Multi-feature fusion sea surface small target detection method based on attention mechanism

Abstract

The invention discloses a multi-feature fusion sea surface small target detection method based on an attention mechanism, and the method comprises the steps: 1, carrying out the preprocessing of radar echo data, and obtaining a combined feature sequence containing rich features; 2, constructing a multi-feature fusion network model based on an attention mechanism, which comprises a feature extraction module, an attention mechanism module, a feature aggregation module and a target detection module, and processing the joint feature sequence to obtain a binary probability matrix sequence; and 3, calculating loss by using a focus loss function, and training the network model to obtain an optimal target detection model. According to the method, through multi-scale feature extraction, an attention mechanism and feature fusion, the detection precision of the radar weak and small target is effectively improved, and the method is suitable for target detection requirements in a complex background and a low signal-to-noise ratio environment.

Description

Technical Field

[0001] This method belongs to the intersection of artificial intelligence, computer vision and radar signal processing, and is a multidisciplinary cross-application technology. Background Art

[0002] Radar technology is widely used in military, civil and industrial fields due to its all-weather working ability and strong robustness to environmental changes. As an important research direction in this field, sea surface radar small target detection has attracted great attention from scholars at home and abroad. Its goal is to accurately locate small targets in complex sea backgrounds, such as small ships, buoys and unmanned boats. However, due to the variability of sea clutter and the low signal-to-noise ratio characteristics of the target, the detection task is full of challenges.

[0003] Traditional radar small target detection methods, such as constant false alarm rate (CFAR) detectors and matched filter-based detection techniques, although they perform well under certain conditions, are highly dependent on statistical models and prior assumptions. In complex sea conditions, these methods show significant limitations in dealing with sea clutter interference and weak target signals, making it difficult to meet the needs of practical applications. For example, CFAR detectors are prone to false alarms in high-clutter environments.

[0004] Current feature fusion methods are mostly based on multi-scale or multi-modal features, which improve the target saliency by integrating feature information from different sources. However, simple feature fusion strategies are prone to information redundancy or feature conflict, affecting the model's discrimination ability and computational efficiency. To this end, the attention mechanism is introduced into the feature fusion process. By dynamically adjusting feature weights, it effectively highlights key features and suppresses redundant information, significantly improving the model's target detection performance in complex backgrounds.

[0005] Although the above methods have made some progress, the existing sea surface small target detection methods based on deep learning methods still have the following shortcomings:

[0006] 1. Low signal-to-noise ratio and complex background: The target signal is weak and easily interfered by sea clutter. The robustness and detection accuracy of existing methods in complex sea conditions still need to be improved.

[0007] 2. Insufficient expression of multi-dimensional features: Current feature fusion methods are mostly based on single-scale or single-modal features, and fail to fully utilize the expressive power of cross-level features. The aggregation of local and global features is insufficient, and the multi-dimensional information expression of the target cannot be fully retained. The feature expression ability of the model still has a lot of room for improvement.

[0008] 3. Category imbalance problem: In the detection of small targets on the sea surface, the target signal usually accounts for a very small proportion, and the category imbalance problem is significant. The existing methods have weak processing capabilities for difficult-to-classify samples, and are prone to high false alarm rates or high missed detection rates, which reduces the reliability of practical applications. Summary of the invention

[0009] In order to solve the shortcomings of the above-mentioned prior art, the present invention proposes a multi-feature fusion sea surface small target detection method based on attention mechanism, in order to achieve cross-level feature fusion, so as to enhance the target characteristics and suppress clutter, thereby effectively improving the detection accuracy of radar weak small targets, and is suitable for target detection needs in complex backgrounds and low signal-to-noise ratio environments.

