A Deep Network-Based Radar Target Detection Method Based on Multi-Branch Combination

Abstract

This invention relates to a deep network-based radar target detection method based on multi-branch integration. The steps include: 1) normalizing the radar echo signal; 2) slicing the data to obtain 10 radar data images; 3) taking one radar data image, copying it twice to increase the number of channels to 3, and then normalizing it again; 4) resizing, performing a flatten operation and decoding branching; then performing Layer Normalization, resizing, and deconvolution operations; 5) after resizing, feeding it into a pyramid convolutional structure for feature extraction; 6) concatenating the data from steps 4 and 5, and convolving to obtain response map data; 7) processing the response map data; 8) repeating steps 3 to 7 to obtain 10 data images from the 10 radar data images, stitching these 10 data images together to obtain the final result, and mapping this final result back to the original image to mark the target location on the original image. This method can effectively capture the subtle features of small radar targets.

Description

Technical Field

[0001] This invention belongs to the field of radar target detection technology, and relates to a deep network radar target detection method based on multi-branch combination. Background Technology

[0002] Radar small target detection refers to the detection and tracking of tiny targets using radar signal processing technology. It is commonly used in military, aviation, aerospace, and geological exploration fields. Compared to traditional detection technologies such as optical and infrared, radar technology is unaffected by external factors such as weather and lighting conditions, and can perform detection in special environments such as at night and in rainy or snowy weather.

[0003] The traditional radar target detection technology process mainly includes: radar signal processing technology, small target detection algorithms, target feature extraction technology, and data fusion technology. Traditional radar small target detection methods are rule-based and feature-based, typically requiring manual design and extraction of target features, and using simple algorithms and features for target detection. These methods have several drawbacks, including the significant manpower and time required for feature engineering, limited detection performance, poor adaptability, and difficulty in handling large-scale data.

[0004] The emergence of deep learning technology has brought new solutions for radar small target detection. Deep learning-based radar small target detection methods can automatically learn target features and use deep learning algorithms such as convolutional neural networks to achieve target detection. The advantages of this method include automatic learning and strong adaptability, ability to handle large-scale data, and high accuracy. However, existing deep learning-based radar target detection methods are mostly based on models for detecting large objects, which leads to low detection performance when applied to small target signals such as radar. Summary of the Invention

[0005] The purpose of this invention is to provide a deep network radar target detection method based on multi-branch combination, which solves the problems of traditional radar target detection methods and existing depth models requiring a lot of manpower and time, having limited detection performance, poor adaptability, and difficulty in handling large-scale data in radar small target detection.

[0006] The technical solution adopted in this invention is a deep network radar target detection method based on multi-branch combination, implemented according to the following steps:

[0007] Step 1: Acquire radar echo signals, normalize the radar echo signals, and obtain radar data. a ;

[0008] Step 2, processing radar data. a Perform data slicing to obtain 10 radar data images. b Set the model's extreme value threshold K;

[0009] Step 3, retrieve radar data. a By copying radar data twice a Increase the number of channels and increase radar data. b The number of channels is 3, and radar data is obtained. c , will radar data c The radar data (Data) is obtained by performing layer normalization again. d ;

[0010] Step 4, transfer radar data. d The data is resized, flattened, and then fed into the Transformer decoder branch; it then passes through Layer Norm to obtain data Output1, which is then resized and deconvolved to obtain data Output. 11 ;

[0011] Step 5, transfer radar data. d The size is adjusted, and the data is fed into a pyramid convolutional structure for feature extraction to obtain data Output2.

[0012] Step 6, Connect Data Output 11 The data Output2 is processed by convolution to obtain the response graph data Output3;

[0013] Step 7: Process each element of the response graph data Output3 to obtain data Output4;

[0014] Step 8: Repeat steps 3 through 7, processing each of the 10 radar data points. b Process the data according to the relevant steps to obtain 10 data Output4 images. Concatenate these 10 data Output4 images to obtain the final output result. Map the final output result back to the original image to mark the target location on the original image.

[0015] The beneficial effects of this invention include the following aspects:

[0016] 1) This invention solves the shortcomings of traditional radar target detection methods, such as requiring a lot of manpower and time, limited detection performance, poor adaptability, and difficulty in processing large-scale data.

[0017] 2) The deep network of the present invention adopts a multi-layer convolutional structure, which can effectively capture the subtle features of small radar targets and thus detect small radar targets.

