A method for detecting target using shipborne ground wave radar spectra

Abstract

A shipborne ground wave radar spectrum target detection method includes establishing a ship target RD spectrum data set, dividing it into a training set and a test set; constructing a target detection model based on an improved YOLOv8 shipborne ground wave radar; extracting features of the ship target in the training set; using the training set and the test set to train the target detection model of the shipborne ground wave radar based on the improved YOLOv8 to obtain a weight file; using the training result and the weight file of the shipborne ground wave radar target detection model based on the improved YOLOv8 to perform target detection on the RD spectrum image to be detected. The present invention combines the motion characteristics of the shipborne platform, mines different sample features of the shipborne ground wave radar target echo spectrum under various backgrounds such as sea clutter broadening area, background noise area and multi-target gathering area, and provides an effective method for realizing the research on shipborne ground wave radar target detection based on improved YOLOv8.

Description

Technical Field

[0001] The invention relates to a target detection method for a shipborne ground wave radar, in particular to a method for detecting a target in a shipborne ground wave radar spectrum, and belongs to the technical field of ground wave radars. Background Art

[0002] Ground wave radars rely on the characteristics of vertically polarized high-frequency electromagnetic waves diffracted along the coastal surface to achieve over-the-horizon target detection and remote sensing of ocean dynamic processes. They have a wide coverage area, long-range detection function and all-weather operation. According to the different installation locations, ground wave radars are mainly divided into shore-based and ship-borne forms. Ship-borne ground wave radars not only have the advantages of shore-based ground wave radars, but can also use the mobility of ship-borne platforms to monitor a wider sea area in real time and continuously. Ship-borne ground wave radars have become a new development direction in the radar field that has attracted much attention because of their good development potential and application prospects. Under the combined influence of multiple influencing factors such as the motion characteristics of the ship-borne platform, the target's own properties, sea clutter and environmental interference, the radar echo signals of marine ship targets are often weakened, obscured, or even submerged in the widened sea clutter. This phenomenon increases the difficulty of target detection by ship-borne ground wave radars, thereby reducing the ability to monitor targets. Under high sea conditions and complex detection backgrounds, the performance of ship-borne ground wave radars is limited. Aggregated ship targets form a multi-target area. In the range-Doppler spectrum, the echo spectra of ship targets overlap with each other, making it difficult to identify echo morphological features and perform target detection. In addition, due to the motion characteristics of the ship-borne platform, the broadened first-order sea clutter submerges a large number of ship target signals. The traditional constant false alarm rate method is prone to produce a large number of false alarms and missed detections under the above environment, and the quality of the target signal echo spectrum is therefore reduced.

[0003] At present, the target detection methods of high-frequency ground wave radar developed at home and abroad are undergoing a major transformation from traditional methods to deep learning methods. In 2017, He Mengke et al. used CNN to classify clutter, preprocessed the RD spectrum and trained and learned it through CNN, and compared and analyzed the final network results with the classical algorithm; in 2020, Zhang L et al. proposed a cascade high-frequency ground wave radar target detection strategy based on a semi-supervised self-distillation learning algorithm. This method is based on the constant false alarm rate detection results and combines semi-supervised learning to classify real targets and other targets. This method has a certain reference value; in 2021, Li Qingzhong et al. proposed a cascade classifier for rapid detection of sea surface targets based on the error self-correcting extreme learning machine network. According to the morphological characteristics of the point target, a first-level classifier based on the point morphological detection operator was designed, and then the error self-correcting extreme learning machine network was used as the second-level classifier to achieve accurate identification of candidate targets. This method is suitable for target detection under a small amount of data. In 2021, Wu M et al. proposed an intelligent detection algorithm for hidden targets in strong clutter and complex interference environments. The algorithm is divided into two stages: preprocessing and target detection. The fused features are sent to the extreme learning machine as a secondary classifier to distinguish between real and false target points. In 2023, Ren Jihong and others combined high-frequency ground wave radar target detection with Yolov3 and achieved target detection by adding a feature pyramid structure and modifying the loss function.

