Ground wave radar clutter classification and identification method based on single-stage convolutional neural network

Abstract

The application provides a ground wave radar clutter classification and identification method based on a single-stage convolutional neural network, obtains radar echo data for training, and constructs a training data set image; radar RD feature maps are constructed using radar echo data images, and the feature maps are input into a YOLOv3 network; a plurality of initial preset identification boxes of different sizes are set for each grid; real radar echo images are classified and marked according to the types of clutter signals, and real marking boxes of each clutter signal are obtained; the size of the real marking box is randomly selected as the initial value of the preset boundary box size; the matching degree d between the real marking box and the preset identification box is calculated; the clutter type corresponding to the real marking box is assigned to the preset identification box with the minimum error as the clutter identification type of the preset identification box; and the process is repeated until the clutter identification types of all preset identification boxes are confirmed.

Description

Technical Field

[0001] This invention belongs to the fields of radio physics and marine science and technology, and relates to a ground wave radar clutter classification and identification method based on a single-level convolutional neural network YOLOv3. Background Technology

[0002] High-frequency surface wave radar (HFSWR), also known as high-frequency ground wave over-the-horizon radar, is an important means of large-scale, continuous detection of maritime targets. It utilizes the characteristics of high-frequency electromagnetic waves (3–30 MHz) diffracting and propagating along the sea surface to achieve large-scale, continuous monitoring of moving targets such as ships and low-flying aircraft, as well as marine dynamic parameters.

[0003] With the continuous advancement of technology and the deepening of research, shore-based high-frequency ground wave radar is constantly being developed and improved, while experiments on shipborne HFSWRs, which are more flexible and mobile and have a longer detection range, are also being carried out one after another.

[0004] High-frequency ground wave radar data typically contains targets, sea clutter, ground clutter, ionospheric clutter, and radio frequency interference. Researchers need to spend a lot of time filtering to quickly obtain the interference data of interest from a large amount of ground wave radar data. Many institutions have carried out research on methods for identifying interference signals in radar signal echoes. They use time-frequency characteristics, distribution characteristics, spatiotemporal characteristics, or modeling analysis to identify interference signals from the data, and have achieved good results.

[0005] However, existing ground wave radar clutter identification methods rely on manually constructed features, and their effectiveness depends on the expressive power of the features. Furthermore, the identification process involves a significant amount of computation, which slows down the identification speed. Current methods for identifying large volumes of data involve researchers generating RD spectra from the radar data and then visually identifying the signals based on experience. This approach is extremely inefficient, has highly subjective judgment criteria, and is prone to mislabeling and omissions, making it difficult to meet engineering requirements. Summary of the Invention

[0006] The purpose of this invention is to provide a ground wave radar clutter classification and identification method with high identification efficiency and accuracy.

[0007] To achieve the above objectives, some embodiments of the present invention provide the following technical solutions:

[0008] A ground wave radar clutter classification and identification method based on a single-level convolutional neural network includes the following steps:

[0009] Acquire radar echo data for training and construct training dataset images;

[0010] Radar RD feature maps are constructed using radar echo data images, and the feature maps are input into the YOLOv3 network;

[0011] According to the clutter signal type, the real radar echo image is classified and labeled to obtain the real label box; the size of the real label box is randomly selected as the initial value of the preset bounding box size; the matching degree d between the real label box and the preset recognition box is calculated; the clutter type corresponding to the real label box is assigned to the preset recognition box with the smallest error, which is used as the clutter recognition type of the preset recognition box.

[0012] Repeat until the clutter recognition type of all preset recognition boxes is confirmed.

[0013] In some embodiments of the present invention, the following steps are further included:

[0014] The median value of the width and height of all real marker boxes corresponding to a single preset recognition box is calculated and used as the size of the preset recognition box.

[0015] In some embodiments of the present invention, the method for obtaining the actual marker box includes:

[0016] Determine the coordinate range on the actual radar echo image of each clutter signal, convert the coordinate range into the length and width of the bounding box, and determine the actual bounding box based on the length and width.

