SAR (Synthetic Aperture Radar) internal wave region feature mining and detecting method based on block segmentation

Abstract

The invention provides an SAR (Synthetic Aperture Radar) internal wave region feature mining and detecting method based on block segmentation, which comprises the following steps: (1) preprocessing: filtering an SAR image; (2) block segmentation: carrying out overall segmentation on the image; (3) feature extraction: extracting statistical features and gradient directivity features for each SAR image block, and cascading the statistical features and the gradient directivity features to form feature vectors capable of describing the image blocks; (4) feature dimensionality reduction: performing dimensionality reduction on the feature vectors by using a principal component analysis method; (5) clustering: on the basis of the features of the samples, clustering the samples by using a clustering algorithm to form a target block region and a background block region; (6) region segmentation: communicating all target block regions to form a target region, and setting mask suppression on all background block regions to form a background region; and (7) internal wave edge detection and extraction: extracting the internal wave of the target area by using an edge detection operator based on CFAR (Constant False Alarm Rate). According to the method, a traditional global feature extraction mode is abandoned, and a new detection thought and direction are provided for SAR internal wave detection.

Description

Technical Field

[0001] The present invention relates to the field of ocean internal wave remote sensing observation, and in particular to a SAR internal wave regional feature mining and detection method based on block segmentation. Background Art

[0002] Ocean internal waves are a ubiquitous marine phenomenon. They are essentially fluctuations of the water body inside the ocean. They are complex in their occurrence mechanism and random in their temporal and spatial distribution. The relevant research on the observation and evolution mechanism of ocean internal waves has important practical significance in both military and civilian fields. Ocean internal waves affect the safety of marine engineering and the development of marine resources. For example, ocean internal waves pose a threat to marine engineering facilities such as offshore operating platforms, oil drilling platforms and submarine oil pipelines. Large-amplitude internal waves can cause strong convergence of seawater and sudden strong currents, posing a serious threat to these facilities. The study of ocean internal waves helps to better understand the marine sedimentation process and the geological structure of the seabed, which is crucial to the exploration and development of marine resources. Ocean internal waves have a great impact on the safe navigation of underwater submarines. Submarines of many countries have fallen due to encountering ocean internal waves. In the field of detection, ocean internal waves affect underwater sonar detection and communication systems. Related research helps to improve sonar technology and enhance submarine detection capabilities, which is an important part of the guarantee of the marine combat environment.

[0003] The observation of ocean internal waves can be divided into field observation and satellite remote sensing observation. Field observation directly observes the fluctuation field changes of seawater by placing temperature probes, which can be done by anchored temperature arrays or underway towing. Such methods have the disadvantages of high observation cost, small detection range, and great influence of water body fluctuations; satellite remote sensing observation uses remote sensing means to observe the distribution of internal waves in a large area. Since the SeaSat satellite used synthetic aperture radar (SAR) to effectively observe ocean internal waves for the first time in 1978, the industry has always recognized that satellite-borne SAR is the most effective and direct means of observing ocean internal waves.

[0004] The characteristics of ocean internal waves in SAR images can be summarized as follows:

[0005] (1) In SAR images, ocean internal waves appear as alternating bright and dark strips, corresponding to the crests and troughs of internal waves, respectively.

[0006] (2) In deep sea areas, the ridge lines of internal waves are often parallel to the depth contours of the seafloor;

[0007] (3) The propagation mode of tidal internal waves is mainly separated wave packets, and the interval between adjacent wave packets is 10-90 km;

[0008] (4) Each wave packet often contains several to dozens of solitary waves with a wavelength between 100m and 20km, and a length of 10-100km in the direction of the wave ridge;

[0009] (5) At the forefront of each wave packet is generally the internal wave with the largest amplitude, wavelength and ridge length, while the subsequent waves tend to attenuate in all aspects.