[0004] The above research results prove the feasibility and effectiveness of deep learning methods in shore-based ground wave radar target detection. However, there are few shipborne ground wave radar target detection methods suitable for sea clutter and multi-target gathering areas at home and abroad. The above studies rarely combine shipborne ground wave radar with deep learning, and there are relatively few discussions on the learning mechanism and detection of shipborne ground wave radar target features. Therefore, it is urgent to develop an intelligent detection method for shipborne ground wave radar ship targets. Summary of the invention

[0005] The purpose of the present invention is to provide a shipborne ground wave radar spectrum target detection method, which takes into account factors such as the interference of sea clutter broadening caused by the ship's motion platform, makes full use of the feature learning mechanism of YOLOv8, solves the problem that a large number of ship targets are submerged in the broadened sea clutter, and thus realizes target detection of shipborne ground wave radar.

[0006] A shipborne ground wave radar spectrum target detection method, characterized in that it comprises the following steps:

[0007] Step 1: Create a ship target RD spectrum dataset.

[0008] The down-converted time-domain digital signals collected by each channel of the shipborne ground wave radar are subjected to pulse compression and coherent accumulation to obtain the range-Doppler spectrum. Matched filtering is performed in the range dimension, and the Doppler frequency is extracted by fast Fourier transform in the Doppler dimension. The range-Doppler spectrum (RD) is plotted to construct a sample data set of ship targets based on the characteristics of the shipborne ground wave radar.

[0009] Step 2: Preprocess the established data set and divide it into training set and test set.

[0010] The RD spectrum images of the shipborne ground wave radar are matched one by one with the corresponding TXT label files, and the training set, test set and validation set are divided in a ratio of 8:1:1.

[0011] 500 RD spectrum images of shipborne ground wave radar were selected with a resolution of 1338 pixels × 1025 pixels.

[0012] Step 3: Build a target detection model based on the improved YOLOv8 shipborne ground wave radar, replace some C2f modules of the backbone network with the convolutional network FasterNet to extract features of ship targets on the initial RD spectrum, and add the BiFormer attention module to the Neck network of the YOLOv8 model to purify the extracted features.

[0013] In step 3, the BiFormer includes a multilayer perceptron (MLP) layer for nonlinear transformation of features; a normalization layer (LN) for standardization of features; a bi-level routing attention (BRA) layer responsible for feature routing according to the top k related connections; and a depthwise separable convolution (DWConv) layer for efficient feature extraction.

[0014] Step 4: Extract features of ship targets in the training set:

[0015] (1) After the image is input, it first passes through a 2-layer CBS module;

[0016] (2) The C2f module is then used to ensure the size consistency of the input and output feature maps of the shipborne ground wave radar RD spectrum, and a CBS module is superimposed on it;

[0017] (3) Then, the shallow and middle layers of feature information are further extracted through the double-layer stacking of FasterNet and CBS, and then through the C2f and SPPF modules;

[0018] (4) The outputs of the two FasterNets and the SPPF output are respectively fed into the Neck network to jointly complete the task of converting and classifying the ship target features.

[0019] In order to solve the problem that a large number of targets are covered in the sea clutter broadening, the BiFormer attention module is added to the Neck network of the YOLOv8 model to purify the extracted features;

[0020] The Neck network constructs an efficient ship target feature pyramid through PAFPN, deeply integrating the shallow, middle and high-level ship target features. In order to improve the performance of the lightweight network in feature fusion and expression ability, the present invention introduces the BiFormer module in the Neck layer.

[0021] The process of integrating the Neck network into BiFormer is as follows:

[0022] (1) When the feature data enters the C2f module, it first undergoes a series of convolution operations to extract preliminary feature representations.

[0023] (2) After standardization processing is achieved through the LN layer, it enters the BRA two-layer routing module.

[0024] (3) The feature data is highly purified through the MLP layer.