[0017] In some embodiments of the present invention, the dataset images include multiple batches of radar echo data. After constructing the training dataset images, the radar echo data images are further preprocessed, and radar RD feature maps are constructed using the preprocessed radar echo data images. The data preprocessing method includes:

[0018] Choose a color gamut length of 64;

[0019] For radar data from the same batch, the maximum and minimum values ​​of radar echo energy are taken; the data between the maximum and minimum echo energy values ​​are then divided into 64 equal parts.

[0020] Determine whether the echo energy interval between adjacent components after equal division is less than 1. If so, perform upward verification for each energy interval; otherwise, expand each energy interval upward to 1.

[0021] The color gamut length is divided into a 1:2:2:2:1 ratio, and the color is applied according to the selected color gamut length and the radar echo image.

[0022] In some embodiments of the present invention, the method for calculating the matching degree between the real bounding box and the preset recognition box includes:

[0023] IoU = Area where the projection of the true bounding box onto the RD feature map intersects with the preset recognition box / Area where the projection of the true bounding box onto the RD feature map merges with the preset recognition box;

[0024] d = 1 - IoU;

[0025] d represents the degree of matching between the actual marker box and the preset marker box.

[0026] In some embodiments of the present invention, the method for setting a preset recognition box includes:

[0027] For each grid in the feature map, set three recognition methods: large, medium, and small, and set three preset bounding boxes for each size.

[0028] In some embodiments of the present invention, the clutter signals include, but are not limited to: sea clutter, ground clutter, ionospheric clutter, and radio frequency interference clutter.

[0029] Compared with the prior art, the beneficial effects of the technical solution of the present invention are as follows:

[0030] The calibration of all preset clutter types is completed by comparing the real marker boxes with the preset marker boxes using real radar echo icons. This eliminates the need for manual feature construction and achieves accurate and efficient clutter identification. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a schematic diagram of the ground wave radar clutter classification and identification method of the present invention;

[0033] Figure 2 This is a schematic diagram of the energy distribution of radar echo data;

[0034] Figure 3 A schematic diagram illustrating the division of radar echo data energy values;

[0035] Figure 4 A schematic diagram of color gamut division;

[0036] Figure 5 This is a schematic diagram of the bounding box offset;

[0037] Figure 6 This is a diagram illustrating IoU calculation.

[0038] Figure 7This is a flowchart of the ground wave radar clutter classification and identification method of the present invention;

[0039] Figure 8 This is a schematic diagram illustrating the model training process.

[0040] Figure 9 Comparison of sea clutter recognition performance before and after model optimization;

[0041] Figure 10 A comparison chart showing the ionosphere recognition performance before and after model optimization. Detailed Implementation

[0042] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0043] It should be noted that the terms "connection" and "communication" can refer to both direct connection and direct communication between components, as well as indirect connection and indirect communication between components.

[0044] This invention provides a ground wave radar clutter classification and identification method based on a single-level convolutional neural network YOLOv3. The overall principle of the method is referenced... Figure 1 Process Reference Figure 7 Specifically, it includes the following steps.

[0045] S1: Acquire radar echo data for training and construct the training dataset.

[0046] The training dataset is used to optimize the neural network model. In this embodiment of the invention, radar echo data is collected in batches to construct the dataset.

[0047] To obtain better training results and improve the recognition rate of clutter and interference, this embodiment of the invention further preprocesses the radar echo data of the dataset to enhance the radar echo data.

[0048] Specifically, the dataset images include multiple batches of radar echo data. After constructing the training dataset images, the radar echo data images are further preprocessed, and radar RD feature maps are constructed using the preprocessed radar echo data images.

[0049] First, clarify the methods to be used for preprocessing.

