A multi-feature fusion sea surface small target detection method based on an attention mechanism

Abstract

The application discloses a kind of based on attention mechanism's multi-feature fusion sea surface small target detection method, comprising:1, radar echo data is preprocessed, and the joint feature sequence containing rich features is obtained;2, construct the multi-feature fusion network model based on attention mechanism, including: feature extraction module, attention mechanism module, feature aggregation module, target detection module, and the joint feature sequence is handled, and the binary probability matrix sequence is obtained;3, loss is calculated using focal loss function, the network model is trained, and the optimal target detection model is obtained.The application effectively improves the detection precision of radar weak small target by multi-scale feature extraction, attention mechanism and feature fusion, and is suitable for the target detection demand in complex background and low signal-to-noise ratio environment.

Description

TECHNICAL FIELD

[0001] The method belongs to the cross field of artificial intelligence and computer vision and radar signal processing, and belongs to a multi-disciplinary application technology. BACKGROUND

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

[0003] Traditional radar small target detection methods, such as constant false alarm rate (CFAR) detectors and matching filter-based detection techniques, although 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, integrating feature information from different sources to improve target saliency. However, simple feature fusion strategies can easily lead to information redundancy or feature conflict, affecting the model's discriminative ability and computational efficiency. Therefore, attention mechanisms are introduced into the feature fusion process, effectively highlighting key features and suppressing redundant information, significantly improving the model's target detection performance in complex backgrounds.

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

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

[0007] 2. Insufficient multi-dimensional feature expression: current feature fusion methods are mostly based on single-scale or single-modal features, failing to fully utilize the expression ability 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 preserved, leaving a large optimization space for the model's feature expression ability.

[0008] 3. Class imbalance problem: in sea surface small target detection, target signals usually account for a very small proportion, and the class imbalance problem is significant. The processing ability of existing methods for difficult-to-classify samples is weak, and false alarm rate or missed detection rate is high, reducing the reliability of practical applications. SUMMARY

[0009] The present application is to solve the above-mentioned deficiencies in the prior art, a multi-feature fusion sea surface small target detection method based on attention mechanism is proposed, in order to realize cross-level feature fusion, to enhance the target characteristics and suppress clutter, so as to effectively improve the detection precision of radar weak small target, suitable for target detection demand in complex background and low signal-to-noise ratio environment.

[0010] In order to achieve the above-mentioned application purposes, the following technical solutions are adopted:

[0011] The multi-feature fusion sea surface small target detection method based on attention mechanism has the characteristics that the following steps are performed:

[0012] Step 1, pre-processing operation is performed on the radar echo data sequence to obtain a joint feature sequence ; wherein, represents the joint feature of the i-th pulse in the k-th channel in the j-th distance unit, N represents the number of pulses, and M represents the number of distance units;

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

[0014] Step 2.1, the feature extraction module performs multi-level and multi-scale feature extraction on , and correspondingly obtains a primary feature , an intermediate feature , and a high-level feature ;

[0015] Step 2.2, the attention mechanism module respectively performs attention weighting on , , , and correspondingly obtains an enhanced primary feature , an enhanced intermediate feature , and an enhanced high-level feature ;

[0016] Step 2.3, the feature aggregation module includes a local feature submodule and a global feature submodule, and is used for performing feature aggregation and fusion processing on , , , to obtain a joint feature sequence ;

[0017] Step 2.4, the target detection module is used for processing , to obtain a binary probability matrix sequence containing a prediction result ;

[0018] Step 3, constructing total loss function ;

[0019] Step 4, training the multi-feature fusion network by gradient descent method, and calculating to update network parameters until converges, thereby obtaining an optimal target detection model for input radar echo data target detection.

[0020] The multi-feature fusion sea surface small target detection method based on attention mechanism has the characteristics that step 1 is performed as follows:

[0021] Step 1.1, obtaining a radar echo data sequence and performing data correction, pulse compression, and normalization operation to obtain a processed radar echo data sequence and its corresponding label sequence , wherein, represents the radar echo signal value of the i-th pulse at the j-th range cell, represents target label;

[0022] Step 1.2, performing fast Fourier transform processing on in the direction of the range cell to obtain an amplitude sequence reflecting the distance information , wherein, represents the amplitude of the i-th pulse at the j-th range cell;

[0023] Step 1.3, performing Fourier transform on in the direction of the pulse to obtain a Doppler spectrum sequence , wherein, represents the Doppler spectrum of the i-th pulse at the j-th range cell;

[0024] Step 1.4, performing centering processing on to obtain a processed Doppler plot sequence , wherein, represents the Doppler plot of the i-th pulse at the j-th range cell;

[0025] Step 1.5, coupling and in the channel dimension to form a joint feature sequence .