[0025] (4) The high-level features are fused with the mid-level and shallow-level features of the ship target through a bottom-up strategy.

[0026] (5) The ship target fusion features are laterally connected through a top-down strategy and output to the Head layer.

[0027] Step 5: Use the training set and the test set to train the target detection model of the shipborne ground wave radar based on the improved YOLOv8 to obtain the weight file;

[0028] Step 6: Use the training results and weight file of the shipborne ground wave radar target detection model based on improved YOLOv8 to perform target detection on the RD spectrum image to be detected.

[0029] In the step 4, the CBS used is composed of a convolution layer (Convolution Layer, Conv), a batch normalization layer (Batch Normalization, BN) and a SiLU (Sigmoid Linear Unit) activation function.

[0030] The SiLU function activation formula used is as follows:

[0031] .

[0032] In step 4, the C2f structure includes 2 CBS convolution blocks and n BottleNeck units, followed by a superimposed CBS module.

[0033] In step 4, the core component of the FasterNet block is a partial convolution layer (PConv), followed by two point-wise convolution (PWConv) layers.

[0034] Compared with the prior art, the innovation of the present invention is embodied in the following aspects:

[0035] 1. Improved the accuracy and robustness of ship target detection. Traditional shipborne ground wave radar target detection methods often rely on complex feature engineering and manually designed feature extractors. This method is not only inefficient, but also difficult to cope with complex and changeable marine environments. In order to overcome these challenges, this study proposed a target detection method based on a lightweight feature extraction network. The network adopts an efficient convolutional neural network structure, which greatly reduces the number of network parameters by introducing technologies such as deep separable convolution, bottleneck layer, and residual connection. At the same time, combined with the semantic feature extraction method, the network can automatically learn and extract the deep-level features of the target, thereby improving the accuracy and robustness of target detection.

[0036] 2. The detection effect is better for submerged targets in the sea clutter broadening area. Due to the motion characteristics of the ship-borne platform, a large number of targets are in the broadened sea clutter. The signal-to-noise ratio of ship targets is reduced due to the broadening of sea clutter, and sea clutter and ship targets are easily confused, making the detection of weak ship targets more complicated and difficult. In addition, the ship targets in the RD spectrum of the shipborne ground wave radar are small in size and difficult to capture, resulting in a lack of sufficient feature information for weak ship targets, which greatly increases the difficulty of detection.

[0037] 3. Improved the detection performance in multi-target clustered areas. As the number of ship detection targets increases, the overlap of their echo ranges becomes more and more significant. The multi-target area shows a significant widening trend in both the distance dimension and the Doppler dimension, and its overall echo morphology is quite different from that of a single target. Therefore, the BiFormer attention module is used to enrich the feature information of ships in the multi-target clustered area, thereby improving the target detection performance in this area.

[0038] Therefore, the present invention combines the motion characteristics of the shipborne platform and mines different sample characteristics of the target echo spectrum of the shipborne ground wave radar under various backgrounds such as the sea clutter broadening area, the background noise area and the multi-target gathering area, so as to provide an effective method for the research on shipborne ground wave radar target detection based on the improved YOLOv8. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The following is a flow chart of target detection of shipborne ground wave radar according to the present invention.

[0040] Figure 2 This is a target network diagram of the shipborne ground wave radar based on the improved YOLOv8 of the present invention.

[0041] Figure 3 It is the RD spectrum of the shipborne ground wave radar of the present invention.

[0042] Figure 4 This is a target detection result diagram of the improved shipborne ground wave radar RD spectrum of the present invention.

[0043] Figure 5 This is a comparison chart of target detection in the sea clutter broadening area of ​​shipborne ground wave radar.