[0050] Determine the energy distribution of radar data. Analyze radar echo data from consecutive batches and determine the energy distribution of the radar echo data based on the echo energy. (Reference) Figure 2The horizontal axis represents radar echo energy, and the vertical axis represents the number of echo data points with the same energy value. Statistical results show that the radar echo energy ranges from 180 to 210, and its characteristics are generally as follows: data points with high or low energy values ​​are fewer and sparsely distributed, while those with intermediate values ​​are more densely distributed. Since most data points in the radar echo data are relatively low-energy background noise, information such as ship targets, clutter, and interference accounts for a small proportion of the total data. Based on this, an adaptive method based on Jet color maps is proposed to enhance radar data.

[0051] Data preprocessing methods include the following steps:

[0052] S11: Select a color gamut length of 64 on the color map;

[0053] In terms of color segmentation, to avoid the overemphasis of some strong background points due to uniform color division, the improved algorithm highlights targets and signals in the foreground by unevenly dividing the color gamut, and enhances the display effect by smoothly transitioning the middle and background colors.

[0054] S12: For radar echo data in the same batch, take the maximum and minimum values ​​of radar echo energy to obtain the echo data range; divide the radar echo data between the maximum and minimum values ​​of echo energy into 64 equal parts.

[0055] refer to Figure 3 After the radar echo data is divided into equal parts, the interval between any two adjacent radar echo data is fixed and equal, and is defined as the energy interval.

[0056] S13: Determine whether the echo energy interval between adjacent components after equal division is less than 1. If so, extend each energy interval upwards; otherwise, it means the span of the data value is insufficient to correspond to the length of the color gamut, and extend each energy interval upwards to 1. To avoid one data value corresponding to multiple colors, each energy interval is extended upwards to ensure the integrity of the color gamut.

[0057] Example explanation: If the energy interval of the radar echo data is 65, then 65 / 64 = 1.015 > 1, so the interval is rounded up to 2; if the energy interval of the radar echo data is 60, then 60 / 64 = 0.93 < 1, so the interval is rounded up to 1.

[0058] S14: Divide the color gamut length into 1:2:2:2:1 and color it according to the selected color gamut length and the radar echo image.

[0059] When compressing the background color, this method of color gamut division can maximize the prominence of the foreground color, and the foreground color transitions smoothly. A diagram illustrating color gamut division is shown below. Figure 4 As shown.

[0060] The algorithm is used to re-divide the color gamut corresponding to energy. It can compress the colors of the densely distributed energy values, thereby increasing the contrast with the background; while the lower background and the higher radar signal or target parts are highlighted, thus enhancing the image quality.

[0061] S2: Construct a radar RD feature map using radar echo data images, and input the feature map into the YOLOv3 network; specifically in this embodiment, the RD feature map is constructed using preprocessed radar echo data images.

[0062] The YOLOv3 network architecture is called Darknet-53, which contains 53 convolutional layers. The main components of the Darknet-53 network architecture that participate in the computation consist of 52 convolutional layers and one fully connected layer used in the output.

[0063] S3: The YOLOv3 network divides the feature map into grids and sets multiple initial preset bounding boxes of different sizes for each grid.

[0064] The YOLOv3 network divides its feature map into a grid, assigning multiple initial bounding boxes of different sizes to each grid. In YOLOv3, the feature map is used for object detection and bounding box prediction. The YOLOv3 network divides the feature map into grids of different sizes and predicts the object's class and location on each grid. Through classification and regression operations on the feature map, the class and bounding box of each object are determined.

[0065] The YOLOv3 network extracts features from feature maps by dividing them into 416x416 grids. Downsampling is commonly used in deep convolutional neural networks to extract deep image features and combat interference from small perturbations in the network.

[0066] Since sea clutter, ground clutter, ionospheric clutter, and radio frequency interference in the RD spectrum feature map of ground wave radar usually appear as stripes or blocks, multiple identification methods are set for each grid in the feature map, and different preset bounding boxes are set for each size.