[0010] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0011] The multi-feature fusion sea surface small target detection method based on the attention mechanism of the present invention is characterized in that it is carried out according to the following steps:

[0012] Step 1: Preprocess the radar echo data sequence to obtain the joint feature sequence ;in, represents the joint feature of the ith pulse in the kth channel at the jth distance unit, N represents the number of pulses, and M represents the number of distance units;

[0013] Step 2: Construct a multi-feature fusion network based on the attention mechanism, including: feature extraction module, attention mechanism module, feature aggregation module, and target detection module;

[0014] Step 2.1: The feature extraction module Perform multi-level and multi-scale feature extraction to obtain primary features accordingly , intermediate features , advanced features ;

[0015] Step 2.2: The attention mechanism module , , After attention weighting, the enhanced primary features are obtained accordingly. , the enhanced intermediate features , enhanced high-level features ;

[0016] Step 2.3: Construct a feature aggregation module including local feature submodule and global feature submodule, and use it to , , Perform feature aggregation and fusion processing to obtain a joint feature sequence ;

[0017] Step 2.4: Use the target detection module to Processing is performed to obtain a binary probability matrix sequence containing the prediction results ;

[0018] Step 3: Construct the total loss function ;

[0019] Step 4: Train the multi-feature fusion network by gradient descent method and calculate To update the network parameters until Until convergence, the optimal target detection model is obtained, which is used to perform target detection on the input radar echo data.

[0020] The multi-feature fusion sea surface small target detection method based on the attention mechanism described in the present invention is characterized in that step 1 is performed according to the following steps:

[0021] Step 1.1: Obtain radar echo data sequence and perform data correction, pulse compression, and normalization operations to obtain the processed radar echo data sequence. And its corresponding label sequence ,in, represents the radar echo signal value of the ith pulse in the jth distance unit, express target marker;

[0022] Step 1.2: Align the distance unit with the Perform fast Fourier transform processing to obtain the amplitude sequence of the response distance information ,in, represents the amplitude of the ith pulse at the jth distance unit;

[0023] Step 1.3: Align the pulse direction Perform Fourier transform to obtain the Doppler spectrum sequence ,in, represents the Doppler spectrum of the ith pulse in the jth range unit;

[0024] Step 1.4: Perform centralization processing to obtain the processed Doppler image sequence ,in, represents the Doppler map of the ith pulse at the jth range unit;

[0025] Step 1.5: and Join on the channel dimension to form a joint feature sequence .

[0026] Further, step 2.1 is performed as follows:

[0027] Step 2.1.1: Use feature extraction module to Perform feature extraction to obtain primary features ={ };in, Indicates The primary characteristics of the ith pulse in the channel at the jth distance unit;

[0028] Step 2.1.1.1. After convolution operation, batch normalization and PReLU activation function processing, the first feature sequence is obtained. ={ },in, Indicates The first feature of the i-th pulse in the channel at the j-th distance unit, express The number of channels;

[0029] Step 2.1.1.2: After performing convolution operation and batch normalization, the second feature sequence is obtained. ,in, Indicates The second characteristic of the i-th pulse in the channel at the j-th distance unit;

[0030] Step 2.1.1.3: Through convolution operation Perform channel adjustment to obtain the residual term sequence ,in, Indicates The residual term of the i-th pulse in the channel at the j-th distance unit;

[0031] Step 2.1.1.4: and After the residual connection, it is processed by the PReLU activation function to obtain the primary features ;

[0032] Step 2.1.2: Follow the process in step 2.1.1 Perform feature extraction to obtain intermediate features ={ },in, Indicates The intermediate features of the channel at the i-th pulse at the j-th distance unit, express The number of channels;

[0033] Step 2.1.3: Follow the process in step 2.1.1 Perform feature extraction to obtain high-level features ={ },in, Indicates The high-level features of the ith pulse in the channel at the jth range unit, express The number of channels.

[0034] Further, step 2.2 is performed as follows:

[0035] Step 2.2.1: Use the attention mechanism module to After attention weighting, the enhanced primary features are obtained ={ | };in, represents the primary feature of the i-th pulse in the k-th channel after enhancement at the j-th distance unit;

[0036] Step 2.2.1.1. After performing average pooling operation and changing its shape, we get the query feature sequence ={ } and key feature sequences ,in, represents the query feature of the i-th pulse in the j-th distance unit, Indicates The key features of the ith pulse in the channel;

[0037] Step 2.2.1.2: Generate the attention weight matrix using formula (1) ={ },in, represents the attention weight matrix of the jth distance unit of the i-th pulse of the k-th channel;