[0044] Figure 6 This is a comparison chart of target detection in multiple target areas of shipborne ground wave radar. DETAILED DESCRIPTION

[0045] The method of the present invention is further described below in combination with formulas, drawings and specific data:

[0046] like Figure 1 As shown, the method of the present invention includes constructing a range-Doppler domain spectrum and forming a shipborne ground wave radar RD spectrum data set suitable for YOLOv8, building a feature extraction network based on an improved YOLOv8, adding a BiFormer attention module to the Neck network, and then training a target detection model of a shipborne ground wave radar based on the improved YOLOv8 and realizing prediction. The specific steps are as follows:

[0047] Step 1: Establish an echo spectrum image dataset of ship target samples. Use the shipborne ground wave radar signal with a period of T seconds, a working carrier frequency of 4.7 MHz, a range resolution of 2.5 km, and a coherent integration time of 262.144 s. After matched filtering and coherent integration, the range-Doppler spectrum is formed. The data of the range-Doppler spectrum is selected, and the echo amplitude value is calculated and rounded to be recorded as s ( r,v ), unit: dB, s The number of n , which is in the form of Figure 3As shown in the figure. The range-Doppler spectrum is obtained by pulse compression and coherent accumulation of the down-converted time-domain digital signals collected by each channel of the radar. Matched filtering is performed in the range dimension, and the Doppler frequency is extracted by fast Fourier transform in the Doppler dimension, and the range-Doppler spectrum (RD) is plotted, and the ship target data set is established based on it.

[0048] Step 2: Preprocess the established data set to create training sets and test sets.

[0049] 500 RD spectrum images of shipborne ground wave radar (resolution 1338 pixels x 1025 pixels) were selected, and the labeled RD spectrum image information and label information were obtained through the adaptive extraction method of ship spectra, and the information was stored in TXT files. Therefore, each RD spectrum corresponds to its corresponding TXT label file one by one, and the training set, test set and validation set are divided in a ratio of 8:1:1.

[0050] Step 3: Build a feature extraction network based on the improved YOLOv8 to extract features of the ship targets on the initial RD spectrum.

[0051] The YOLOv8 network structure used in the present invention is as follows: Figure 2 As shown, the BackBone network is responsible for feature extraction, and the specific steps are as follows.

[0052] Step 4: (1) After the image is input, it first passes through a 2-layer CBS module. CBS consists of a convolution layer (ConvolutionLayer, Conv), a batch normalization layer (Batch Normalization, BN) and a SiLU (Sigmoid Linear Unit) activation function. This module first uses the convolution layer to preliminarily extract the texture and color of the ship target echo morphology. Secondly, the normalization process solves the problem of gradient vanishing and explosion during the back propagation process to obtain better convergence speed and effect. Finally, the SiLU activation function is used to solve the problems of saturated output and gradient vanishing. The SiLU function activation formula is as follows, where x represents the input signal:

[0053] .

[0054] In CBS, the shipborne ground wave radar feature map undergoes two rounds of 3 × 3 convolution kernels with step sizes of 1 and 2 respectively, and the number of channels becomes 4 times the original one.

[0055] (2) Secondly, through the C2f module, C2f integrates the C3 module of YOLOv5 and the ELAN idea of ​​YOLOv7. By adding multiple shortcut paths, the shallow features of the ship target are reused, and the gradient problem in the deep network is effectively alleviated. The C2f network ensures the size consistency of the input and output feature maps of the shipborne ground wave radar RD spectrum. By finely controlling the length of the gradient path, the convergence problem caused by the increase of the depth of the model is effectively solved. The C2f structure includes 2 CBS convolution blocks and n ( ) BottleNeck unit, followed by a CBS module.

[0056] (3) Then, some C2f modules in the original YOLOv8 BackBone network are replaced with the lightweight feature extraction network FasterNet. The shallow and middle layer feature information is further extracted by stacking FasterNet and CBS, and then fed into the Neck network respectively. The core component of each FasterNet block is a partial convolution layer (PConv), followed by two point-wise convolution (PWConv) layers. These three layers are connected in series to form the structure of the Inverted Residual Block, in which the middle layer improves the representation capability of ship target features by expanding the number of channels. At the same time, the connection of shortcuts in the FasterNet block realizes the reuse of input ship target features, further improving the utilization efficiency of ship target feature information.