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

[0027] Step 2.1.1, using a feature extraction module to perform feature extraction to obtain the primary feature ={ }; wherein, denotes the primary feature of the i-th pulse in the j-th distance unit in the k-th channel; denotes the primary feature of the i-th pulse in the j-th distance unit in the k-th channel;

[0028] Step 2.1.1.1, performing convolution operation on , and then processing through batch normalization and PReLU activation function to obtain the first feature sequence ={ }; wherein, denotes the first feature of the i-th pulse in the j-th distance unit in the k-th channel, denotes the number of channels of ;

[0029] Step 2.1.1.2, performing convolution operation on , and then processing through batch normalization to obtain the second feature sequence , wherein, denotes the second feature of the i-th pulse in the j-th distance unit in the k-th channel;

[0030] Step 2.1.1.3, performing channel adjustment on through convolution operation to obtain the residual term sequence , wherein, denotes the residual term of the i-th pulse in the j-th distance unit in the k-th channel;

[0031] Step 2.1.1.4, performing residual connection on and , and then processing through PReLU activation function to obtain the primary feature ;

[0032] Step 2.1.2, performing feature extraction on according to the process of Step 2.1.1 to obtain the intermediate feature ={ }; wherein, denotes the intermediate feature of the i-th pulse in the j-th distance unit in the k-th channel, denotes the number of channels of ;

[0033] Step 2.1.3, performing feature extraction on according to the process of Step 2.1.1 to obtain the high-level feature ={ }; wherein, ​​​​the i-th pulse in the k-th channel in the j-th distance unit, the i-th pulse in the k-th channel in the j-th distance unit, the i-th pulse in the k-th channel in the j-th distance unit, the number of channels.

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

[0035] Step 2.2.1, after attention weighting of the primary feature using an attention mechanism module, the enhanced primary feature is obtained | }; wherein, the i-th pulse in the k-th channel in the j-th distance unit,

[0036] Step 2.2.1.1, after average pooling operation on and changing its shape, the query feature sequence is obtained and the key feature sequence , wherein, the i-th pulse in the j-th distance unit, the i-th pulse in the k-th channel in the j-th distance unit, the key feature of the i-th pulse in the k-th channel.

[0037] Step 2.2.1.2, the attention weight matrix is generated using formula (1) , wherein, the i-th pulse in the k-th channel in the j-th distance unit,

[0038] (1)

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

[0040] Step 2.2.1.3, after applying the weight matrix to , the enhanced primary feature is obtained | ;

[0041] Step 2.2.2, after attention weighting of according to the process of step 2.2.1, the enhanced secondary feature is obtained | , wherein, the i-th pulse in the k-th channel in the j-th distance unit, The enhanced intermediate features of the i-th pulse in the channel 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 The enhanced high-level features of the i-th pulse in the channel at the j-th range unit.

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

[0044] Step 2.3.1, feature aggregation module Processing to obtain global primary features ={ };in, Indicates the The global primary characteristics of the i-th pulse in the channel at the j-th 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 the Channel No. The local primary characteristics of the pulse at 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 i-th pulse in the j-th distance unit in the k-th channel Global high-level features of the i-th pulse in the j-th distance unit in the k-th channel

[0049] Step 2.3.4, the target detection module performs multiple convolution operations on , , After splicing in the channel dimension, the joint feature sequence ={ | } is obtained, wherein represents the joint feature of the i-th pulse in the j-th distance unit in the k-th channel, and C represents the number of channels of the joint feature sequence.

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

[0051] Step 2.4.1, the target detection module performs multiple convolution operations on After splicing in the channel dimension, the joint feature sequence is obtained, wherein represents the joint feature of the i-th pulse in the j-th distance unit in the k-th channel, and C represents the number of channels of the joint feature sequence.

[0052] Step 2.4.2, the target detection module performs multiple convolution operations on After splicing in the channel dimension, the joint feature sequence ={ } is obtained, wherein represents the joint feature of the i-th pulse in the j-th distance unit in the k-th channel, and C represents the number of channels of the joint feature sequence. represents the prediction probability of the i-th pulse in 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 non-target prediction probability of the i-th pulse in the j-th distance unit.

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

[0054] Step 3.1, a standard cross-entropy loss is constructed using formula (2) :

[0055] (2)

[0056] Step 3.2, a cross-entropy loss is constructed using formula (3) :

[0057] (3)

[0058] Step 3.3, a total loss function is constructed using formula (4) :​​

[0059] (4)

[0060] In formula (4), is the class weight of the jth distance unit, is the adjustment factor.