[0010] Internal wave detection based on SAR images has always been a hot research topic in the field of ocean internal waves. In limited samples, how to take advantage of data and construct more data to detect targets in images is a problem worthy of in-depth consideration. In the field of image processing, internal wave detection can be regarded as a detection problem of stripe distribution in images, and traditional image processing methods can generally be used to achieve internal wave detection. For example, the edge features of SAR images are used to identify the boundaries of internal waves. Commonly used algorithms include Canny edge detection and Sobel operator. Internal waves can also be detected by analyzing the texture features of SAR images. Commonly used methods include gray-level co-occurrence matrix (GLCM) and wavelet transform. For the detection of areas with obvious internal wave distribution, SAR images can be segmented based on intensity thresholds to identify internal wave areas. Commonly used threshold methods include Otsu method and adaptive threshold method. Traditional methods of internal wave detection generally adopt a global detection strategy. False alarms and missed alarms are prone to occur when there is strong interference in the image. It is generally suitable for detection under conditions of strong internal wave intensity and uniform sea surface background. In recent years, with the continuous development of deep learning, internal wave detection based on convolutional neural networks has also emerged. By building an internal wave sample data set and training the convolutional neural network model, detection is achieved. However, the deep learning-based detection method relies on a large number of data samples. In the absence of a large number of samples, the model is very likely to fall into local optimality and overfitting, and the detection method lacks adaptability. Summary of the invention

[0011] In view of the above problems, the present invention provides a SAR internal wave region feature mining and detection method based on block segmentation, which applies the core algorithms in feature engineering such as feature mining, feature dimension reduction, and feature clustering to establish a detection algorithm model for SAR image target feature extraction, and performs feature dimension reduction and feature selection by constructing multiple sample image blocks and multi-feature mining to achieve effective identification of internal wave regions. The technical solution adopted by the present invention is:

[0012] A SAR internal wave regional feature mining and detection method based on block segmentation includes the following steps:

[0013] (1) Preprocessing: filtering the SAR image;

[0014] (2) Block segmentation: the image is segmented as a whole according to a window of a certain size to form several SAR image blocks;

[0015] (3) Feature extraction: For each SAR image block, the GLCM gray-level co-occurrence matrix is ​​used to extract the statistical features in the image block, and the HOG directional gradient algorithm is used to extract the gradient directional features in the image block. The above two feature vectors are cascaded to form a feature vector that can describe the image block;

[0016] (4) Feature dimensionality reduction: PCA principal component analysis is used to reduce the dimensionality of feature vectors;

[0017] (5) Clustering: according to the features after dimensionality reduction, the number of cluster centers is set. Based on the features of the samples, the samples are clustered using a clustering algorithm to form target block regions and background block regions. The number of target block regions and background block regions is consistent with the number of block segmentations.

[0018] (6) Region segmentation: using the clustering results, all target blocks are connected to form a target area, and all background blocks are masked to form a background area;

[0019] (7) Internal wave edge detection and extraction: The internal waves in the target area are extracted using the CFAR-based edge detection operator to form the internal wave detection results.

[0020] The filtering processing method includes LEE filtering and mean filtering.

[0021] The block segmentation uses a non-overlapping segmentation method to process the image into blocks. Assuming the size of the image I is W×H and the block size is block×block, the calculation formula is:

[0022] B ij =I[(i-1)×block:i×block,(j-1)×blaock:j×block],

[0023] The image is divided into blocks by setting different block values. The maximum number of blocks is floor(W / block)×floor(H / block), where floor(·) means rounding down.

[0024] The GLCM grayscale co-occurrence matrix takes the original image as input, designs different grayscale co-occurrence directions θ, considers the spatial distance d when the grayscale values ​​co-occur, and sets the grayscale level L; initializes and creates an initial grayscale co-occurrence matrix P of size L×L; traverses each pixel (x, y) in the image, and determines the neighboring pixel point (x', y') of (x, y) according to the co-occurrence direction θ and the spatial distance d. Let g1 and g2 be the grayscale values ​​of the current pixel (x, y) and the neighboring pixel (x', y'), respectively. The grayscale values ​​of the two points can be represented by (g1, g2). Let (x, y) move in the entire image to obtain different (g1, g2) values ​​to form the grayscale co-occurrence matrix P. The normalized grayscale co-occurrence matrix P' is obtained by the following formula:

[0025]

[0026] Furthermore, the normalized GLCM matrix has the following statistical distribution characteristics: contrast characteristics, Dissimilar features, Homogeneity characteristics, The second-order moment energy characteristic, Energy characteristics, Correlation characteristics, Where μ is the mean of the gray level co-occurrence matrix, and σ is the variance of the gray level co-occurrence matrix.

