High-precision video sar moving target detection method based on information geometry theory

Abstract

The application discloses a high-precision video SAR moving target detection method based on information geometry theory in the technical field of SAR moving target shadow detection, and comprises the following steps: shadow enhancement and background suppression on a video SAR image sequence; ROI region extraction based on the information geometry theory; and false ROI region removal through false alarm suppression. The high-precision video SAR moving target detection method based on the information geometry theory improves the detection probability of the video SAR moving target shadow under the premise of ensuring the detection precision.

Description

Technical Field

[0001] This invention relates to the field of SAR moving target shadow detection technology, and more specifically, to a high-precision video SAR moving target detection method based on information geometry theory. Background Technology

[0002] Video SAR, due to its ability to continuously monitor changes in target areas, has wide applications in military decision-making, smart cities, and traffic management. In video SAR systems, because the radar carrier frequency is high, the Doppler modulation of moving target echoes is highly sensitive to target movement. Even small movements can cause significant shifts and defocusing in SAR images of moving targets. However, the shadows left by moving targets can reflect their actual position and state information within the SAR image sequence. Therefore, compared to traditional SAR systems, the shadows of moving targets are easier to observe in video SAR image sequences, making target shadow detection a more direct and effective method.

[0003] Currently, there are three main types of methods in traditional video SAR moving target shadow detection technology: differential detection, constant false alarm rate (CFAR) detector, and threshold segmentation-based moving target detection methods.

[0004] Differential-based detection methods first estimate the background image using multiple consecutive frames of video SAR images. Then, they perform a difference operation between the background image and the current image to be detected to obtain a binary image of the foreground. Next, an inter-frame difference algorithm is used to extract motion information from previous and subsequent frames, and this information is then ANDed with the binary image obtained from the background difference. Finally, morphological processing is performed on the results from the previous step to extract the shadow region of the moving target. Although differential-based detection methods are effective, they have some limitations: First, they place high demands on the background modeling algorithm, and the detection bounding boxes may lag due to the use of estimations from previous frames. Second, because the inter-frame difference algorithm only compares the differences between two frames and ignores information about overlapping areas, it is easy to see holes and double-sided phenomena in the shadow region of the moving target, leading to missed detections or false alarms.

[0005] CFAR-based detection methods can effectively circumvent the aforementioned limitations. CFAR can select a suitable region within the background noise of a video SAR image and determine an appropriate detection threshold based on the noise statistical characteristics within that region. When the target region exceeds this threshold, the target is considered detected. However, in multi-target environments, the reference region may contain multiple target signals, which can reduce the detection performance of the method; simultaneously, CFAR performance is also affected in non-uniform noise environments.

[0006] Threshold-based detection methods utilize image thresholding techniques (such as the OTSU method and the minimum error method) to binarize video SAR images and extract the shadow regions of moving targets. However, thresholding methods segment based on the image's histogram information, ignoring the spatial and edge information of the video SAR image, thus easily leading to missed detections and excessively high false alarm rates. Summary of the Invention

[0007] To address the aforementioned problems, the present invention aims to provide a high-precision video SAR moving target detection method based on information geometry theory, thereby solving the technical problems of shortcomings in traditional video SAR moving target shadow detection methods and improving the detection probability of video SAR moving target shadows while ensuring detection accuracy.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] This invention provides a high-precision video SAR moving target detection method based on information geometry theory, comprising the following steps:

[0010] S1. Perform shadow enhancement and background suppression on video SAR image sequences;

[0011] S2. ROI region extraction based on information geometry theory;

[0012] S3. By suppressing false detections, false ROI regions are effectively removed.

[0013] Furthermore, S1 includes the following steps:

[0014] S11. Calculate the initial subspace {U,V} of the video SAR image sequence using the principal component analysis algorithm;

[0015] S12. Set the number of Gaussian mixtures and model the data matrix X composed of video SAR image sequences using a low-rank representation model based on Gaussian mixtures, i.e.:

[0016]

[0017] in, This represents the set of weights for all Gaussian components. This represents the set of variances of all Gaussian components;

[0018] S13. Update the low-rank matrix U*V using online subspace learning techniques. T Subtracting the original video SAR image from the foreground image yields a foreground image containing the shadows of moving targets in the video SAR;

[0019] S14. The alternating direction multiplier algorithm is used to remove strong scattered objects in the foreground image to obtain a video SAR image sequence that has undergone shadow enhancement and background suppression, i.e., enhanced image.