[0061] The electronic device comprises a memory and a processor, and the memory is used to store a program supporting the processor to execute the sea surface small target detection method, and the processor is configured to execute the program stored in the memory.

[0062] The computer readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the sea surface small target detection method are executed.

[0063] Compared with the prior art, the present application has the following advantages:

[0064] 1. The present application can enhance the target features by multi-level and multi-scale feature extraction and attention mechanism, highlight the key features and suppress the clutter information, significantly improve the detection probability of small targets, and is particularly suitable for complex background and low signal-to-noise ratio scenes.

[0065] 2. The present application can retain the multi-dimensional information expression of the target by aggregating local features and global features, and combining cross-level feature splicing, and optimize the feature expression ability of the model. Compared with the traditional single feature fusion method, the aggregation of multi-level features enables the model to consider information of different scales, thereby improving the accurate positioning and classification effect of the target.

[0066] 3. The present application introduces a focal loss function to give higher weight to difficult-to-classify samples for the class imbalance problem, effectively reduces the false alarm rate and the missed detection rate. Compared with the conventional loss function, the focal loss can better focus on the samples difficult to classify, especially in target detection, effectively reducing the false alarm and missed detection phenomenon, thereby significantly improving the detection accuracy and ensuring the reliability of the detection result. BRIEF DESCRIPTION OF DRAWINGS

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

[0068] Figure 2 is the structure diagram of the feature extraction module in the present application;

[0069] Figure 3 is the structure diagram of the attention mechanism module in the present application;

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

[0071] Figure 5 Flow chart of target detection module in the present application. DETAILED DESCRIPTION

[0072] In this embodiment, the multi-feature fusion sea surface small target detection method with attention mechanism is realized through a series of optimization designs, including radar data preprocessing, multi-scale feature extraction, introduction of attention mechanism and cross-level feature fusion, which realizes efficient detection of radar weak small targets in complex sea surface environment. Specifically, as shown in the figure, the method is performed according to the following steps: Figure 1

[0073] Step 1, pre-process the radar echo data sequence to obtain the joint feature sequence ; wherein, represents the joint feature of the i-th pulse in the k-th channel at the j-th range cell, N represents the number of pulses, and M represents the number of range cells;

[0074] Step 1.1, obtain the radar echo data sequence and perform data correction, pulse compression and normalization operation to obtain the processed radar echo data sequence and its corresponding label sequence , wherein, represents the radar echo signal value of the i-th pulse at the j-th range cell, represents the target label of . In this embodiment, the size of each radar echo data is .

[0075] Step 1.2, perform fast Fourier transform processing on along the direction of the range cell to obtain the amplitude sequence reflecting the distance information , wherein, represents the amplitude of the i-th pulse at the j-th range cell;

[0076] Step 1.3, perform Fourier transform on along the pulse direction to obtain the Doppler spectrum sequence , wherein, represents the Doppler spectrum of the i-th pulse at the j-th range cell;

[0077] Step 1.4, centering processing is performed on to obtain the processed Doppler map sequence , wherein, represents the Doppler map of the i-th pulse at the j-th range cell;

[0078] Step 1.5, the 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 the feature extraction module to Perform feature extraction to obtain primary features ={ };in, Indicates the The primary characteristics 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 The first characteristic 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, Perform convolution operation and then process it through batch normalization to obtain the second feature ,in Indicates the The second characteristic of the i-th pulse in the channel at the j-th range unit;

[0084] Step 2.1.1.3, through the convolution operation Perform channel adjustment to obtain the residual term ,in, Indicates the 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 to Perform feature extraction to obtain intermediate features ={ },in, Indicates the The intermediate features of the channel at the i-th pulse in the j-th distance unit, express In this embodiment, =32;

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

[0088] Step 2.2, attention mechanism module , , After performing attention weighting respectively, the enhanced level features are obtained accordingly , the enhanced mid-level 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, Perform average pooling operation and change its shape to obtain 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 i-th 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 The enhanced intermediate features of the i-th pulse in the channel 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 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 The global primary characteristics of the i-th pulse in the channel at the j-th 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 the Channel No. The local primary characteristics of the pulse at 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 splicing in the channel dimension, the joint feature sequence is obtained ={ | },in, Indicates the The joint signature of the i-th pulse in the channel at the j-th distance unit, C represents the number of channels in the joint signature sequence. In this embodiment, C=112.

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

[0105] Step 2.4.1, after the target detection module performs multiple convolution operations on , a target prediction sequence is obtained, wherein, represents the target prediction value of the i th pulse in the j th distance unit in the k th channel;

[0106] Step 2.4.2, after is subjected to a softmax activation function, a binary probability matrix sequence containing prediction results ={ } is obtained, wherein, represents the prediction probability of the i th pulse in the j th distance unit in the k th channel; when k = 1, represents the target prediction probability of the i th pulse in the j th distance unit, when k = 2, represents the non-target prediction probability of the i th pulse in the j th distance unit.