[0027] The HOG directional gradient algorithm for image extraction includes the following steps:

[0028] (1) Image window grayscale and gamma correction: The HOG operator extracts features from the grayscale image and converts the RGB color image into a grayscale image for processing; gamma correction is introduced to reduce the impact of local shadows and lighting changes in the image and suppress noise interference;

[0029] (2) Gradient calculation: divide the detection window into different cells and calculate the horizontal and vertical gradients of each pixel in the cell;

[0030] (3) Divide the cell histogram, divide the gradient direction in the cell [0°~180°] into different intervals at a certain interval angle in the gradient direction, examine the gradient direction and gradient amplitude of each position in the block, and superimpose the gradient amplitude value in the gradient direction interval according to the gradient direction falling in different intervals;

[0031] (4) Feature vector cascading: multiple cells are grouped into a block, and the feature descriptors of all cells in a block are cascaded to obtain the HOG feature descriptor of the block; the HOG feature descriptors of all blocks in the image detection window are cascaded to obtain the HOG feature descriptor of the image, thereby forming a feature vector that describes the image detection window.

[0032] Furthermore, the gamma correction is Among them I in is the original input image, I out is the corrected image, γ is the correction coefficient, which is [0.5, 1.5].

[0033] Furthermore, the horizontal gradient and vertical gradient are G x (x,y)=H(x+1,y)-H(x-1,y),G y (x,y)=H(x,y+1)-H(x,y-1), in the formula, G x (x,y) represents the horizontal gradient, G y (x,y) represents the vertical gradient, H(x+1,y), H(x-1,y), H(x,y+1), and H(x,y-1) represent the grayscale values ​​of the corresponding points in the image at the pixel position; the global gradient amplitude G and gradient direction are calculated from the horizontal gradient and the vertical gradient The value range of α is [0°~180°].

[0034] Furthermore, the interval angle is 20°.

[0035] The PCA principal component analysis method dimensionality reduction process is: suppose that the n-dimensional features composed of m samples constitute the feature matrix X m×n =[x1,x2,...,x m ], where each x i is an n-dimensional column eigenvector for the feature matrix X m×n The following processing is performed: (1) Decentralization, (2) Calculate the covariance matrix, (3) Perform eigenvalue decomposition on the covariance matrix to obtain a new eigenvalue matrix. The eigenvalues ​​are arranged in columns from large to small, and the first k columns are taken to form a new matrix P. n×k , P n×k Equivalent to a new coordinate system, P n×k Each column in is a coordinate axis; (4) Project the original data onto the new P n×k In the coordinate system, we get the data after dimensionality reduction.

[0036] The clustering algorithms include fuzzy C-means clustering and K-means clustering.

[0037] Furthermore, the calculation steps of the K-means clustering algorithm include: (1) initialization, at the beginning of the algorithm, randomly selecting k data points as initial cluster centers; (2) data sample allocation, for each data point, using the Euclidean distance metric to calculate its distance to each cluster center, and assigning it to the cluster center closest to it. The Euclidean distance calculation formula is: In the formula, x is the sample data point, c i is the i-th cluster center, d is the dimension of the data, x j and c ij are x and c respectively i The value in the jth dimension; (3) Update the cluster center. The cluster center is updated after each calculation. The updated cluster center is the mean of the data points in the cluster, that is, In the formula, S i is the set of data points of the ith cluster, |S i | is the number of data points in the set; (4) Iterate and update until termination, repeating the allocation and update steps until the termination condition is met.

[0038] Furthermore, the termination conditions include: ① the cluster center no longer changes significantly, and the distance between the new cluster center and the old cluster center is less than a preset threshold; ② the clustering algorithm reaches the maximum number of iterations.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention abandons the traditional global feature extraction method, and adopts the block segmentation method to segment the SAR image into multiple equal-sized blocks; adopts the gray-level co-occurrence matrix and the directional gradient histogram algorithm to extract the features in the blocks, and cascades them to form a feature vector; adopts the principal component analysis method to reduce the dimension of the features in the blocks, applies the K-menas clustering algorithm to adaptively cluster and merge the blocks to form an internal wave area and a background area, and uses the edge detection algorithm based on constant false alarm to extract the internal wave edge features, providing a new detection idea and direction for SAR internal wave detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments or the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0041] Figure 1 It is a flow chart of SAR internal wave regional feature mining and detection based on block segmentation;