[0038] (1)

[0039] In formula (1), represents the activation function; T represents transposition,

[0040] Step 2.2.1.3: Apply weight matrix After that, the enhanced primary features are obtained. ={ | };

[0041] Step 2.2.2: Follow the process in step 2.2.1 to After attention weighting, the enhanced intermediate features are obtained ={ | },in, Indicates The mid-level features of the i-th pulse in the channel after enhancement at the j-th distance unit;

[0042] Step 2.2.3: Follow the process in step 2.2.1 to After attention weighting, enhanced high-level features are obtained ={ | };in, Indicates The enhanced high-level features of the i-th pulse in the channel at the j-th range unit.

[0043] Further, step 2.3 is performed as follows:

[0044] Step 2.3.1: Feature aggregation module Processing to obtain global primary features ={ };in, Indicates The global primary features of the ith pulse in the channel at the jth distance unit;

[0045] Step 2.3.1.1: Local feature submodule pair After multiple convolution operations and compressing its pulse dimension, local primary features are obtained ={ },in, Indicates Channel No. The local primary characteristics of a pulse in the jth distance unit, Indicates the number of pulses after compression;

[0046] Step 2.3.1.2: Global feature submodule pair Transpose, and then perform multiple convolution operations, and then transpose again to restore the original shape to obtain the global primary features ={ };

[0047] Step 2.3.2: Follow the process in 2.3.1 After processing, the global intermediate features are obtained ={ },in, For the The global intermediate features of the i-th pulse in the channel at the j-th distance unit;

[0048] Step 2.3.3: Follow the process in 2.3.1 After processing, we get the global high-level features ={ },in, For the The global high-level features of the i-th pulse in the channel at the j-th distance unit;

[0049] Step 2.3.4: , , After concatenation in the channel dimension, the joint feature sequence is obtained ={ | },in, Indicates The joint feature of the i-th pulse in the channel at the j-th distance unit, C represents the number of channels in the joint feature sequence.

[0050] Further, step 2.4 is performed as follows:

[0051] Step 2.4.1: Target detection module After multiple convolution operations, the target prediction sequence is obtained ,in, Indicates The target prediction value of the i-th pulse in the channel at the j-th distance unit;

[0052] Step 2.4.2: After passing the softmax activation function, we get a binary probability matrix sequence containing the prediction results. ={ },in, Indicates The predicted probability of the i-th pulse in the channel at the j-th distance unit; when k=1, represents the target prediction probability of the i-th pulse in the j-th distance unit, when k=2, represents the predicted probability of non-target of the i-th pulse at the j-th distance unit.

[0053] Further, step 3 is performed as follows:

[0054] Step 3.1: Use formula (2) to construct the standard cross entropy loss :

[0055] (2)

[0056] Step 3.2: Use formula (3) to construct the cross entropy loss :

[0057] (3)

[0058] Step 3.3: Use formula (4) to construct the total loss function :

[0059] (4)

[0060] In formula (4), is the category weight of the j-th distance unit, is the adjustment factor.

[0061] An electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the radar small target detection method, and the processor is configured to execute the program stored in the memory.

[0062] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the radar small target detection method when the computer program is executed by a processor.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] 1. The present invention uses multi-level and multi-scale feature extraction and attention mechanism to refine and enhance target features, highlight key features and suppress clutter information, significantly improving the detection probability of weak targets, and is particularly suitable for scenes with complex backgrounds and low signal-to-noise ratio.

[0065] 2. The present invention retains the multi-dimensional information expression of the target and optimizes the feature expression ability of the model by aggregating local features and global features and combining cross-level feature splicing. Compared with the traditional single feature fusion method, aggregating multi-level features enables the model to comprehensively consider information at different scales, thereby improving the precise positioning and classification effect of the target.