[0107] Step 3, the loss is calculated by using a focal loss function, and the network model is trained;

[0108] Step 3.1, a standard cross-entropy loss is constructed by using formula (2) :

[0109] (2)

[0110] Step 3.2, a cross-entropy loss is constructed by using formula (3) :

[0111] (3)

[0112] Step 3.3, a total loss function is constructed by using formula (4) :

[0113] (4)

[0114] In formula (4), is the class weight of the j th distance unit, is an adjustment factor. In the embodiment, is set to 2, .

[0115] Step 4, the multi-feature fusion network is trained by using a gradient descent method, and is calculated to update the network parameters until​​ Converge until the optimal target detection model is obtained for 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 configured to store a program supporting the processor to execute the sea surface small target detection method, and the processor is configured to execute the program stored in the memory.

[0117] In this embodiment, a computer readable storage medium has a computer program stored thereon, characterized in that the computer program, when executed by a processor, performs the steps of the sea surface small target detection method.

Claims

1. A multi-feature fusion sea surface small target detection method based on attention mechanism, characterized by: 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 i-th pulse in the k-th channel at the j-th range unit, N represents the number of pulses, and M represents the number of range 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 mid-level features , enhanced high-level features ; 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, Perform average pooling operation and change its shape to obtain 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 The key features of the i-th 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 The enhanced intermediate features of the i-th pulse in the channel 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 The enhanced high-level features of the i-th pulse in the channel at the j-th range unit; 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 2.3.1, feature aggregation module Processing to obtain global primary features ={ };in, Indicates the The global primary characteristics of the i-th pulse in the channel at the j-th 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 the Channel No. The local primary characteristics of the pulse at 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 splicing in the channel dimension, the joint feature sequence is obtained ={ | },in, Indicates the 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; Step 2.4: Use the target detection module to Processing is performed to obtain a binary probability matrix sequence containing the prediction results ; Step 2.4.1, target detection module After multiple convolution operations, the target prediction sequence is obtained ,in, Indicates the The target prediction value of the i-th pulse in the channel at the j-th range 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 The predicted probability of the i-th pulse in the channel at the j-th distance unit; when k=1, Indicates 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 in the j-th distance unit; Step 3: Construct the total loss function ; Step 3.1: Use formula (2) to construct the standard cross entropy loss : (2) Step 3.2: Use formula (3) to construct 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 regulating factor; 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.

2. The method for detecting small sea targets using multi-feature fusion based on attention mechanism according to claim 1 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 tag sequence ,in, represents the radar echo signal value of the i-th pulse in the j-th range 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 i-th pulse at the j-th 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 i-th pulse at the j-th range unit; Step 1.4, Perform centralization processing to obtain the processed Doppler image sequence ,in, represents the Doppler map of the i-th pulse at the j-th range unit; Step 1.5, and Join on the channel dimension to form a joint feature sequence .

3. The method for detecting small sea targets based on multi-feature fusion of attention mechanism according to claim 2 is characterized in that: Step 2.1 is performed as follows: Step 2.1.1, use the feature extraction module to Perform feature extraction to obtain primary features ={ };in, Indicates the The primary characteristics of the i-th pulse in the channel at the j-th distance unit; Step 2.1.1.1, After performing convolution operation and then processing with batch normalization and PReLU activation function, the first feature sequence is obtained. ={ },in, Indicates the The first characteristic of the i-th pulse in the channel at the j-th distance unit, express The number of channels; Step 2.1.1.2, Perform convolution operation and then batch normalization to obtain the second feature sequence ,in, Indicates the The second characteristic of the i-th pulse in the channel at the j-th range unit; Step 2.1.1.3, through the convolution operation Perform channel adjustment to obtain the residual term sequence ,in, Indicates the 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 and the primary features are obtained. ; Step 2.1.2: Follow the process in step 2.1.1 to Perform feature extraction to obtain intermediate features ={ },in, Indicates the The intermediate features of the channel at the i-th pulse in the j-th distance unit, express The number of channels; Step 2.1.3: Follow the process of step 2.1.1 to Perform feature extraction to obtain high-level features ={ },in, Indicates the The high-level features of the i-th pulse in the channel at the j-th range unit, express The number of channels.

4. 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 sea surface small target detection method according to any one of claims 1 to 3, and the processor is configured to execute the program stored in the memory.

5. 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 method for detecting small sea targets in any one of claims 1 to 3 are executed.

Images