[0042] Figure 2 is a satellite SAR observation image selected in an embodiment of the present invention;

[0043] Figure 3 It is the result obtained by internal wave detection and background area mask processing in one embodiment of the present invention;

[0044] Figure 4 This is a sample processing result of retaining cluster categories 1 and 2 according to an embodiment of the present invention;

[0045] Figure 5 This is the result obtained by performing edge detection using the CFAR detection algorithm according to an embodiment of the present invention. DETAILED DESCRIPTION

[0046] In the following description, a large number of specific details are provided to provide a more thorough understanding of the present invention. However, it is apparent to those skilled in the art that the present invention can be implemented without one or more of these details. In other examples, in order to avoid confusion with the present invention, some technical features well known in the art are not described.

[0047] In order to fully understand the present invention, detailed steps and detailed structures will be proposed in the following description to illustrate the technical solution of the present invention. The preferred embodiments of the present invention are described in detail below, but in addition to these detailed descriptions, the present invention may also have other implementations.

[0048] like Figure 1 As shown in FIG, a SAR internal wave regional feature mining and detection method based on block segmentation includes the following steps: (1) preprocessing, performing LEE filtering, mean filtering and other preprocessing on the SAR image to suppress the influence of coherent speckle noise; (2) block segmentation, segmenting the image as a whole according to a window of a certain size to form a number of SAR image blocks; (3) feature extraction, for each image block, using the GLCM gray-level co-occurrence matrix to extract the statistical features in the image block, using the HOG directional gradient algorithm to extract the robust gradient directional features in the image block, and concatenating the above two feature vectors to form a feature vector that can describe the image block; (4) feature dimensionality reduction, using the PCA main feature extraction algorithm to extract the robust gradient directional features in the image block. The component analysis method is used to reduce the dimension of the lengthy feature vector and extract the valuable main features as the object of cluster analysis; (5) Clustering: According to the features after dimensionality reduction, the number of cluster centers is set. Based on the characteristics of the samples, the K-means algorithm is used to cluster the samples to form target block areas and background block areas. The number of target block areas and background block areas is consistent with the number of block segmentations; (6) Region segmentation: Using the clustering results, all target block areas are connected to form a target area, and all background block areas are set with masks to suppress them to form a background area; (7) Internal wave edge detection and extraction: Using the CFAR-based edge detection operator, the internal waves of the target area are extracted to form the final internal wave detection result.

[0049] The experimental data are selected from the SAR images obtained by the Yunhai 3A satellite. Since its launch in November 2022, the satellite has carried out routine ocean observations. The ocean SAR observation images obtained on a certain day in 2023 are selected, such as Figure 2 As shown, the image contains internal wave areas.

[0050] 1 Block segmentation of image

[0051] The image block segmentation algorithm is a sample-based image processing strategy that suppresses global interference by segmenting the image into several image blocks to achieve the purpose of region detection. The image is segmented into blocks using a non-overlapping segmentation method. Assume that the size of image I is W×H and the block size is block×block. The calculation formula is shown in formula (1):

[0052] B ij =I[(i-1)×block:i×block,(j-1)×blaock:j×block] (1)

[0053] The image can be divided into blocks by setting different block values. The maximum number of blocks is floor(W / block)×floor(H / block), where floor(·) means rounding down.

[0054] 2 Gray-level co-occurrence matrix feature mining

[0055] Gray-level Co-occurrence Matrix (GLCM) is a feature extraction algorithm based on statistics, which is used to quantify the spatial distribution relationship of image gray values. It takes the original image as input, designs the co-occurrence direction θ of different gray values, considers the spatial distance d when gray values ​​co-occur, and sets the gray level L.

[0056] Initialize and create an initial gray-level co-occurrence matrix P of size L×L.