[0020] Furthermore, S2 includes the following steps:

[0021] S21. Construct a neighborhood system by taking all pixels in the neighborhood of the pixel at coordinate point in the enhanced image, and build a statistical distribution describing that point;

[0022] S22. Calculate the JS divergence to describe the geometric distance in the Gaussian statistical manifold, perform pattern analysis, and capture local information features in the image;

[0023] S23. Traverse each pixel of the image to obtain the corresponding JS divergence matrix of the enhanced image, and calculate the joint probability distribution JPDF;

[0024] S24. Use the minimum cross-entropy algorithm to select a threshold and extract the ROI region of the enhanced image.

[0025] Furthermore, S21 includes the following steps:

[0026] S211. Construct a neighborhood system of all pixels within an N*N neighborhood of the (x,y) pixel in the enhanced image. Calculate the mean μ and variance σ of all pixels within this neighborhood system, and describe the statistical distribution of this neighborhood system using a Gaussian distribution:

[0027]

[0028] S212. Given the mean μ′ of the gray intensity at the current (x,y) pixel and a very small variance σ′, construct a statistical distribution describing this point:

[0029]

[0030] Furthermore, S22 includes the following steps:

[0031] S221. From the parameters θ=(μ,σ) T The family of probability distributions formed together constitutes a statistical model S = {P(x|θ)|θ∈Θ}, and S forms a differentiable Gaussian statistical manifold under a certain topological structure.

[0032] S222. Calculate the JS divergence, and use the JS divergence to describe the geometric distance between different probability distributions in the Gaussian statistical manifold to perform pattern analysis;

[0033] S223. Normalize the numerical value of JS divergence. On a Gaussian statistical manifold, JS divergence explains the geometric relationship between the pixel gray value at the current location and the distribution of its neighborhood system, helping to capture local information features in the image.

[0034] Furthermore, the formula for calculating the JS divergence in S222 is as follows:

[0035]

[0036] in, It is the average distribution of two distributions, and KL(pq) is defined as:

[0037]

[0038] Furthermore, S23 includes the following steps:

[0039] S231. Traverse each pixel of the image to obtain the corresponding JS divergence matrix of the enhanced image;

[0040] S232. Let n ij Let n be the number of times the grayscale value and JS divergence occur together. ij Dividing each element in the algorithm by the total number of pixels in the enhanced image yields the normalized joint probability distribution JPDF.

[0041] Furthermore, S3 includes the following steps:

[0042] S31. Calculate the gradient image of the current frame, and use the gradient image to remove erroneous ROI regions;

[0043] S32. Use data association algorithms to suppress false alarms across multiple frames.

[0044] Furthermore, S31 includes the following steps:

[0045] S311. The Sobel operator is used to detect the edges of the image, and the relative change of gray levels of the original video SAR image is calculated to obtain the gradient image of the current original SAR image.

[0046] S312. Introduce the gradient image, calculate the overlap area between the gradient image and the ROI region, and remove ROI regions with an overlap area greater than the threshold.

[0047] Furthermore, S32 includes the following steps:

[0048] S321. Extract the centroid position and corresponding time frame number of each ROI region as a feature vector;

[0049] S322. Set a spatial distance threshold T Iand time threshold T T By associating the detection points within the time threshold frame using these two thresholds, possible real targets can be identified.

[0050] S323. Detection points with strong consistency across multiple frames are considered valid targets, and only detection points that meet certain correlation counts and spatial changes in centroid are retained.

[0051] By adopting the above technical solution, the present invention has the following advantages:

[0052] 1. This invention provides a high-precision video SAR moving target detection method based on information geometry theory. By performing shadow enhancement and background suppression operations on the video SAR image sequence, the representation of moving target shadows is enhanced, while effectively suppressing strong scattered objects and low scattering areas in the background.

[0053] 2. This invention provides a high-precision video SAR moving target detection method based on information geometry theory. Based on information geometry theory, the geometric distance between the current pixel position and its neighborhood system distribution is calculated, revealing the changes in local structure. It also forms a joint probability distribution with the gray value distribution. The minimum cross-entropy is used to complete the extraction of the ROI region, which greatly improves the detection performance of moving target shadows.