[0057] (4) In order to maintain feature diversity and reduce the delay of model inference, C2f and SPPF (Spatial Pyramid Pooling - Fast) are set after the second group of FasterNet and CBS are stacked to jointly complete the conversion and classification tasks of ship target features. The concat module of SPPF will fuse the feature maps of the four dimensions after three pooling operations and before pooling, and finally perform another conv module to increase the dimension, thereby obtaining high-level ship target feature information.

[0058] Step 5: In view of the fact that a large number of targets are covered in the sea clutter broadening, a BiFormer attention module is added to the Neck network of the YOLOv8 model to purify the extracted features;

[0059] The improved YOLOv8 lightweight feature extraction network inputs three effective feature layers into the Neck layer after feature extraction. The Neck part constructs an efficient ship target feature pyramid through PAFPN, deeply integrates the shallow, middle and high-level ship target features, and realizes the effective fusion of multi-scale and multi-level ship target information.

[0060] In order to improve the performance of lightweight networks in terms of feature fusion and expression capabilities, the present invention introduces a BiFormer module in the Neck layer. This module divides the input image into The BiFormer attention module builds and optimizes a region-level association graph of target vessel features to improve the pertinence and efficiency of feature extraction. Specifically, the module prunes the association graph to ensure that each node retains only the top k most relevant connections ( ). This design ensures that each routing region can focus on the k regions that are most relevant to it, thereby improving the pertinence and efficiency of feature processing.

[0061] BiFormer contains a multilayer perceptron (MLP) layer for nonlinear transformation of features; a layer normalization (LN) layer for standardization of features; a bi-level routing attention (BRA) layer for feature routing based on the top k related connections; and a depthwise separable convolution (DWConv) layer for efficient feature extraction. The Neck network flow integrated into BiFormer is as follows:

[0062] (1) When the feature data enters the C2f module, it first undergoes a series of convolution operations to extract preliminary feature representations. Subsequently, these features are sent to the concat module for feature fusion to integrate information at different levels and scales. After fusion, the feature data undergoes depthwise separable convolution again to further reduce the number of model parameters and computational complexity while maintaining the discriminative power of ship target features.

[0063] (2) Standardization is achieved through the LN layer to ensure that the feature data remains stable in subsequent processing. Subsequently, the ship target feature data enters the BRA double-layer routing module. In this module, the model collects the key-value pairs in the first k relevant windows and uses sparsity operations to directly skip the calculations that are least relevant to the current area, thereby significantly reducing the computational complexity and saving parameter resources.

[0064] (3) The MLP layer highly purifies the ship target feature data. Since the MLP layer adopts a fully connected structure, it can effectively integrate and strengthen the ship target feature information, and finally output a highly purified ship target feature representation, providing strong support for subsequent ship target detection tasks.

[0065] (4) Subsequently, the high-level features are fused with the middle-level and shallow-level features of the ship target through a bottom-up strategy. On this basis, the upsampled high-level features of the ship target are directly concat-superimposed with the middle-level features and sent to the BiFormer attention module. Subsequently, two paths are branched out, one of which is combined with the shallow features of C3 through upsampling, achieving a deep fusion of shallow details of the target features and high-level semantic information, thereby enhancing the model's detection accuracy for ship targets.

[0066] (5) Finally, the fusion features of the ship target are horizontally connected through a top-down strategy and output to the Head layer. After the shallow feature C3 passes through the BiFormer attention module branch, one path is directly output to the P3 layer, while the other path achieves dimensionality reduction through a 3 × 3 convolution operation with a step size of 2. This step solves the problem of feature discontinuity and image distortion that may be introduced by upsampling, further enriches the resolution and semantic level of each layer of the shipborne ground wave radar feature map, and ensures the model's effective detection of ship targets at different resolutions. Then, the target feature information is transmitted to P4 and the next layer CBS, and finally transmitted to P5 through the same steps.

[0067] So far, all steps of YOLOv8 in the Neck ship target feature fusion network have been completed.