[0057] Traverse each pixel (x, y) in the image, determine the neighboring pixel point (x', y') of (x, y) according to the co-occurrence direction θ and the spatial distance d, let g1 and g2 be the grayscale values ​​of the current pixel (x, y) and the neighboring pixel (x', y'), respectively, then the grayscale values ​​of the two points can be represented by (g1, g2), let (x, y) move in the entire image, get different (g1, g2) values, form the grayscale co-occurrence matrix P, and get the matrix P' represented by the grayscale co-occurrence matrix:

[0058]

[0059] The normalized GLCM matrix is ​​obtained, and the contrast, dissimilarity, homogeneity, second-order moment energy, energy, correlation and other features of the normalized GLCM are extracted through equations (3) to (8) as statistical distribution features in the image.

[0060] Contrast characteristics:

[0061]

[0062] Dissimilar features:

[0063]

[0064] Homogeneity characteristics:

[0065]

[0066] Second-order moment energy characteristics:

[0067]

[0068] Energy characteristics:

[0069]

[0070] Correlation characteristics:

[0071]

[0072] Where μ is the mean of the gray level co-occurrence matrix, and σ is the variance of the gray level co-occurrence matrix.

[0073] The gray-level co-occurrence matrix is ​​used to extract the above six-dimensional features from each block. The first-order moment or high-order moment of the above features can also be extracted as needed to form features of different dimensions.

[0074] 3-directional gradient algorithm feature mining

[0075] The Histogram of Oriented Gradients (HOG) is a feature descriptor. Its basic working principle is to create a histogram of the distribution of gradient directions in an image and normalize it in a special way. This special normalization allows HOG to effectively detect target edge features, and it can also achieve efficient edge detection even in low contrast conditions. The standardized histogram is composed of the HOG feature vector (called HOG descriptor), which converts the target information in the image into a describable feature vector, which can be used to train machine learning algorithms for image recognition and object detection tasks in computer vision. The steps of image feature extraction of the HOG directional gradient algorithm are as follows:

[0076] 3.1 Image Window Grayscale and Gamma Correction

[0077] The HOG operator extracts features from grayscale images, so weak RGB color images need to be converted into grayscale images for processing. At the same time, in order to reduce the impact of local shadows and lighting changes in the image and suppress noise interference, the HOG operator introduces gamma correction, which is calculated as follows:

[0078]

[0079] Among them I in is the original input image, I out is the corrected image, γ is the correction coefficient, usually [0.5, 1.5].

[0080] 3.2 Gradient Calculation

[0081] In order to capture contour information and further weaken the interference of illumination, the detection window is divided into different cells, and the horizontal gradient and vertical gradient of each pixel in the cell are calculated using equations (10) and (11):

[0082] G x (x,y)=H(x+1,y)-H(x-1,y) (10)

[0083] G y (x,y)=H(x,y+1)-H(x,y-1) (11)

[0084] In the formula, G x (x,y) represents the horizontal gradient, G y (x,y) represents the vertical gradient, H(x+1,y), H(x-1,y), H(x,y+1), and H(x,y-1) represent the grayscale values ​​of the corresponding points in the image at the pixel position.

[0085] The global gradient magnitude G and gradient direction α are calculated from the horizontal gradient and the vertical gradient, as shown in equations (12) and (13):

[0086]

[0087]

[0088] According to the principle of symmetry, the value range of α is [0°~180°].

[0089] 3.3 Dividing Cell Histogram

[0090] The gradient direction of the cell is divided into different intervals within the range of [0°~180°] at a certain interval angle in the gradient direction. For example, 20° can be taken as the interval. The gradient direction and gradient amplitude of each position in the block are examined. The gradient direction falls into different intervals, and the gradient amplitude value is superimposed in the interval of the gradient direction.

[0091] 3.4 Feature Vector Concatenation

[0092] Multiple cells are grouped into a block, and the feature descriptors of all cells in a block are cascaded to obtain the HOG feature descriptor of the block. The HOG feature descriptors of all blocks in the image detection window are cascaded to obtain the HOG feature descriptor of the image, thereby forming a feature vector that can describe the image detection window, which can be used for subsequent classification or clustering tasks.

[0093] 4 Feature Dimensionality Reduction Algorithm

[0094] High-dimensional feature vectors will lead to more redundant information in the sample, which will have a negative impact on data classification. Therefore, dimensionality reduction can be used to reduce the dimensionality of high-dimensional feature vectors. In this paper, the principal component analysis (PCA) dimensionality reduction method is selected. Through PCA, a set of orthogonal basis vectors are found. These vectors can represent the main changes in the data set, remove redundant features, and find the best descriptive feature vector. PCA can also be applied to denoising, feature extraction, etc. The most essential function of PCA is to extract the main information of the data. High-dimensional data has dimensionality disasters, and there may be correlations and noise between many variables. Extracting the main information to achieve dimensionality reduction is an effective solution.