[0067] Specifically, for the YOLOv3 network's feature maps, it sets three recognition methods (small, medium, and large) for each grid in the feature map, and sets three preset bounding boxes for each size. That is, the size of the preset bounding boxes is fixed.

[0068] With an input of 416*416, after downsampling at 32x, 16x, and 8x, each grid cell generates a total of (52*52 + 26*26 + 13*13)*3 = 10647 proposal boxes. The parameters are shown in the table below:

[0069] Table 1. Conversion Table between Feature Map Size and Preset Bounding Box Parameters

[0070]

[0071]

[0072] Smaller feature layers (13*13) typically identify more pixels corresponding to objects in the original input image, resulting in a larger receptive field and thus higher sensitivity for large objects during object recognition. Conversely, larger feature layers (e.g., 52*52) identify fewer pixels corresponding to objects in the original image, resulting in a smaller receptive field and thus stronger ability to recognize small objects. However, the prediction accuracy of bounding boxes (anchor boxes) is not high. Therefore, model optimization is necessary to predict object boundaries more accurately.

[0073] S4: Classify and label the real radar echo images according to the clutter signal type, obtain the coordinate range on the real radar echo image of each clutter signal, convert the coordinate range into the length and width of the label box, and determine the real label box based on the length and width.

[0074] Specifically, the process of obtaining the ground truth bounding boxes is as follows: By analyzing the radar echo data image, the coordinate range of the clutter signal on the image is determined, namely the coordinates of the upper left and lower right corners. Then, the coordinate range is converted into the length and width of the bounding box, and the difference is calculated. Finally, based on the calculated length and width, the position (center coordinates) and size of the ground truth bounding box are determined. This allows the coordinate range of the clutter signal to be converted into the corresponding ground truth bounding box, enabling target detection and bounding box prediction in YOLOv3.

[0075] Randomly select the size of the real marker box as the initial value of the preset bounding box size; calculate the matching degree d between the real marker box and the preset recognition box; assign the clutter type corresponding to the real marker box to the preset recognition box with the smallest error, and use it as the clutter recognition type of the preset recognition box.

[0076] The real radar echo image mentioned here refers to an echo image formed directly from radar echo data in the dataset without any data preprocessing or image preprocessing.

[0077] The purpose of model optimization is to more clearly define the boundaries of various clutter signals on RD images. This paper uses the K-means clustering algorithm to calculate the preset bounding boxes for cluster initialization, aiming to achieve more accurate prediction of object boundaries and thus optimize the model. When using K-means clustering to calculate the size of an object, it can be obtained by calculating the intersection-union ratio (IUU) of two bounding boxes, thereby predicting the shape of the object. The specific method is as follows.

[0078] Clutter signal types include: sea clutter, ground clutter, ionospheric clutter, radio frequency interference clutter, etc., which are signal types known in the prior art. All types of clutter signals appear as stripes on the RD feature map, and the position of each type of clutter on the RD feature map differs depending on the clutter type. Therefore, all different types of clutter signals can be marked with rectangular boxes.

[0079] Each signal image corresponds to an extreme value in the horizontal and vertical coordinates, and the bounding box can be determined based on the coordinate values.

[0080] w = b x -a x ;

[0081] h = a y -b y ;

[0082] Among them, a x b is the minimum value of the x-coordinate of the rectangle. x a is the maximum value of the x-coordinate of the rectangle. y b is the maximum value of the rectangle's ordinate. y This represents the minimum value of the rectangle's ordinate, where w and h are the lengths of the rectangle's two sides. (See reference.) Figure 5 The coordinates shown.

[0083] Since the same type of clutter and different types of clutter are scattered across the real radar echo map, multiple real bounding boxes can be obtained based on the above calculations.