[0066] 3. The present invention introduces the focal loss function to give higher weights to difficult-to-classify samples to address the problem of class imbalance, effectively reducing the false alarm rate and missed detection rate. Compared with conventional loss functions, focal loss can better focus on difficult-to-classify samples, especially effectively reducing false alarms and missed detections in target detection, thereby significantly improving detection accuracy and ensuring the reliability of detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 is a flow chart of the method of the present invention;

[0068] Figure 2 It is a structural diagram of the feature extraction module in the present invention;

[0069] Figure 3 This is a structural diagram of the attention mechanism module in the present invention;

[0070] Figure 4 It is a schematic diagram of the feature aggregation module in the present invention;

[0071] Figure 5 This is a flow chart of the target detection module in the present invention. DETAILED DESCRIPTION

[0072] In this embodiment, a method for detecting small sea targets using multi-feature fusion with an attention mechanism is implemented through a series of optimization designs, including radar data preprocessing, multi-scale feature extraction, the introduction of an attention mechanism, and cross-level feature fusion, to achieve efficient detection of small radar targets in a complex sea environment. Specifically, Figure 1 As shown, the method is performed in the following steps:

[0073] Step 1: Preprocess the radar echo data sequence to obtain the joint feature sequence ;in, represents the joint feature of the ith pulse in the kth channel at the jth distance unit, N represents the number of pulses, and M represents the number of distance units;

[0074] Step 1.1: Obtain radar echo data sequence and perform data correction, pulse compression, and normalization operations to obtain the processed radar echo data sequence. And its corresponding label sequence ,in, represents the radar echo signal value of the ith pulse in the jth distance unit, express In this embodiment, the size of each radar echo data is .

[0075] Step 1.2: Align the distance unit with the Perform fast Fourier transform processing to obtain the amplitude sequence of the response distance information ,in, represents the amplitude of the ith pulse at the jth distance unit;

[0076] Step 1.3: Align the pulse direction Perform Fourier transform to obtain the Doppler spectrum sequence ,in, represents the Doppler spectrum of the ith pulse in the jth range unit;

[0077] Step 1.4: Perform centralization processing to obtain the processed Doppler image sequence ,in, represents the Doppler map of the ith pulse at the jth range unit;

[0078] Step 1.5: and Join on the channel dimension to form a joint feature sequence .

[0079] Step 2: Construct a multi-feature fusion network based on the attention mechanism, including: feature extraction module, attention mechanism module, feature aggregation module, and target detection module;

[0080] Step 2.1: Feature extraction module Perform multi-level and multi-scale feature extraction to obtain primary features accordingly , intermediate features , advanced features The structure is as follows Figure 2 As shown;

[0081] Step 2.1.1: Use feature extraction module to Perform feature extraction to obtain primary features ={ };in, Indicates The primary characteristic of the i-th pulse in the channel at the j-th distance unit. In this embodiment, =16.

[0082] Step 2.1.1.1. After performing convolution operation and then processing with batch normalization and PReLU activation function, the first feature is obtained. ={ },in Indicates The first feature of the i-th pulse in the channel at the j-th distance unit, express The number of channels;

[0083] Step 2.1.1.2: After performing convolution operation and batch normalization, the second feature is obtained. ,in Indicates The second characteristic of the i-th pulse in the channel at the j-th distance unit;

[0084] Step 2.1.1.3: Through convolution operation Perform channel adjustment to obtain the residual term ,in, Indicates The residual term of the i-th pulse in the channel at the j-th distance unit;

[0085] Step 2.1.1.4: and After the residual connection, it is processed by the PReLU activation function to obtain the primary feature sequence .

[0086] Step 2.1.2: Follow the process in step 2.1.1 Perform feature extraction to obtain intermediate features ={ },in, Indicates The intermediate features of the channel at the i-th pulse at the j-th distance unit, express In this embodiment, =32;

[0087] Step 2.1.3: Follow the process in step 2.1.1 Perform feature extraction to obtain high-level features ={ },in, Indicates The high-level features of the ith pulse in the channel at the jth range unit, express In this embodiment, =64.