[0018] 3) In the process of feature extraction of radar targets, the present invention incorporates a Transformer encoder structure. This structure not only helps to extract the features of the target, but also enriches the feature information due to the cross-correlation operation between various image elements. Attached Figure Description

[0019] Figure 1 This is a network structure block diagram of the method of the present invention;

[0020] Figure 2 This is a schematic diagram illustrating an embodiment of the method of the present invention;

[0021] Figure 3 This is a visualization of the original radar test data input to the method of this invention;

[0022] Figure 4 This is a visualization of the test data output results obtained by the method of this invention;

[0023] Figure 5 This is a schematic diagram of the pyramid convolution structure used in the method of this invention;

[0024] Figure 6 This is a schematic diagram of the Transformer encoder used in the method of the present invention. Detailed Implementation

[0025] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0026] Reference Figure 1 The deep network radar target detection method of the present invention is based on multi-branch combination and is implemented in accordance with the following steps:

[0027] Step 1: Acquire radar echo signals, normalize the radar echo signals to obtain radar data of size 260*4000*1. a ,

[0028] Radar equipment transmits pulse signals, usually multiple pulses, to improve signal quality and reduce the error rate. After propagating through the air, the transmitted pulse signals are scattered at the target, forming echo signals. These echo signals are then received by a receiving antenna and transmitted to the radar receiver. Upon receiving the echo signals, they undergo filtering, demodulation, and digitization processes to obtain the radar's RD matrix data. ori The data is 260*4000*1 in size. A visualization of the original radar test data can be found here. Figure 3 Data information of RD matrix ori Radar data is obtained after layer normalization. a ,

[0029] The specific operations for layer normalization are as follows:

[0030] 1.1) For a given sample, calculate its mean and variance using the following expressions:

[0031]

[0032] Where d represents Data ori The amount of data contained in a channel, here Data ori The data size of one channel is 260*4000, x i Represents Data ori The numerical value of each data item in a channel;

[0033] 1.2) Calculate Data ori Each data x in a channel i The result after normalization y i The expression is:

[0034]

[0035] Where i = 1, 2, ..., 260 * 4000, y i It is x i The result of normalization is that γ and β are scaling factors and offsets used for learning, and ò is a small constant used to avoid division by zero;

[0036] Data ori After the above processing, normalized radar data is obtained. a .

[0037] Step 2, processing radar data. a Data slicing was performed to obtain 10 radar data images, each with a size of 260*400. b Set the model's extreme value threshold K.

[0038] The radar data obtained in step 1 a Perform data slicing, specifically slicing the data of size 260*4000*1.

[0039] The specific operation process is as follows: Maintain radar data. a With the number of rows remaining unchanged, the columns are divided into equidistant segments of length 400, resulting in 10 radar data images of size 260*400*1. b Meanwhile, when training the model, a threshold K is set for the extreme value of the model. This threshold is a hyperparameter. When the element of the model's final output data reaches the hyperparameter K, it is considered as the target element.

[0040] Step 3, retrieve radar data. a By copying radar data twice a Increase the number of channels and increase radar data. b The number of channels is 3, and radar data is obtained. c Radar data c The current size is 260*400*3; [Data from radar] c Based on formulas (1) and (2), layer normalization is performed again to obtain the radar data Data input into the model. d .

[0041] Step 4, transfer radar data. d The size is adjusted to 224*224*3, and after a Flatten operation, it is fed into the Transformer decoder branch; then, after Layer Norm, a data Output1 of size 197*768 is obtained. Data Output1 is then resized and deconvolved to obtain an Output of size 260*400*1. 11 ,

[0042] The specific process is as follows:

[0043] 4.1) Radar Data d The PIL library's reshape function is used to resize the image to 224*224*3. The resize() function performs interpolation on the image to generate new pixel values ​​and rearranges the pixels according to the specified size to generate a new image.

[0044] 4.2) After resizing, the data is sliced ​​into 16*16 image segments and then flattened into a data sequence, which is then fed into the decoder branch of the Transformer.

[0045] The Transformer's Positional Encoding Layer adds positional information to each word in the input sequence so that the model can learn the positional relationships within the sequence.

[0046] Multi-Head Self-Attention Layer: This layer uses a self-attention mechanism to interact with the input sequences, thereby obtaining the importance weight of each sequence in the overall sequence.

[0047] Encoder-Decoder Attention Layer: This layer uses the output of the encoder and the output of the previous decoder to interact with each other and obtain relevant information between the input and output sequences.

[0048] Feedforward Neural Network Layer: A feedforward neural network layer is a fully connected feedforward neural network that typically contains two linear layers and an activation function. Its function is to transform the embedding vector at each location into a higher-dimensional hidden representation for further processing.