[0054] 3. This invention provides a high-precision video SAR moving target detection method based on information geometry theory. By utilizing edge information and dynamic information between multiple frames, it effectively removes false ROI regions and improves the detection accuracy of moving target shadows. Attached Figure Description

[0055] Figure 1 This is a flowchart of the high-precision video SAR moving target detection method based on information geometry theory of the present invention;

[0056] Figure 2 This is a flowchart of the ROI region extraction process of the present invention;

[0057] Figure 3(a) shows the original video SAR image of the present invention;

[0058] Figure 3(b) is an enhanced image of the present invention;

[0059] Figure 4 This refers to the ROI region extracted in this invention;

[0060] Figure 5 This refers to the detection results of moving targets in each frame of this invention. Detailed Implementation

[0061] The technical solution of the present invention will be specifically described below with reference to the accompanying drawings. It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0062] Figure 1 The flowchart illustrates a high-precision video SAR moving target detection method based on information geometry theory. This method can be divided into three parts, including the following steps: Figure 1 As shown:

[0063] S1. Perform shadow enhancement and background suppression on video SAR image sequences;

[0064] S1 includes the following specific steps:

[0065] S11. Calculate the initial subspace {U,V} of the video SAR image sequence using the principal component analysis algorithm;

[0066] S12. Set the number of Gaussian mixtures and model the data matrix X composed of video SAR image sequences using a low-rank representation model based on Gaussian mixtures, i.e.:

[0067]

[0068] in, This represents the set of weights for all Gaussian components. This represents the set of variances of all Gaussian components;

[0069] S13. Update the low-rank matrix U*V using online subspace learning techniques. T Subtracting the original video SAR image from the foreground image yields a foreground image containing the shadows of moving targets in the video SAR;

[0070] S14. The alternating direction multiplier algorithm is used to remove strong scattered objects in the foreground image to obtain a video SAR image sequence that has been enhanced by shadow enhancement and background suppression, which is simply referred to as the enhanced image.

[0071] S2. ROI region extraction based on information geometry theory;

[0072] S2 includes the following steps, specifically as follows: Figure 2As shown:

[0073] S21. Construct a neighborhood system by taking all pixels in the neighborhood of the pixel at coordinate point in the enhanced image, and build a statistical distribution describing that point;

[0074] S21 includes the following specific steps:

[0075] S211. Construct a neighborhood system of all pixels within an N*N neighborhood of the (x,y) pixel in the enhanced image. Calculate the mean μ and variance σ of all pixels within this neighborhood system, and describe the statistical distribution of this neighborhood system using a Gaussian distribution:

[0076]

[0077] S212. Given the mean μ′ of the gray intensity at the current (x,y) pixel and a very small variance σ′, construct a statistical distribution describing this point:

[0078]

[0079] S22. Calculate the JS divergence to describe the geometric distance in the Gaussian statistical manifold, perform pattern analysis, and capture local information features in the image;

[0080] S22 includes the following specific steps:

[0081] S221. From the parameters θ=(μ,σ) T The family of probability distributions formed can form a statistical model S = {P(x|θ)|θ∈Θ}. S forms a differentiable Gaussian statistical manifold under a certain topological structure. Compared with Euclidean space and Euclidean distance metric, statistical manifold and geodesic distance-based metric on it can better reflect the geometric position and similarity of each point.

[0082] S222. The Jensen–Shannon divergence (JS divergence) is used to calculate the geometric distance between different probability distributions in the Gaussian statistical manifold for pattern analysis.

[0083]

[0084] in, The average distributions of the two distributions are given. KL(p||q) is the Kullback-Leibler divergence, defined as:

[0085]

[0086] S223. Normalize the JS divergence value to 0-255. On a Gaussian statistical manifold, the JS divergence provides a non-Euclidean "distance" measure. Its performance on the manifold can explain the geometric relationship between the pixel gray value at the current location and the distribution of its neighborhood system. It can reveal changes in local structure and help capture local information features in the image, which are often "target regions" that are different from the background.

[0087] S23. Traverse each pixel of the image to obtain the corresponding JS divergence matrix of the enhanced image, and calculate the joint probability distribution JPDF;

[0088] S23 includes the following specific steps:

[0089] S231. Traverse each pixel of the image to obtain the corresponding JS divergence matrix of the enhanced image;

[0090] S232. Let n ij Let n be the number of times the grayscale value and JS divergence occur together. ij Dividing each element in the algorithm by the total number of pixels in the enhanced image yields the normalized joint probability distribution JPDF.