[0068] Step 6: Use the ship target RD spectrogram dataset to train the target detection model of the shipborne ground wave radar based on the improved YOLOv8.

[0069] The RD spectrum dataset of ship-borne ground wave radar ship targets is input into the improved YOLOv8 network for target detection model training. This experiment is completed in the operating system Windows 10, and PyTorch is used as the deep learning framework. The processor is Intel(R) Xeon(R) Gold 5122 CPU @ 3.60 GHz, the memory is 16G, and the NVIDIA TITAN RTX graphics processor is used for accelerated computing. The GPU acceleration is CUDA 11.3+cuDNN. According to the configuration requirements of the experimental training data set platform, the network model configuration parameters are continuously adjusted. The final model training configuration parameters are set as follows: the number of target categories is 1, the detection category is "target", and the main parameters are shown in the following table.

[0070] Table 1 SSWR_YOLO model parameters

[0071] Parameter name Value / Form Batch 16 Epoch 1000 pretrained YOLOv8s.pt workers 8 Lr0 0.001 Lf 0.01 Warmup_epochs 3 Warmup_momentum 0.8 imgsz 640 .

[0072] Through the above experimental training, the weight model file of the PyTorch version based on the improved YOLOv8 shipborne ground wave radar is obtained, and this file is used for target detection later.

[0073] Step 6: Use the training results of the shipborne ground wave radar target detection model based on the improved YOLOv8 to perform target detection on the RD spectrum image to be detected.

[0074] Use the PyTorch version of the weight model file obtained in step 5 to perform shipborne ground wave radar target detection. Figure 3 This is a measured RD spectrum of a shipborne ground wave radar. Figure 4 This is the overall detection result diagram based on the improved YOLOv8, where the red solid line box is the target detection result, and the detection category and confidence value are marked on it.

[0075] Figure 5 This is a comparison chart of target detection in the sea clutter broadening area of ​​shipborne ground wave radar. Figure 5 (a) is the sea clutter broadening area before detection, which contains complex sea clutter interference and potential ship targets. Figure 5 (b) To improve the detection results of YOLOv8, since there are many ship targets in the area and they are densely distributed, the category labels and confidence information generated during the target detection process occlude each other. In order to maintain the clarity and readability of the image, these detailed information are hidden in the comparison image. Figure 5 (c) is the detection result of the traditional constant false alarm rate (CFAR) algorithm, and the black circle is the CFAR detection result. By comparison, it can be found that although the CFAR algorithm has a wide range of applications in the field of radar signal processing, its performance is relatively limited in target detection in the sea clutter broadening area. In the sea clutter broadening area, the radar echo signal of the ship target at sea is often weakened, obscured, or even submerged in the broadened sea clutter. The improved YOLOv8 algorithm can accurately identify the number of targets and accurately locate them, so it has more advantages.

[0076] Figure 6 This is a comparison chart of target detection in multiple target areas of shipborne ground wave radar. Figure 6 (a) is the multi-target area before detection, which includes multiple ship targets and possible background noise or interference. Figure 6(b) shows the detection results of the improved YOLOv8 algorithm. It can be seen that the improved YOLOv8 algorithm not only detects more real ship targets, but also can accurately identify the position, size and echo shape of ship targets. In addition, for targets with overlapping echoes, the improved YOLOv8 algorithm can identify and distinguish these overlapping targets one by one. Figure 6 (c) is the detection result of the CFAR algorithm. Its performance is relatively limited in target detection in multi-target areas. Compared with the improved YOLOv8 algorithm, the CFAR algorithm has a certain gap in the number of target recognition, position accuracy, and overlapping target recognition. By comparison, it can be seen that the improved YOLOv8 algorithm has more advantages in target detection in multi-target areas of shipborne ground wave radar.

