectral anomaly detection algorithm based on the compressed domain with the benchmark RX anomaly detection algorithm.

[0094] 1. Experimental conditions

[0095] Data: Real hyperspectral image dataset HYDICE urban dataset, image length H=80, width W=100, number of bands B=162;

[0096] Hardware requirements: GPU model GTX4060, CPU model Intel i9-14900K;

[0097] 2. Experiment Content

[0098] Based on the real hyperspectral image dataset Urban Dataset, the benchmark RX processing program and the compressed domain processing program using the method of this invention were run respectively, and the computation time of the two implementations was statistically analyzed. The results are shown in Table 1.

[0099] Table 1. Time Comparison of Rapid Hyperspectral Anomaly Target Detection Method Based on Compressed Domain with RX Benchmark Algorithm

[0100]

[0101] 3. Results Analysis

[0102] Reference Figure 3 A ground truth map to indicate the locations of anomalous targets in the hyperspectral data to be detected. (Refer to...) Figure 5a and Figure 5b The images shown are the results obtained using the RX benchmark processing method and the method used in this patent, respectively. The calculation results obtained by this method are better.

[0103] Table 1 shows the time consumption of the dual-window guided filtering hyperspectral anomaly detection method based on the compressed domain and the RX benchmark algorithm. It can be seen that the method proposed in this invention reduces the running time from 0.121 to 0.083, which improves the detection speed and the detection effect at the same time.

[0104] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A rapid detection method for hyperspectral anomalous targets based on compressed domain, characterized in that, Includes the following steps: Step 1: Multidimensional hyperspectral data compression based on Haar wavelet transform; Step 2: Weighted fusion of compressed high and low frequency subbands and band selection; Step 3: Highlighting hyperspectral image anomalies based on a dual-window guided filter; A dual-window guided filter process is designed, which simultaneously performs guided filtering from two windows of different sizes (inner and outer), and subtracts the results to obtain the final result, including: The compressed high- and low-frequency subbands are input band by band into the guidance filter as guidance images p. Then, the mean value of guidance images p is calculated in two local windows of different sizes. and variance The formulas for calculating the mean and variance are as follows: ； ； Then, input the same compressed spectral image as the image to be processed, I, and calculate the covariance between the guiding image p and the image to be processed I. The calculation formula is as follows: ； The specific parameters a and b used for filtering are determined by the above variables, and the specific formula is as follows: ； ； in To prevent division by zero regularization parameters, the image is output according to the output rules. ; Finally, the final output image is obtained by subtraction. ; Step 4: Rapid anomaly degree calculation; Step 5: Weighted fusion of high and low frequency anomalies.

2. The rapid detection method for hyperspectral anomalous targets based on compressed domain according to claim 1, characterized in that, A fast detection algorithm for hyperspectral anomalous targets based on compressed domain is designed. This algorithm can be embedded in the high-dimensional hyperspectral data compression process. The intermediate results of the compression process are used to highlight and detect anomalous targets in the spatial-spectral dimensions using a guided filter with a dual-window structure. This can significantly improve the detection speed while improving the accuracy of anomalous target detection.

3. A rapid detection method for hyperspectral anomalous targets based on a compressed domain, as described in claim 1 or 2, characterized in that, Hyperspectral image data I The image is decomposed into four sub-bands: LL, LH, HL, and HH using discrete wavelet transform. The LL sub-band is used as the low-frequency sub-band, while the LH, HL, and HH sub-bands are weighted and fused to form the high-frequency sub-band. M and N represent the length and width of the image, respectively, and L represents the number of image bands.

4. A rapid detection method for hyperspectral anomalous targets based on a compressed domain, as described in claim 1 or 2, characterized in that, In step two, principal component analysis is used to process the high-frequency and low-frequency subbands separately to screen characteristic bands. This process includes: The compressed high-frequency and low-frequency subbands are reconstructed into matrices respectively. and , No. The average value of the band is calculated as follows: ； Then, this average value is copied to the original image size, and the average matrix is ​​further expanded as follows: ； but covariance matrix The calculation is as follows: ； covariance matrix eigenvalues With feature vectors The following relationship exists: ； in This represents the transpose operation; we sort the eigenvectors in descending order of their eigenvalues, and select the b eigenvectors corresponding to the first b eigenvalues ​​to construct a new matrix. Finally, the original hyperspectral image is projected into the new feature space, as follows: ； Finally, the reduction results for the high-frequency and low-frequency subbands were obtained respectively. and ; Finally, the final output image is obtained by subtraction. .

5. A rapid detection method for hyperspectral anomalous targets based on a compressed domain, as described in claim 1 or 2, characterized in that, In step four, to quickly obtain the degree of pixel anomaly, the filtered single-band image is reconstructed according to wavelength, and the degree of anomaly for each pixel is calculated using a diagonal matrix, i.e., through the formula... The degree of abnormality is obtained for each pixel.

6. A rapid detection method for hyperspectral anomalous targets based on a compressed domain according to claim 1 or 2, characterized in that, Step five involves fusing and segmenting target and background pixels based on the degree of high and low frequency anomalies, including: The abnormality levels of high-frequency and low-frequency subbands are statistically analyzed and then weighted and fused, thereby utilizing both high-frequency and low-frequency information of the compressed image. Then, based on the degree of abnormality of each sample to be detected, the detection rate and false alarm rate are calculated under different threshold conditions, and the false alarm rate is determined according to the specified false alarm rate. The corresponding global threshold T is adaptively calculated, and the final detection result is as follows: ； in This indicates that the pixel is an anomalous pixel. This indicates that the pixel is normal.

Images