[0095] Assume that the n-dimensional features composed of m samples constitute the feature matrix X m×n =[x1,x2,...,x m ], where each x i is an n-dimensional column eigenvector. m×n Perform the following processing:

[0096] (1) Decentralization,

[0097] (2) Calculate the covariance matrix,

[0098] (3) Perform eigenvalue decomposition on the covariance matrix to obtain a new eigenvalue matrix. The eigenvalues ​​are arranged in columns from large to small, and the first k columns are taken to form a new matrix P. n×k .P n×k Equivalent to a new coordinate system, Pn×k Each column in is an axis.

[0099] (4) Project the original data to the new P n×k In the coordinate system, the dimensionality-reduced data can be obtained.

[0100] Depending on the value of k, feature matrices with different dimensions can be obtained, but overall it is possible to reduce the feature dimension without changing the overall feature distribution.

[0101] right Figure 2 Select block size 64 and cluster number 3 as the main parameters for internal wave detection, and use the background area as a mask to obtain the detection results as follows Figure 3 shown.

[0102] 5-Sample Adaptive Clustering

[0103] In actual SAR image detection tasks, it is usually difficult to obtain enough prior information to distinguish the target and the background in advance. Therefore, on the basis of establishing image block features, unsupervised clustering can be used to achieve feature-based clustering to achieve the purpose of distinguishing possible target areas and background areas. Fuzzy C-means clustering, K-means clustering, etc. can be used in SAR image processing. In the present invention, the K-means clustering algorithm is used to achieve the segmentation of the background area and the internal wave area.

[0104] K-means algorithm is an iterative cluster analysis algorithm, and its core idea is to divide the n objects in the data set into K clusters so that the sum of the distances from each object to the center of the cluster to which it belongs (or called the mean point, centroid) is minimized. The distance mentioned here usually refers to the Euclidean distance, but it can also be other types of distance metrics. K-means algorithm continuously optimizes the clustering results in an iterative manner so that the objects in each cluster are as close as possible, while the objects between different clusters are as separated as possible. This optimization process is usually based on some objective function, such as the sum of squared errors, mean square error, root mean square error, etc. In the present invention, the sum of squared errors is used as the objective function to measure the sum of the distances of all objects to the centers of the clusters to which they belong. The calculation steps of the K-means algorithm can be summarized as follows:

[0105] (1) Initialization

[0106] At the beginning of the algorithm, k data points are randomly selected as initial cluster centers. The selection of initial cluster centers has a certain impact on the clustering results. In practical applications, heuristic methods are usually used to select better initial cluster centers.

[0107] (2) Data Sample Allocation

[0108] For each data point, the distance between it and each cluster center is calculated, and it is assigned to the cluster center closest to it. The distance measurement method can usually be used to measure the sample distance. In the present invention, the Euclidean distance measurement method is used, and it is calculated by the following formula:

[0109]

[0110] Where x is the sample data point, c i is the i-th cluster center, d is the dimension of the data, x j and c ij are x and c respectively i The value in the jth dimension.

[0111] (3) Update cluster centers

[0112] The cluster center is updated after each calculation. The updated cluster center is the mean of the data points in the cluster, which is calculated by the following formula:

[0113]

[0114] Among them, S i is the set of data points of the ith cluster, |S i | is the number of data points in the set.

[0115] (4) Iterate and update until termination

[0116] Repeat the allocation and update steps until some termination condition is met. Common termination conditions include:

[0117] ① The cluster center no longer changes significantly, that is, the distance between the new cluster center and the old cluster center is less than a preset threshold;

[0118] ② Reaching the maximum number of iterations. In order to prevent the algorithm from falling into an infinite loop, a maximum number of iterations is usually set as the termination condition.

[0119] During the iteration process, the algorithm will continuously optimize the clustering results, making the objects within each cluster more compact and the objects between different clusters more dispersed. Finally, when the termination condition is met, the algorithm stops iterating and outputs the final clustering results.