[0091] S24. Based on the obtained joint probability distribution JPDF, calculate JPDF = {jpdf} using the minimum cross-entropy algorithm. ij The optimal grayscale threshold T for {i,j = 0, 1, ..., L-1} G With JS divergence threshold T JS Extract the ROI region of the enhanced image.

[0092] S3. By suppressing false detections (assisted by road edge information and inter-frame correlation), false ROI regions are effectively removed.

[0093] S3 includes the following specific steps:

[0094] S31. Calculate the gradient image of the current frame, and use the gradient image to remove erroneous ROI regions;

[0095] S31 includes the following specific steps:

[0096] S311. The ROI regions extracted from the ROI regions do not contain the shadows of moving targets. Most of these erroneous ROI regions are distributed in the edge regions of the image. The Sobel operator is used to detect the image edges and calculate the relative change of gray levels in the original video SAR image to obtain the gradient image of the current original SAR image.

[0097] S312. Introduce the gradient image, calculate the overlap area between the gradient image and the ROI region, and for overlap areas greater than a threshold T...A The ROI region is removed.

[0098] S32. Most of the erroneous ROI regions existing at the edges have been removed, but a small number still exist. Therefore, a data association algorithm is used to suppress false alarms across multiple frames.

[0099] Specifically, S32 includes the following steps:

[0100] S321. Extract the centroid position and corresponding time frame number of each ROI region as a feature vector;

[0101] S322. Set a spatial distance threshold T I and time threshold T T By associating the detection points within the time threshold frame using these two thresholds, possible real targets can be identified.

[0102] S323. Detection points with strong consistency across multiple frames are considered valid targets, and only detection points that meet certain correlation counts and spatial changes in centroid are retained.

[0103] Ultimately, by going through the above three parts, high-precision moving target detection of video SAR can be achieved.

[0104] In a specific embodiment, the implementation steps are as follows:

[0105] S1. A low-rank representation model based on Gaussian mixture is used to perform shadow enhancement and background suppression on the original video SAR image to obtain an enhanced image sequence. In this example, the 56th frame video SAR image and the enhanced image are selected as examples, as shown in Figure 3(a) and Figure 3(b). From the comparison of Figure 3(a) and Figure 3(b), it can be seen that the shadow of the moving target in the enhanced image is more obvious and the contrast with its surroundings is higher. At the same time, the background of the original video SAR image is suppressed so that the subsequent ROI region extraction operation will not be interfered with by its background.

[0106] S2. Set a neighborhood size of 20*20, extract the JS divergence feature matrix of the image using information geometry theory, combine it with the image grayscale values, calculate the joint probability distribution JPDF of the image, and use the minimum cross-entropy algorithm to calculate the optimal grayscale threshold T. G With JS divergence threshold T JS Pixels with positions smaller than both threshold values ​​are considered Regions of Interest (ROIs). Connectivity analysis is performed on the ROIs, and connected components with more than 500 pixels or less than 100 pixels are deleted. The centroid positions and frame numbers of the connected components are recorded before proceeding to the next step. The overall process for this step is as follows: Figure 2 As shown in the example image, the extracted ROI region is shown below. Figure 4 As shown, from Figure 4 As can be seen, almost all shadow areas of moving targets were detected.

[0107] S3. Calculate the gradient magnitude of the current original video SAR image using the Sobel operator, generating gradient images with a gradient magnitude greater than 50. Calculate the overlap area between the gradient image and the ROI region. Regions with an overlap area greater than 50% of their own ROI area are considered erroneous ROIs and are removed. Then, extract the centroid position and time frame number of the ROI region for each frame, and calculate the Euclidean distance between the detection results of the current frame and the detection results of the next frame frame by frame. If the distance between a target and the target in the next frame is less than the spatial distance threshold (T...),... I If the time threshold (T) is 30, then the two targets are considered to belong to the same trajectory and are marked as the same group. If not, the search continues in subsequent frames until the time threshold (T) is exceeded. T =5). Finally, the association count for each target is calculated, and the change between the first and last centroid positions is detected. If the motion distance is large and the association count is greater than or equal to 8, the target is considered a valid target. The final detection results for moving targets in each frame are as follows: Figure 5 As shown, the red box is the detection box of the present invention, and the green box is the actual label box.

