A sea clutter suppression method based on a scalable codec enhancement network
By directly performing sea clutter suppression on the original range term of the radar echo through the CEDAN network, the problem that existing technologies cannot directly suppress and adapt to time requirements is solved. This achieves effective suppression of strong sea clutter and flexible tailoring of the network structure, improving the effect and efficiency of clutter suppression.
Patent Information
- Application Number
- CN202310064391.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-01-12
- Publication Date
- 2025-09-16
- Estimated Expiration
- 2043-01-12
AI Technical Summary
Existing technologies cannot directly suppress sea clutter based on the original distance term of radar echoes, and the fixed network structure makes it impossible to adapt to different time requirements.
Based on the pruning codec enhancement network (CEDAN), the network is trained by constructing a loss function and a back-propagation algorithm, and clutter suppression is performed directly on the original range term of the radar echo. The feature extraction capability of the network is enhanced by ResUblock, and the network structure is tailored to adapt to different time requirements.
It achieves effective suppression of strong sea clutter, improves the effect and consistency of clutter suppression, reduces network operation time, and maintains efficient clutter suppression performance.
Smart Images
Figure CN116243265B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of radar clutter suppression, and in particular relates to a sea clutter suppression technology. Background Art
[0002] Marine environmental monitoring holds significant research value in areas such as marine environmental awareness, marine ecological protection, and marine security and defense. Ocean surveillance radar is a microwave sensor system capable of all-weather sea surface observation. It dynamically monitors the sea surface and detects targets, and is widely used in both civilian and military ocean surveillance. When radar detects surface targets, particularly those with small radar cross-sections such as fishing vessels and small boats, the presence of sea clutter can dramatically reduce the signal-to-clutter ratio (SCR) of the radar echo, lowering the probability of target detection. Consequently, sea clutter can severely impact and degrade the detection performance of ocean surveillance radars. Therefore, research on sea clutter suppression is essential to improve target detection performance.
[0003] Many domestic universities and research institutions have conducted research on radar clutter suppression technology based on deep learning and have achieved some results. The paper "Sea clutter suppression for radar PPI images based on SCS-GAN" (IEEE Geosci. Remote Sens. Lett., vol. 18, no. 11, pp. 1886–1890, 2020) designed a residual attention generator and a sea clutter discriminator based on a generative adversarial network (GAN). A GAN for sea clutter suppression was then constructed for clutter suppression in radar planar position indicator images. The paper "A sea clutter suppression method based on a machine learning approach for marine surveillance radar" (IEEE J. Sel. Top. Appl. Earth Observations Remote Sens, vol. 15, pp. 3120–3130, 2022) designed a recurrent GAN consisting of two GANs to suppress sea clutter and proposed a target consistency loss to preserve target information.
[0004] These methods alleviate the problem of traditional clutter suppression methods, which require pre-modeling and estimation of sea clutter parameters, resulting in insufficient robustness in sea clutter suppression. However, they all perform clutter suppression based on radar images and cannot directly suppress the original range term. Furthermore, due to their fixed network structure, it is impossible to change the network structure to meet the time requirements of different tasks when performing clutter suppression. Therefore, the research of a clutter suppression network that can meet different time requirements and directly act on the original range term has important practical significance and application value. Summary of the Invention
[0005] To solve the above technical problems, the present invention proposes a sea clutter suppression method based on a tailorable codec enhancement network, which can directly perform clutter suppression on the original range term of the radar echo.
[0006] The technical solution adopted by the present invention is: a sea clutter suppression method based on a tailorable codec enhancement network, comprising:
[0007] S1. Pulse compress the received radar echo signal and construct a training set, a validation set, and a test set based on the pulse-compressed radar echo signal and the corresponding label;
[0008] S2. Build a tailorable codec enhancement network;
[0009] S3. Construct loss function;
[0010] S4, using the training set constructed in step S1 and the loss function constructed in step S3 to train the clippable codec enhancement network constructed in step S2 through a back propagation algorithm; and using the validation set constructed in step S1 to update the hyperparameters of the clippable codec enhancement network constructed in step S2;
[0011] S5. Input the pulse-compressed radar echo signal in the test set constructed in step S1 into the tailorable codec enhancement network trained in step S4. The output result is the final sea clutter suppression result.