[0088] Step 2.2: Attention mechanism module , , After weighting the attention, the enhanced level features are obtained accordingly. , the enhanced intermediate features , enhanced high-level features The structure is as follows Figure 3 As shown;

[0089] Step 2.2.1: Use the attention mechanism module to After attention weighting, the enhanced primary features are obtained ={ | };in, represents the primary feature of the i-th pulse in the k-th channel after enhancement at the j-th distance unit;

[0090] Step 2.2.1.1. After performing average pooling operation and changing its shape, we get the query feature sequence ={ } and key feature sequences ,in, represents the query feature of the i-th pulse in the j-th distance unit, Indicates Key features of the ith pulse in the channel.

[0091] Step 2.2.1.2: Generate the attention weight matrix using formula (1) ={ },in, represents the attention weight matrix of the jth distance unit of the i-th pulse of the k-th channel;

[0092] (1)

[0093] In formula (1), represents the activation function; T represents transposition,

[0094] Step 2.2.1.3: Apply weight matrix After that, the enhanced primary features are obtained. ={ | }.

[0095] Step 2.2.2: Follow the process in step 2.2.1 to After attention weighting, the enhanced intermediate features are obtained ={ | },in, Indicates The mid-level features of the i-th pulse in the channel after enhancement at the j-th distance unit;

[0096] Step 2.2.3: Follow the process in step 2.2.1 to After attention weighting, enhanced high-level features are obtained ={ | };in, Indicates The enhanced high-level features of the i-th pulse in the channel at the j-th range unit.

[0097] Step 2.3: Construct a feature aggregation module including local feature submodule and global feature submodule, and use it to , , Perform feature aggregation and fusion processing to obtain a joint feature sequence The structure is as follows Figure 4 As shown;

[0098] Step 2.3.1: Feature aggregation module Processing to obtain global primary features ={ };in, Indicates The global primary features of the ith pulse in the channel at the jth distance unit;

[0099] Step 2.3.1.1: Local feature submodule pair After multiple convolution operations and compressing its pulse dimension, local primary features are obtained ={ },in, Indicates Channel No. The local primary characteristics of a pulse in the jth distance unit, Indicates the number of pulses after compression. In this embodiment, ;

[0100] Step 2.3.1.2: Global feature submodule pair Transpose, and then perform multiple convolution operations, and then transpose again to restore the original shape to obtain the global primary features ={ }.

[0101] Step 2.3.2: Follow the process in 2.3.1 After processing, the global intermediate features are obtained ={ },in, For the The global intermediate features of the i-th pulse in the channel at the j-th distance unit;

[0102] Step 2.3.3: Follow the process in 2.3.1 After processing, we get the global high-level features ={ },in, For the The global high-level features of the i-th pulse in the channel at the j-th distance unit;

[0103] Step 2.3.4: , , After concatenation in the channel dimension, the joint feature sequence is obtained ={ | },in, Indicates The joint feature of the ith pulse in the channel at the jth distance unit, C represents the number of channels in the joint feature sequence. In this embodiment, C=112.

[0104] Step 2.4: Use the target detection module to Processing is performed to obtain a binary probability matrix sequence containing the prediction results The structure is as follows Figure 5 As shown;

[0105] Step 2.4.1: Target detection module After multiple convolution operations, the target prediction sequence is obtained ,in, Indicates The target prediction value of the i-th pulse in the channel at the j-th distance unit;

[0106] Step 2.4.2: After passing the softmax activation function, we get a binary probability matrix sequence containing the prediction results. ={ },in, Indicates The predicted probability of the i-th pulse in the channel at the j-th distance unit; when k=1, represents the target prediction probability of the i-th pulse in the j-th distance unit, when k=2, represents the predicted probability of non-target of the i-th pulse at the j-th distance unit.

[0107] Step 3: Use the focal loss function to calculate the loss and train the network model;

[0108] Step 3.1: Use formula (2) to construct the standard cross entropy loss :

[0109] (2)

[0110] Step 3.2: Use formula (3) to construct the cross entropy loss :

[0111] (3)

[0112] Step 3.3: Use formula (4) to construct the total loss function :

[0113] (4)

[0114] In formula (4), is the category weight of the j-th distance unit, is the adjustment factor. In this embodiment, Set to 2, .