[0108] Finally, it should be noted that although the present invention has been described with reference to specific embodiments, those skilled in the art should recognize that the above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Various equivalent changes or substitutions can be made without departing from the concept of the present invention. Therefore, any changes or modifications to the above embodiments within the essential spirit of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A high-precision video SAR moving target detection method based on information geometry theory, characterized in that, Includes the following steps: S1. Perform shadow enhancement and background suppression on video SAR image sequences; S2. ROI region extraction based on information geometry theory; S2 includes the following steps: S21. Construct a neighborhood system by combining all pixels in the neighborhood of the enhanced image at the coordinate pixel, and build a statistical distribution describing the pixel at that coordinate. S21 includes the following steps: S211. Construct a neighborhood system of all pixels within an N*N neighborhood of the (x,y) pixel in the enhanced image. Calculate the mean μ and variance σ of all pixels within this neighborhood system, and describe the statistical distribution of this neighborhood system using a Gaussian distribution: S212. Given a mean μ′ of the gray intensity value at the current (x,y) pixel and a very small variance σ′, construct a statistical distribution describing the current pixel: S22. Calculate the JS divergence to describe the geometric distance in the Gaussian statistical manifold, perform pattern analysis, and capture local information features in the image; S22 includes the following steps: S221. From the parameters θ=(μ,σ) T The family of probability distributions formed together constitutes a statistical model S = {P(x|θ)|θ∈Θ}, and S forms a differentiable Gaussian statistical manifold under a certain topological structure. S222. Calculate the JS divergence, and use the JS divergence to describe the geometric distance between different probability distributions in the Gaussian statistical manifold to perform pattern analysis; S223. Normalize the numerical value of JS divergence. On a Gaussian statistical manifold, JS divergence explains the geometric relationship between the pixel gray value at the current position and the distribution of its neighborhood system, helping to capture local information features in the image. S23. Traverse each pixel of the image to obtain the corresponding JS divergence matrix of the enhanced image, and calculate the joint probability distribution JPDF; S23 includes the following steps: S231. Traverse each pixel of the image to obtain the corresponding JS divergence matrix of the enhanced image; S232. Let n ij Let n be the number of times the grayscale value and JS divergence occur together. ij Dividing each element in the matrix by the total number of pixels in the enhanced image yields the normalized joint probability distribution JPDF; S24. Use the minimum cross-entropy algorithm to select a threshold and extract the ROI region of the enhanced image; S3. By suppressing false detections, false ROI regions are effectively removed.

2. The high-precision video SAR moving target detection method based on information geometry theory according to claim 1, characterized in that, S1 includes the following steps: S11. Calculate the initial subspace {U,V} of the video SAR image sequence using the principal component analysis algorithm; S12. Set the number of Gaussian mixtures and model the data matrix X composed of video SAR image sequences using a low-rank representation model based on Gaussian mixtures, i.e.: in, This represents the set of weights for all Gaussian components. This represents the set of variances of all Gaussian components; S13. Update the low-rank matrix U*V using online subspace learning techniques. T Subtracting the original video SAR image from the foreground image yields a foreground image containing the shadows of moving targets in the video SAR; S14. The alternating direction multiplier algorithm is used to remove strong scattered objects in the foreground image to obtain a video SAR image sequence that has undergone shadow enhancement and background suppression, i.e., enhanced image.

3. The high-precision video SAR moving target detection method based on information geometry theory according to claim 1, characterized in that, The formula for calculating the JS divergence in S222 is as follows: in, The average distribution of two distributions is given by KL(p||q), which is defined as:

4. The high-precision video SAR moving target detection method based on information geometry theory according to claim 1, characterized in that, S3 includes the following steps: S31. Calculate the gradient image of the current frame, and use the gradient image to remove erroneous ROI regions; S32. Use data association algorithms to suppress false alarms across multiple frames.

5. The high-precision video SAR moving target detection method based on information geometry theory according to claim 4, characterized in that, S31 includes the following steps: S311. The Sobel operator is used to detect the edges of the image, and the relative change of gray levels of the original video SAR image is calculated to obtain the gradient image of the current original SAR image. S312. Introduce the gradient image, calculate the overlap area between the gradient image and the ROI region, and remove ROI regions with an overlap area greater than the threshold.

6. The high-precision video SAR moving target detection method based on information geometry theory according to claim 5, characterized in that, S32 includes the following steps: S321. Extract the centroid position and corresponding time frame number of each ROI region as a feature vector; S322. Set a spatial distance threshold T I and time threshold T T By associating the detection points within the time threshold frame using these two thresholds, possible real targets can be identified. S323. Detection points with strong consistency across multiple frames are considered valid targets, and only detection points that meet certain correlation counts and spatial changes in centroid are retained.

Images