[0012] The beneficial effects of the present invention include: The CEDAN-based clutter suppression method of the present invention can effectively suppress radar echoes with strong sea clutter; the encoder and decoder constructed using ResUblock can enhance the network's ability to extract contextual information; the enhancement module improves the clutter suppression effect by generating intermediate suppression results and fusing them; the result consistency loss of the present invention can improve the suppression performance of all intermediate suppression results, making CEDAN customizable. The method of the present invention can directly perform clutter suppression on the original distance term of the radar echo, effectively suppressing strong sea clutter, and can also tailor the CEDNA network structure to different time requirements. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Figure 1 The overall framework diagram of the invention method;
[0014] Figure 2 This is the network structure diagram of ResUblock;
[0015] in, Figure 2 (a) is the overall structure diagram of ResUblock, Figure 2 (b) and Figure 2 (c) is the network structure diagram of two different implementations of ResUblock;
[0016] Figure 3 The network structure diagram of the feature mapping layer and the inhibition layer;
[0017] Figure 4 The original radar echo and simulation data;
[0018] Figure 5 The output of CEDAN and each intermediate result;
[0019] in, Figure 5 (a) is the output and intermediate results of CEDAN using the result consistency loss function. Figure 5 (b) Output and intermediate results of CEDAN without using the result consistency loss function;
[0020] Figure 6 The output results of different methods under different signal-to-noise ratios of the input data;
[0021] Figure 7 The signal-to-noise ratio and signal-to-noise ratio gain of the output results of different methods change with the signal-to-noise ratio of the input data;
[0022] in, Figure 7 (a) shows the change of the signal-to-noise ratio of the output results of different methods with the signal-to-noise ratio of the input data. Figure 7(b) The signal-to-noise ratio gain of different methods changes with the signal-to-noise ratio of the input data;
[0023] Figure 8 Schematic diagram of the scene for measured data;
[0024] in, Figure 8 (a) with Figure 8 (b) The original echoes obtained in different scenarios;
[0025] Figure 9 Training samples and labels for measured data;
[0026] in, Figure 9 (a) is the training sample, Figure 9 (b) is the label of the training sample;
[0027] Figure 10 Output results of different measured data under different methods;
[0028] in, Figure 10 (a) is the output result of the measured data of target 1 under different methods. Figure 10 (b) is the output results of the measured data of target 2 under different methods. DETAILED DESCRIPTION
[0029] To facilitate those skilled in the art to understand the technical content of the present invention, the present invention is further explained below with reference to the accompanying drawings.
[0030] The overall structure of the method of the present invention is shown in FIG. Figure 1 The specific steps are as follows:
[0031] Step 1: Build ResUblock
[0032] ResUblock consists of feature extraction layer, U-block and feature fusion layer, such as Figure 2 (a) shows the feature extraction layer used to extract the H×W×C in Input data characteristics Figure X The local feature Y, represents the field of real numbers, It consists of a (C in ,3×3,C out ) convolutional layer, batch normalization layer and ReLU activation function; H×W×C in Indicates that the length, width and height of the input feature are H, W and C respectively in ;(C in ,3×3,C out ) indicates that the number of input channels of the convolutional layer is C in , the number of output channels is Cout , the convolution kernel size is 3×3. U-block is composed of adding convolution layer, downsampling operation, upsampling operation and splicing operation to the U-shaped structure with height L, which is used to extract the multi-scale feature Z of Y, where like Figure 2 (c) Figure 2 The (M, 3×3, d = 2, M) in the figure indicates that the convolution used is a dilated convolution and the number of input channels is C. in , the number of output channels is C out , the convolution kernel size is 3×3, and the hole rate is 2. The feature fusion layer fuses Y and Z through residual connection operation to obtain rich context information O. So the fusion operation can be expressed as:
[0033] O=Y+Z
[0034] ResUblock extracts contextual information from input feature maps without changing their size. Therefore, it is a plug-and-play module that can be used to build any part of a network to enhance its feature extraction capabilities. Furthermore, since most of its operations are performed on the downsampled feature maps, the computational overhead is minimal. Figure 2 (b) with Figure 2 (c) are two different implementations of ResUblock. The difference between them is that, Figure 2 (b) All ordinary convolutions in (c) are replaced by dilated convolutions, and upsampling and downsampling operations are no longer used. This can avoid the loss of useful contextual information caused by direct upsampling and downsampling.