[0115] Step 4: Train the multi-feature fusion network using the gradient descent method and calculate To update the network parameters until Until convergence, the optimal target detection model is obtained, which is used to perform target detection on the input radar echo data.

[0116] In this embodiment, an electronic device includes a memory and a processor, characterized in that the memory is used to store a program that supports the processor to execute any radar small target detection method as described in claim 1, and the processor is configured to execute the program stored in the memory.

[0117] In this embodiment, a computer-readable storage medium stores a computer program, wherein the computer program executes the steps of any one of the radar small target detection methods in claim 1 when executed by a processor.

Claims

1. A multi-feature fusion sea surface small target detection method based on attention mechanism, characterized in that: The steps are as follows: Step 1: Preprocess the radar echo data sequence to obtain the joint feature sequence ;in, represents the joint feature of the ith pulse in the kth channel at the jth distance unit, N represents the number of pulses, and M represents the number of distance units; Step 2: Construct a multi-feature fusion network based on the attention mechanism, including: feature extraction module, attention mechanism module, feature aggregation module, and target detection module; Step 2.1: The feature extraction module Perform multi-level and multi-scale feature extraction to obtain primary features accordingly , intermediate features , advanced features ; Step 2.2: The attention mechanism module , , After attention weighting, the enhanced primary features are obtained accordingly. , the enhanced intermediate features , enhanced high-level features ; Step 2.3: Construct a feature aggregation module including local feature submodule and global feature submodule, and use it to , , Perform feature aggregation and fusion processing to obtain a joint feature sequence ; Step 2.4: Use the target detection module to Processing is performed to obtain a binary probability matrix sequence containing the prediction results ; Step 3: Construct the total loss function ; Step 4: Train the multi-feature fusion network by gradient descent method and calculate To update the network parameters until Until convergence, the optimal target detection model is obtained, which is used to perform target detection on the input radar echo data.

2. According to the multi-feature fusion sea surface small target detection method based on attention mechanism in claim 1, it is characterized in that: Step 1 is performed as follows: Step 1.1: Obtain radar echo data sequence and perform data correction, pulse compression, and normalization operations to obtain the processed radar echo data sequence. And its corresponding label sequence ,in, represents the radar echo signal value of the ith pulse in the jth distance unit, express target marker; Step 1.2: Align the distance unit with the Perform fast Fourier transform processing to obtain the amplitude sequence of the response distance information ,in, represents the amplitude of the ith pulse at the jth distance unit; Step 1.3: Align the pulse direction Perform Fourier transform to obtain the Doppler spectrum sequence ,in, represents the Doppler spectrum of the ith pulse in the jth range unit; Step 1.4: Perform centralization processing to obtain the processed Doppler image sequence ,in, represents the Doppler map of the ith pulse at the jth range unit; Step 1.5: and Join on the channel dimension to form a joint feature sequence .

3. According to the multi-feature fusion sea surface small target detection method based on attention mechanism in claim 2, it is characterized in that: Step 2.1 is performed as follows: Step 2.1.1: Use feature extraction module to Perform feature extraction to obtain primary features ={ };in, Indicates The primary characteristics of the ith pulse in the channel at the jth distance unit; Step 2.1.1.

1. After convolution operation, batch normalization and PReLU activation function processing, the first feature sequence is obtained. ={ },in, Indicates The first feature of the i-th pulse in the channel at the j-th distance unit, express The number of channels; Step 2.1.1.2: After performing convolution operation and batch normalization, the second feature sequence is obtained. ,in, Indicates The second characteristic of the i-th pulse in the channel at the j-th distance unit; Step 2.1.1.3: Through convolution operation Perform channel adjustment to obtain the residual term sequence ,in, Indicates The residual term of the i-th pulse in the channel at the j-th distance unit; Step 2.1.1.4: and After the residual connection, it is processed by the PReLU activation function to obtain the primary features ; Step 2.1.2: Follow the process in step 2.1.1 Perform feature extraction to obtain intermediate features ={ },in, Indicates The intermediate features of the channel at the i-th pulse at the j-th distance unit, express The number of channels; Step 2.1.3: Follow the process in step 2.1.1 Perform feature extraction to obtain high-level features ={ },in, Indicates The high-level features of the ith pulse in the channel at the jth range unit, express The number of channels.