[0084] Within the actual bounding boxes, K boxes are randomly selected as the initial values ​​for K preset recognition boxes. There will be a displacement deviation between the actual bounding boxes and the preset recognition boxes. (Reference) Figure 4 The dashed rectangles represent the projections of predefined bounding boxes (with initial values ​​assigned to the ground truth bounding boxes) onto the RD feature map, obtained using the K-means distance algorithm. The solid rectangles represent the predefined bounding boxes obtained after image preprocessing. Based on IoU calculation, the matching degree d between the ground truth bounding boxes and the predefined bounding boxes is calculated.

[0085] IoU = Area where the projection of the true bounding box onto the RD feature map intersects with the preset recognition box / Area where the projection of the true bounding box onto the RD feature map merges with the preset recognition box.

[0086] d = 1 - IoU.

[0087] Specifically, the Intersection over Union (IoU) ratio is used to measure the degree of matching between the projection of the ground truth bounding box onto the RD feature map and the preset recognition box. IoU can be calculated as the area where the projection of the ground truth bounding box onto the RD feature map intersects with the preset recognition box, divided by the area where the projection of the ground truth bounding box onto the RD feature map merges with the preset recognition box.

[0088] Based on the IoU calculation, a matching degree can be obtained. Let d represent the matching degree, which can be calculated as d = 1 - IoU. Thus, the closer d is to 0, the higher the matching degree; conversely, the closer d is to 1, the lower the matching degree.

[0089] This metric can be used to measure the degree of matching between the preset recognition box and the ground truth bounding box, thereby adjusting and optimizing the box during training to more accurately predict the location and bounding box of the target.

[0090] refer to Figure 6 The vertically oriented rectangles represent the preset recognition boxes, and the horizontally oriented rectangles represent the actual marking boxes.

[0091] Sort the errors d between each preset recognition box and the real mark box obtained in the previous step in descending order, select the preset recognition box with the smallest error, assign the real mark box classification to it, and then perform the same operation on all real mark boxes, and record the correspondence between the preset recognition boxes and the real mark boxes.

[0092] Repeat step S4 until all preset recognition boxes have corresponding real clutter types.

[0093] S5: Calculate the median of the width and height of all real marker boxes corresponding to a single preset recognition box, and use it as the size of the preset recognition box.

[0094] Corrosion protection can be verified by the IoU values ​​of the true bounding boxes. Calculate the IoU value for each true bounding box, select the one with the highest IoU value (representing its class of anchor boxes), and then calculate the average of all the highest IoU values ​​for true bounding boxes; this average represents the anchor box accuracy. Anchor boxes for various clutter types can be completely calibrated using the YOLO algorithm. The closer the value is to 1, the higher the accuracy.

[0095] The above training process yields a radar clutter signal identification model. This method enables rapid classification and identification of ionospheric clutter, sea clutter, and radio frequency interference in large amounts of ground-wave radar data.

[0096] The dataset was expanded using conventional methods such as scaling, flipping, cropping, and translation to obtain as many samples of ground wave radar frequency domain data as possible. Training was performed using as many samples as possible that contained sea clutter, ground clutter, ionospheric interference, and radio frequency interference. The training parameters are set as shown in the table below:

[0097] Table 2 Training Parameter Settings

[0098]

[0099]

[0100] The weight decay regularization term is set to 0.0005 to proportionally reduce the parameters during the learning process to prevent overfitting; momentum is set to 0.9 to control the speed of gradient descent; the initial learning rate is set to 0.001. After the number of iterations reaches the burn_in setting of 1000, the learning rate is updated according to the adjustment strategy set in the policy, and after the number of iterations reaches the value set in steps, the learning rate is adjusted according to the decay coefficient set in scales; the entire training process stops when the number of iterations reaches max_batches. At 1600 iterations, the learning rate decays tenfold, and at 1800 iterations, the learning rate decays tenfold again based on the previous learning rate.

[0101] refer to Figure 8 At 2000 iterations, the overall model showed a generally stable convergence trend. By 800 iterations, the model's loss value had stabilized and eventually remained around 0.8. The average accuracy also improved continuously as the model converged. However, since some categories with low signal-to-noise ratios in the labeled RD images were difficult to label, and some sample categories overlapped in the RD images, the average accuracy was affected. By 1000 iterations, the average accuracy had reached over 80%, and finally stabilized at 84.01%.