[0035] Step 2: Construct CDEAN (Cuttable Encoder-Decoder-Augmentation Network) for clutter suppression
[0036] CDEAN consists of an encoder, a decoder, and a contrast information enhancement module. Figure 1 The encoder is composed of six ResUblocks of different heights and a downsampling operation, where the downsampling operation is used to reduce the size of the feature map. The specific number of ResUblocks in the encoder of this embodiment is determined by the size of the input data:
[0037]
[0038] Where n represents the number of ResUblocks in the encoder; ResUblock is used to extract the contextual information of the feature map; the encoder can obtain rich multi-scale contextual information about the input radar echo by alternating ResUblock and downsampling operations.
[0039] Since the decoder needs to fuse the features extracted by each ResUblock of the encoder, the decoder consists of five ResUblocks of different heights, upsampling operations, and splicing operations. The upsampling operation is used to increase the size of the feature map and gradually increase it to the same size as the input. The splicing operation is used to fuse the output of the previous ResUblock and the symmetrical ResUblock in the encoder and input it into the next ResUblock to fuse the context information of different scales extracted by the encoder and extract deeper context information at the same time. Fusion through this process can reduce the loss of detail in the feature map caused by multiple direct upsampling operations; the decoder can fuse the features extracted by decoding through the cooperation of these modules, and further extract multi-scale features containing multi-context information.
[0040] The enhancement module consists of a suppression layer, a splicing operation, and multiple feature mapping layers. It first converts the features of each scale extracted by the decoder into intermediate suppression results through the feature mapping layer, then splices all the intermediate suppression results together through the splicing operation and sends them to the suppression layer. Finally, the suppression layer is used to obtain the final sea clutter suppression result (output). The feature mapping layer consists of a convolution layer with a convolution kernel size of 3×3, a sigmoid activation function, and an upsampling operation. The sampling operation is to upsample the output to the same size as the input data. The suppression layer consists of a convolution layer and a sigmoid activation function, as shown in the attached figure. Figure 3 shown.
[0041] The specific structural parameters of the network are shown in Table 1:
[0042] Table 1 Structural parameters of CEDAN
[0043]
[0044] Step 3: Construct the loss function
[0045] The loss function consists of two parts: the main loss function and the auxiliary loss function. The main loss function is the cross entropy loss l between the output and the label R , the calculation formula is:
[0046]
[0047] Among them, (h,w) is the coordinate index of the two-dimensional data, h is the length index, w is the width index, R is the output of CEDAN, L is the label, H and W are the length and width of L respectively.
[0048] The auxiliary loss function consists of the result consistency loss. It increases the contrast information between the target and the clutter of each intermediate suppression result by calculating the loss between all intermediate suppression results and the label, thereby improving the clutter suppression performance of each intermediate suppression result, so that the intermediate suppression results can also suppress sea clutter. In this way, the network can trim some high-level ResUblocks of the decoder, the suppression layers of the enhancement module, and some feature mapping layers to reduce the running time without significantly reducing the suppression performance of the network. The calculation method of the result consistency loss can be expressed as:
[0049]
[0050]
[0051] Among them, S i is the i-th intermediate suppression result, It's S i The loss between the labels, yes The weight of .
[0052] So the loss function can be expressed as:
[0053]
[0054] Among them, w R and w aux They are l R and l aux The weight of .
[0055] Step 4: Preprocess the radar echo signal
[0056] Consider an ocean surveillance radar that continuously transmits M linear frequency modulation signal pulses to detect a target moving toward the radar at a speed v at position R0. The echo signal received by the radar is:
[0057]
[0058] Among them, σ q is the scattering coefficient of the target in the mth pulse, f c is the carrier frequency, T is the pulse repetition interval, τ0 = 2R0 / c represents the round-trip time delay of the radar signal, c is the speed of light, and k is the frequency modulation slope of the radar.
[0059] The received signal r(t) is subjected to pulse compression processing.
[0060] The following uses semi-measured data and full-measured data as examples to illustrate the clutter suppression effect of the method of the present invention:
[0061] Step 5: Create a semi-measured dataset
[0062] The data set is constructed using a semi-measured method, that is, several simulated target signals are generated and pulse compressed, and then the simulated target signals are superimposed on the measured clutter data to construct data samples, where the size of each data sample is 400×400, such as Figure 4 As shown in the figure, the target signal superimposed on the measured clutter data is the label of the corresponding data, that is, the label is the target signal without clutter. The measured clutter used is 138 frames of data from the 2020 issue 2 of the Journal of Radars (20200722150408_798_scanning.mat-20200722150931_935_scanning.mat). The data in the dataset is divided into training, validation, and test sets in a ratio of 6:2:2.