4. According to claim 3, a multi-feature fusion sea surface small target detection method based on attention mechanism is characterized in that: Step 2.2 is performed as follows: Step 2.2.1: Use the attention mechanism module to After attention weighting, the enhanced primary features are obtained ={ | };in, represents the primary feature of the i-th pulse in the k-th channel after enhancement at the j-th distance unit; Step 2.2.1.

1. After performing average pooling operation and changing its shape, we get the query feature sequence ={ } and key feature sequences ,in, represents the query feature of the i-th pulse in the j-th distance unit, Indicates The key features of the ith pulse in the channel; Step 2.2.1.2: Generate the attention weight matrix using formula (1) ={ },in, represents the attention weight matrix of the jth distance unit of the i-th pulse of the k-th channel; (1) In formula (1), represents the activation function; T represents transposition, Step 2.2.1.3: Apply weight matrix After that, the enhanced primary features are obtained. ={ | }; Step 2.2.2: Follow the process in step 2.2.1 to After attention weighting, the enhanced intermediate features are obtained ={ | },in, Indicates The mid-level features of the i-th pulse in the channel after enhancement at the j-th distance unit; Step 2.2.3: Follow the process in step 2.2.1 to After attention weighting, enhanced high-level features are obtained ={ | };in, Indicates The enhanced high-level features of the i-th pulse in the channel at the j-th range unit.

5. According to the multi-feature fusion sea surface small target detection method based on attention mechanism in claim 4, it is characterized in that: Step 2.3 is performed as follows: Step 2.3.1: Feature aggregation module Processing to obtain global primary features ={ };in, Indicates The global primary features of the ith pulse in the channel at the jth distance unit; Step 2.3.1.1: Local feature submodule pair After multiple convolution operations and compressing its pulse dimension, local primary features are obtained ={ },in, Indicates Channel No. The local primary characteristics of a pulse in the jth distance unit, Indicates the number of pulses after compression; Step 2.3.1.2: Global feature submodule pair Transpose, and then perform multiple convolution operations, and then transpose again to restore the original shape to obtain the global primary features ={ }; Step 2.3.2: Follow the process in 2.3.1 After processing, the global intermediate features are obtained ={ },in, For the The global intermediate features of the i-th pulse in the channel at the j-th distance unit; Step 2.3.3: Follow the process in 2.3.1 After processing, we get the global high-level features ={ },in, For the The global high-level features of the i-th pulse in the channel at the j-th distance unit; Step 2.3.4: , , After concatenation in the channel dimension, the joint feature sequence is obtained ={ | },in, Indicates The joint feature of the i-th pulse in the channel at the j-th distance unit, C represents the number of channels in the joint feature sequence.

6. According to claim 5, a multi-feature fusion sea surface small target detection method based on attention mechanism is characterized in that: Step 2.4 is performed as follows: Step 2.4.1: Target detection module After multiple convolution operations, the target prediction sequence is obtained ,in, Indicates The target prediction value of the i-th pulse in the channel at the j-th distance unit; Step 2.4.2: After passing the softmax activation function, we get a binary probability matrix sequence containing the prediction results. ={ },in, Indicates The predicted probability of the i-th pulse in the channel at the j-th distance unit; when k=1, represents the target prediction probability of the i-th pulse in the j-th distance unit, when k=2, represents the predicted probability of non-target of the i-th pulse at the j-th distance unit.

7. The method for detecting small sea targets by multi-feature fusion based on attention mechanism according to claim 6 is characterized in that: Step 3 is performed as follows: Step 3.1: Use formula (2) to construct the standard cross entropy loss : (2) Step 3.2: Use formula (3) to construct the cross entropy loss : (3) Step 3.3: Use formula (4) to construct the total loss function : (4) In formula (4), is the category weight of the j-th distance unit, is the adjustment factor.

8. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the radar small target detection method according to any one of claims 1 to 7, and the processor is configured to execute the program stored in the memory.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the radar small target detection method according to any one of claims 1 to 7 are executed.

Images