[0102] refer to Figure 10 Figures (a) and (c) show the recognition results before optimization, while figures (b) and (d) show the recognition results after optimization. Figures (a) and (c) show that the unoptimized model performed poorly in recognizing low-energy radio frequency signals when the radio frequency signal and sea clutter signal overlapped, almost failing to recognize them correctly. Figures (b) and (d) show a significant improvement in the model's ability to recognize both overlapping and low-energy radio frequency signals.

[0103] refer to Figure 10Figures (a) and (c) show the recognition results before optimization, while figures (b) and (d) show the recognition results after optimization. Figures (a) and (c) show that the model before optimization had problems recognizing or not recognizing the ionosphere across the image completely, while figures (b) and (d) show that the optimized model significantly improved upon these problems.

[0104] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for classifying and identifying ground wave radar clutter based on a single-level convolutional neural network, characterized in that, Includes the following steps: Acquire radar echo data for training and construct training dataset images; Radar RD feature maps are constructed using radar echo data images, and the feature maps are input into the YOLOv3 network; The YOLOv3 network divides the radar RD feature map into grids and sets multiple initial preset recognition boxes of different sizes for each grid. According to the clutter signal type, the real radar echo image is classified and labeled to obtain the real label box, including: determining the coordinate range of each clutter signal on the real radar echo image, converting the coordinate range into the length and width of the label box, and determining the real label box based on the length and width. Randomly select the size of the real-world bounding box as the initial value for the preset bounding box size; calculate the matching degree between the real-world bounding box and the preset recognition box. ; Arrange the matching degree d between each preset recognition box and the real mark box in descending order, and select the preset recognition box with the smallest error; assign the clutter signal type corresponding to the real mark box to the preset recognition box with the smallest error as the clutter recognition type of the preset recognition box; Repeat until the clutter recognition type of all preset recognition boxes is confirmed; The methods for calculating the matching degree between the real bounding box and the preset recognition box include: = Area where the projection of the true bounding box onto the RD feature map intersects with the preset recognition box / Area where the projection of the true bounding box onto the RD feature map merges with the preset recognition box; ； This indicates the degree of matching between the actual marker box and the preset marker box.

2. The ground wave radar clutter classification and identification method based on a single-level convolutional neural network as described in claim 1, characterized in that, Further steps include: The median value of the width and height of all real marker boxes corresponding to a single preset recognition box is calculated and used as the size of the preset recognition box.

3. The ground wave radar clutter classification and identification method based on a single-level convolutional neural network as described in claim 1, characterized in that, The dataset images include multiple batches of radar echo data. After constructing the training dataset images, the radar echo data images are further preprocessed, and radar RD feature maps are constructed using the preprocessed radar echo data images. Data preprocessing methods include: Choose a color gamut length of 64; For radar data from the same batch, the maximum and minimum values ​​of radar echo energy are taken; the data between the maximum and minimum echo energy values ​​are then divided into 64 equal parts. Determine whether the echo energy interval between adjacent components after equal division is less than 1. If so, perform upward verification for each energy interval; otherwise, expand each energy interval upward to 1. The color gamut length is divided into a 1:2:2:2:1 ratio, and the color is applied according to the selected color gamut length and the radar echo image.

4. The ground wave radar clutter classification and identification method based on a single-level convolutional neural network as described in claim 1, characterized in that: Methods for setting preset recognition boxes include: For each grid in the feature map, set three recognition methods: large, medium, and small, and set three preset bounding boxes for each size.

5. The ground wave radar clutter classification and identification method based on a single-level convolutional neural network as described in claim 1, characterized in that: The clutter signals include, but are not limited to: sea clutter, ground clutter, ionospheric clutter, and radio frequency interference clutter.

Images