[0063] Step 6: Model training on semi-real data sets
[0064] The training set constructed in step 5 and the loss function constructed in step 3 are used to train the CDEAN constructed in step 2 through the back propagation algorithm, and the hyperparameters of CDEAN are updated through the validation set constructed in step 5.
[0065] Step 7: Test the model on a semi-real dataset
[0066] The model trained in step 6 is tested on the test set data constructed in step 5. First, the effect of the result consistency loss on the clutter suppression performance of CEDAN is as follows: Figure 5 As shown in the figure, the model that does not use the result consistency loss as an auxiliary function for training is recorded as CEDAN w / o loss. It is obvious that the result consistency loss can improve the suppression performance of the intermediate results of CEDAN, so that the intermediate results can be directly used as the final output, thereby cutting out some redundant layers to reduce the inference time of the model. Then compare the clutter suppression performance of different models, as shown in the figure. Figure 6 and Figure 7 As shown, Figure 7 The σ is the signal-to-noise ratio improvement factor, which is defined as:
[0067] σ=SCR out -SCR in ,
[0068] Among them, SCR out SCR is the signal-to-noise ratio of the output data. inis the signal-to-noise ratio of the output data. It can be seen that CEDAN has a stronger clutter suppression capability than UNet and CV-UNet methods, and the clutter consistency loss improves CEDAN's performance. The number of parameters, number of multiplications and additions, and inference time of different models are shown in Table 2:
[0069] Table 2 Parameters, multiplication and addition times, and inference time of different models
[0070] Model CEDAN CEDAN S2 UNet CV-UNet Parameter quantity <![CDATA[44.01×10 6 ]]> <![CDATA[43.86×10 6 ]]> <![CDATA[31.04×10 6 ]]> <![CDATA[62.06×10 6 ]]> Number of multiplications and additions <![CDATA[91.89×10 9 ]]> <![CDATA[75.09×10 9 ]]> <![CDATA[133.45×10 9 ]]> <![CDATA[532.89×10 9 <!-- 5 -->]]> Inference time 19.50ms 14.83ms 16.59ms 63.26ms
[0071] CEDAN S2 uses the intermediate result S2 of CEDAN as the final output, pruning DE1, Fm1, Fm3, Fm4, Fm5, Fm6, and the suppression layer. The number of multiplications and additions can be expressed as the computational complexity of the model. Table 2 shows that CEDAN outperforms UNet and CV-UNet in terms of computational complexity. CEDAN S2, obtained after pruning, outperforms UNet and CV-UNet in terms of parameter count, computational complexity, and inference time.
[0072] Step 8: Create a full measurement dataset
[0073] The dataset is constructed using a fully measured method, that is, the data with target samples are directly selected from the radar echo and the corresponding labels are made. The labels are specifically extracted from the measured dataset by manual methods without any mixed measured target signals. The data used is the data published in the first issue of the Journal of Radar in 2020, as shown in Table 3 and Figure 8 As shown in . The size of each data sample is 400×400, as shown in Figure 9 As shown in Figure 2, the dataset is divided into training set, validation set, and test set in a ratio of 6:2:2.
[0074] Table 3 Dataset information
[0075] Serial number Sea state level Description Data files 1 Level 3-4 There is a channel buoy at 4.84km 20210106150614_01_staring.mat 2 Level 3-4 There is a channel buoy at 4.84km 20210106150614_02_staring.mat 3 Level 3-4 There is a channel buoy at 4.84km 20210106150614_03_staring.mat 4 Level 3-4 There are two channel buoys at 2.81km and 4.16km 20210106160919_01_staring.mat 5 Level 3-4 There are two channel buoys at 2.81km and 4.16km 20210106160919_02_staring.mat
[0076] Step 9: Model training on the full measured dataset
[0077] The training set constructed in step 8 and the loss function constructed in step 3 are used to train the CDEAN constructed in step 2 through the back propagation algorithm, and the hyperparameters of CDEAN are updated through the validation set constructed in step 5.
[0078] Step 10: Test the model on the full measured dataset
[0079] Use the model trained in step 9 to test the test data constructed in step 8. The test results are as follows Figure 10As shown in Figure 2, it can be seen that the suppression effect of CEDAN is better than that of UNet and CV-UNet; the result consistency loss can improve the suppression performance of CENAN and make the network scalable.
[0080] In summary, the method of the present invention has a good clutter suppression effect on both full-test and semi-test data.