[0120] right Figure 3 Select samples with cluster categories 1 and 2 as the retention, and the results are as follows Figure 4 shown.

[0121] right Figure 4 The CFAR detection algorithm is used to perform edge detection and obtain the final internal wave detection result, as shown in the following figure. Figure 5 shown.

[0122] The preferred embodiments of the present invention are described above. It should be understood that the present invention is not limited to the above-mentioned specific embodiments, nor is it limited to the SAR internal wave regional features. The feature extraction and recognition methods that are not described in detail should be understood to be implemented in a common manner in the art; any technician familiar with the art can make many possible changes and modifications to the technical solutions of the present invention using the above-disclosed methods and technical contents without departing from the scope of the technical solutions of the present invention, or modify them into equivalent embodiments of equivalent changes, which does not affect the essential contents of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the contents of the technical solutions of the present invention are still within the scope of protection of the technical solutions of the present invention.

Claims

1. A SAR internal wave regional feature mining and detection method based on block segmentation, characterized in that: The steps include: (1) Preprocessing: filtering the SAR image; (2) Block segmentation: the image is segmented as a whole according to a window of a certain size to form several SAR image blocks; (3) Feature extraction: For each SAR image block, the GLCM gray-level co-occurrence matrix is ​​used to extract the statistical features in the image block, and the HOG directional gradient algorithm is used to extract the gradient directional features in the image block. The above two feature vectors are cascaded to form a feature vector that can describe the image block; (4) Feature dimensionality reduction: PCA principal component analysis is used to reduce the dimensionality of feature vectors; (5) Clustering: according to the features after dimensionality reduction, the number of cluster centers is set. Based on the features of the samples, the samples are clustered using a clustering algorithm to form target block regions and background block regions. The number of target block regions and background block regions is consistent with the number of block segmentations. (6) Region segmentation: using the clustering results, all target blocks are connected to form a target area, and all background blocks are masked to form a background area; (7) Internal wave edge detection and extraction: The internal waves in the target area are extracted using the CFAR-based edge detection operator to form the internal wave detection results.

2. The SAR internal wave regional feature mining and detection method based on block segmentation as claimed in claim 1, characterized in that: The filtering processing method includes LEE filtering and mean filtering.

3. The SAR internal wave regional feature mining and detection method based on block segmentation as claimed in claim 1, characterized in that: The block segmentation uses a non-overlapping segmentation method to process the image into blocks. Assuming the size of the image I is W×H and the block size is block×block, the calculation formula is: B ij =I[(i-1)×block:i×block,(j-1)×blaock:j×block], The image is divided into blocks by setting different block values. The maximum number of blocks is floor(W / block)×floor(H / block), where floor(·) means rounding down.

4. The SAR internal wave regional feature mining and detection method based on block segmentation as claimed in claim 1, characterized in that: The GLCM grayscale co-occurrence matrix takes the original image as input, designs different grayscale co-occurrence directions θ, considers the spatial distance d when the grayscale values ​​co-occur, and sets the grayscale level L; initializes and creates an initial grayscale co-occurrence matrix P of size L×L; traverses each pixel (x, y) in the image, and determines the neighboring pixel point (x', y') of (x, y) according to the co-occurrence direction θ and the spatial distance d. Let g1 and g2 be the grayscale values ​​of the current pixel (x, y) and the neighboring pixel (x', y'), respectively. The grayscale values ​​of the two points can be represented by (g1, g2). Let (x, y) move in the entire image to obtain different (g1, g2) values ​​to form the grayscale co-occurrence matrix P. The matrix P' represented by the grayscale co-occurrence matrix is ​​obtained by the following formula:

5. The SAR internal wave regional feature mining and detection method based on block segmentation as claimed in claim 4, characterized in that: The normalized GLCM matrix has the following statistical distribution characteristics: contrast characteristics, Dissimilar features, Homogeneity characteristics, The second-order moment energy characteristic, Energy characteristics, Correlation characteristics, Where μ is the mean of the gray level co-occurrence matrix, and σ is the variance of the gray level co-occurrence matrix.