Claims

1. A shipborne ground wave radar spectrum target detection method, characterized in that The following steps are involved: Step 1: Establish a ship target sample RD spectrum dataset; Step 2: Preprocess the established data set and divide it into training set and test set: Step 3: Build a target detection model based on the improved YOLOv8 shipborne ground wave radar: Some C2f modules of the BackBone network are replaced with the convolutional network FasterNet, and the BiFormer attention module is added to the Neck network to purify the extracted features; Step 4: Extract features of ship targets in the training set: (1) After the image is input, it first passes through a 2-layer CBS module; (2) The C2f module is then used to ensure the size consistency of the input and output feature maps of the shipborne ground wave radar RD spectrum, and a CBS module is superimposed on it; (3) Then, the shallow and middle layers of feature information are further extracted through the double-layer stacking of FasterNet and CBS, and then through the C2f and SPPF modules; (4) The outputs of the two FasterNets and the SPPF are respectively fed into the Neck network to jointly complete the task of converting and classifying the ship target features; Step 5: Use the training set and the test set to train the target detection model based on the improved YOLOv8 shipborne ground wave radar and obtain the weight file; Step 6: Use the training results and weight file of the improved YOLOv8 shipborne ground wave radar target detection model to perform target detection on the RD spectrum image to be detected.

2. The shipborne ground wave radar spectrum target detection method according to claim 1, characterized in that The step 1 is as follows: The down-converted time-domain digital signals collected by each channel of the shipborne ground wave radar are subjected to pulse compression and coherent accumulation to obtain the range-Doppler spectrum. Matched filtering is performed in the range dimension, and the Doppler frequency is extracted by fast Fourier transform in the Doppler dimension and plotted into a range-Doppler spectrum. In this way, a RD spectrum dataset of ship target samples based on the characteristics of the shipborne ground wave radar is constructed.

3. The shipborne ground wave radar spectrum target detection method as claimed in claim 1, characterized in that The step 2 includes making a one-to-one correspondence between the RD spectrum image of the shipborne ground wave radar and the matching TXT tag file, and dividing the training set, the test set and the validation set in a ratio of 8:1:

1.

4. The shipborne ground wave radar spectrum target detection method as claimed in claim 3, characterized in that In the step 2, the number of RD spectrum images of the shipborne ground wave radar selected is 500, and the resolution is 1338 pixels×1025 pixels.

5. The shipborne ground wave radar spectrum target detection method as claimed in claim 1, characterized in that In step 3, the BiFormer includes a multi-layer perceptron layer, i.e., an MLP layer, for nonlinear transformation of features; A normalization layer, namely the LN layer, is used for feature standardization; a double-layer routing attention layer, namely the BRA layer, is responsible for feature routing according to the top k related connections; and a depthwise separable convolutional layer, namely the DWConv layer, is used to achieve efficient feature extraction.

6. The shipborne ground wave radar spectrum target detection method as claimed in claim 1, characterized in that In step 4, the Neck network process described in step (4) is as follows: (1) When the feature data enters the C2f module, it first undergoes a series of convolution operations to extract preliminary feature representations; (2) After being standardized by the LN layer, it enters the BRA two-layer routing module; (3) The feature data is highly purified through the MLP layer; (4) The high-level features are integrated with the mid-level and shallow-level features of the ship target through a bottom-up strategy; (5) The ship target fusion features are laterally connected through a top-down strategy and output to the Head layer.

7. The shipborne ground wave radar spectrum target detection method as claimed in claim 1, characterized in that In step 4, the CBS used consists of a convolutional layer, a batch normalization layer and a SiLU activation function.

8. The shipborne ground wave radar spectrum target detection method as claimed in claim 7, characterized in that The SiLU function activation formula used is as follows: , where x represents the input signal.

9. The shipborne ground wave radar spectrum target detection method as claimed in claim 1, characterized in that In step 4, the C2f module includes 2 CBS convolution blocks and n BottleNeck units, followed by a superimposed CBS module.

10. The shipborne ground wave radar spectrum target detection method according to claim 1, characterized in that In step 4, the core component of the FasterNet is a partial convolutional layer, followed by two point-by-point convolutional layers.

Images