[0081] Those skilled in the art will appreciate that the embodiments described herein are intended to aid the reader in understanding the principles of the present invention, and it should be understood that the scope of the present invention is not limited to such specific descriptions and embodiments. Various modifications and variations are readily apparent to those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims.
Claims
1. A sea clutter suppression method based on a scalable codec enhancement network, characterized in that: include: S1. Pulse compress the received radar echo signal and construct a training set, a validation set, and a test set based on the pulse-compressed radar echo signal and the corresponding label; S2. Build a tailorable codec enhancement network; The pruneable codec enhancement network includes: encoder, decoder and contrast information enhancement module; The encoder includes multiple ResUblocks of different heights and multiple downsampling operations; ResUblock is used to extract the context information of the input feature map without changing the size of the input feature map; downsampling operation is used to reduce the size of the feature map; The decoder is used to extract multi-scale features containing multi-context information. The decoder includes multiple ResUblocks of different heights, multiple upsampling operations, and multiple splicing operations. The upsampling operation is used to increase the size of the feature map. The splicing operation is used to fuse the output of the previous ResUblock in the decoder and the symmetrical ResUblock in the encoder and input it into the next ResUblock in the decoder to fuse the context information of different scales extracted by the encoder. ResUblock includes: feature extraction layer, U block and feature fusion layer; The feature extraction layer is used to extract the H×W×C in The local feature Y of the input data feature map X, the feature extraction layer specifically includes a (C in ,3×3,C out )’s convolutional layer, batch normalization layer, and ReLU activation function; represents the field of real numbers, H×W×C in Indicates that the length, width and height of the input feature are H, W and C respectively in ;(C in ,3×3,C out ) indicates that the number of input channels of the convolutional layer is C in , the number of output channels is C out , the convolution kernel size is 3×3; The U block is used to extract the multi-scale feature Z of Y, where The U block is specifically composed of adding convolutional layers, downsampling operations, upsampling operations, and splicing operations to a U-shaped structure with a height of L; The feature fusion layer fuses Y and Z through residual connection operations to obtain rich context information O. The enhancement module consists of a suppression layer, a splicing operation, and multiple feature mapping layers. The feature mapping layer converts the features of each scale extracted by the decoder into intermediate suppression results. The splicing operation then concatenates all the intermediate suppression results and feeds them into the suppression layer. Finally, the suppression layer obtains the final sea clutter suppression result. S3. Construct loss function; S4, using the training set constructed in step S1 and the loss function constructed in step S3 to train the clippable codec enhancement network constructed in step S2 through a back propagation algorithm; and using the validation set constructed in step S1 to update the hyperparameters of the clippable codec enhancement network constructed in step S2; S5. Input the pulse-compressed radar echo signal in the test set constructed in step S1 into the tailorable codec enhancement network trained in step S4. The output result is the final sea clutter suppression result.
2. The sea clutter suppression method based on a scalable codec enhancement network according to claim 1, characterized in that: The label is a radar echo signal without any impurities extracted from the received radar echo signal.
3. The sea clutter suppression method based on a tailorable codec enhancement network according to claim 2 is characterized in that There are two implementation methods of ResUblock, namely ResUblock-L and ResUblock-LD. The difference between ResUblock-LD and ResUblock-L is that the U block of ResUblock-LD is implemented by void convolution, and upsampling and downsampling operations are no longer used.
4. The sea clutter suppression method based on a tailorable codec enhancement network according to claim 3, characterized in that: The number of ResUblocks in the encoder is determined by the size of the input feature map of the codec enhancement network that can be cropped.
5. The sea clutter suppression method based on a scalable codec enhancement network according to claim 4, characterized in that: The loss function expression constructed in step S3 is: Among them, w R and w aux They are the main loss functions l R And the auxiliary loss function l aux The weight of .
6. The sea clutter suppression method based on a scalable codec enhancement network according to claim 5, characterized in that: Main loss function l R is the cross entropy loss l between the output and the label R , the calculation formula is: Among them, (h,w) is the coordinate index of the two-dimensional data, h is the length index, w is the width index, R is the output of the croppable codec enhancement network, L is the label, H and W are the length and width of L respectively.
7. The sea clutter suppression method based on a scalable codec enhancement network according to claim 6, characterized in that: The auxiliary loss function consists of the result consistency loss, which is calculated as: Among them, S i is the i-th intermediate suppression result, It's S i The loss between the labels, yes The weight of .