6. The SAR internal wave regional feature mining and detection method based on block segmentation as claimed in claim 1, characterized in that: The HOG directional gradient algorithm for image extraction includes the following steps: (1) Image window grayscale and gamma correction: The HOG operator extracts features from the grayscale image and converts the RGB color image into a grayscale image for processing; gamma correction is introduced to reduce the impact of local shadows and lighting changes in the image and suppress noise interference; (2) Gradient calculation: divide the detection window into different cells and calculate the horizontal and vertical gradients of each pixel in the cell; (3) Divide the cell histogram, divide the gradient direction in the cell [0°~180°] into different intervals at a certain interval angle in the gradient direction, examine the gradient direction and gradient amplitude of each position in the block, and superimpose the gradient amplitude value in the gradient direction interval according to the gradient direction falling in different intervals; (4) Feature vector cascading: multiple cells are grouped into a block, and the feature descriptors of all cells in a block are cascaded to obtain the HOG feature descriptor of the block; the HOG feature descriptors of all blocks in the image detection window are cascaded to obtain the HOG feature descriptor of the image, thereby forming a feature vector that describes the image detection window.

7. The SAR internal wave regional feature mining and detection method based on block segmentation as claimed in claim 6, characterized in that: The gamma correction is Among them I in is the original input image, I out is the corrected image, γ is the correction coefficient, which is [0.5, 1.5].

8. The SAR internal wave regional feature mining and detection method based on block segmentation as claimed in claim 6, characterized in that: The horizontal gradient and vertical gradient are G x (x,y)=H(x+1,y)-H(x-1,y),G y (x,y)=H(x,y+1)-H(x,y-1), in the formula, G x (x,y) represents the horizontal gradient, G y (x,y) represents the vertical gradient, H(x+1,y), H(x-1,y), H(x,y+1), and H(x,y-1) represent the grayscale values ​​of the image at the corresponding pixel positions respectively; the global gradient amplitude G and gradient direction α are calculated from the horizontal gradient and vertical gradient. The value range of α is [0°~180°].

9. The SAR internal wave regional feature mining and detection method based on block segmentation as claimed in claim 6, characterized in that: The interval angle is 20°.

10. The SAR internal wave regional feature mining and detection method based on block segmentation according to claim 1, characterized in that: The PCA principal component analysis method dimensionality reduction process is: suppose that the n-dimensional features composed of m samples constitute the feature matrix X m×n =[x1,x2,...,x m ], where each x i is an n-dimensional column eigenvector for the feature matrix X m×n The following processing is performed: (1) Decentralization, (2) Calculate the covariance matrix, (3) Perform eigenvalue decomposition on the covariance matrix to obtain a new eigenvalue matrix. The eigenvalues ​​are arranged in columns from large to small, and the first k columns are taken to form a new matrix P. n×k , P n×k Equivalent to a new coordinate system, P n×k Each column in is a coordinate axis; (4) Project the original data to the new P n×k In the coordinate system, we get the data after dimensionality reduction.

11. The SAR internal wave regional feature mining and detection method based on block segmentation according to claim 1, characterized in that: The clustering algorithms include fuzzy C-means clustering and K-means clustering.

12. The SAR internal wave regional feature mining and detection method based on block segmentation as claimed in claim 11, characterized in that: The calculation steps of the K-means clustering algorithm include: (1) initialization, at the beginning of the algorithm, randomly select k data points as the initial cluster centers; (2) data sample allocation, for each data point, use the Euclidean distance metric to calculate its distance to each cluster center, and assign it to the cluster center with the closest distance. The Euclidean distance calculation formula is: In the formula, x is the sample data point, c i is the i-th cluster center, d is the dimension of the data, x j and c ij are x and c respectively i The value in the jth dimension; (3) Update the cluster center. The cluster center is updated after each calculation. The updated cluster center is the mean of the data points in the cluster, that is, In the formula, S i is the set of data points of the ith cluster, |S i | is the number of data points in the set; (4) Iterate and update until termination, repeating the allocation and update steps until the termination condition is met.

13. The SAR internal wave regional feature mining and detection method based on block segmentation as claimed in claim 12, characterized in that: The termination conditions include: ① the cluster center no longer changes significantly, and the distance between the new cluster center and the old cluster center is less than a preset threshold; ② the clustering algorithm reaches the maximum number of iterations.

Images