[0049] 4.3) After Layer Norm, we get a 197*768 data Output1. Output1 is then resized and deconvolved to get a data Output of size 260*400*1. 11 .

[0050] Step 5, transfer radar data. d The size was adjusted to 224*224*3, and then fed into a pyramid convolutional structure for feature extraction, resulting in feature data Output2 with a size of 260*400*3.

[0051] The specific process is: transferring radar data... d The size is adjusted to 224*224*3 and fed into a pyramid convolutional structure to extract features. This pyramid convolutional structure is a five-layer feature extraction network, consisting of five, four, three, two, and one layers from top to bottom. This pyramid convolutional structure obtains local detail features of the radar target through upsampling and downsampling, resulting in feature data Output2 with a size of 260*400*3.

[0052] Step 6, Connect Data Output 11 Output2 and the response graph Output3 are obtained through a convolution operation.

[0053] The specific process is as follows: using the `cat` function of the PyTorch library, the data is connected to the Output channel based on the third channel of the data. 11 Output2 is used to obtain fused data that combines the features of the two branches. This data has a size of 260*400*4. Then, after convolution, the response map data Output3 is obtained, which has a size of 260*400*1.

[0054] Step 7: Process each element of the response map data Output3. If the element of the response map data is greater than the model extreme value threshold K, set the value of the element to 1, indicating that the location is the location of a small radar target; otherwise, set the value of the element to 0, indicating that the location is not the location of a small radar target. After the above processing, the data Output4 is obtained.

[0055] Step 8: Repeat steps 3 through 7, processing each of the 10 radar data points. b Following the relevant steps, 10 data outputs (Output4) are obtained. These 10 Output4 data are then concatenated to obtain the final output. This final output is mapped back to the original image to mark the target location on the original image. The model detection results for the test data are shown below. Figure 4 .

[0056] Experimental verification:

[0057] Reference Figure 2 The method embodiments of the present invention are implemented according to the following steps:

[0058] Step 1: Acquire and preprocess radar data to obtain the radar's RD matrix data information. ori The data is 260*4000*1 in size. A visualization of the original radar test data can be found here. Figure 3 Set the model's extreme value threshold K to 0.5, and simultaneously adjust the Data... ori Radar data obtained by performing layer normalization processing. a The specific operation of layer normalization is as follows:

[0059] 1.1) For a given sample, calculate its mean and variance:

[0060]

[0061] Where d represents Data ori The amount of data contained in a channel, here Data ori The data size of one channel is 260*4000, x i Represents Data ori The numerical value of each data point in a channel.

[0062] 1.2) Calculate Data ori Each data x in a channel i The result after normalization y i :

[0063]

[0064] Where i = 1, 2, ..., 260 * 4000, y i It is x i The result of normalization, where γ and β are scaling factors and offsets used for learning, and ò is a small constant used to avoid division by zero; Data ori After the above processing, normalized radar data is obtained. a .

[0065] Step 2, process the obtained radar data. a Perform data slicing. Specifically, the radar data queue with a size of 260*4000*1 is sliced ​​at equal intervals with a length of 300, resulting in 10 radar data images of size 260*400*1. b At the same time, when training the model, a threshold K is set for the model's extreme values. This threshold is a hyperparameter. When an element in the model's final output data reaches this hyperparameter, that element is considered the target.

[0066] Step 3, retrieve radar data. a By copying radar data twice a Increase the number of channels and increase radar data. b The number of channels is 3, and radar data is obtained. c Radar data c The size at this point is 260*400*3, radar data. c Based on formulas (1) and (2), layer normalization is performed again to obtain the radar data Data input into the model. d ;

[0067] Step 4, transfer radar data. d The PIL library's `reshape` function resizes the image to 224*224*3. The `resize()` function performs interpolation on the image, generating new pixel values ​​and rearranging the pixels according to the specified size to create a new image. Interpolation operations are performed using different algorithms as needed, such as nearest-neighbor interpolation, bilinear interpolation, and bicubic interpolation.

[0068] Reference Figure 6The resized data is sliced ​​into 16*16 image segments and then flattened to form a data sequence, which is then fed into the Transformer's decoder branch. The Transformer's Positional Encoding Layer adds positional information to each word in the input sequence, allowing the model to learn the positional relationships between sequences. The Multi-Head Self-Attention Layer interacts with the input sequence using self-attention, acquiring the importance weight of each sequence within the overall sequence. The Encoder-Decoder Attention Layer interacts with the output of the encoder and the output of the previous decoder layer to acquire relevant information between the input and output sequences, which is then fed into the feedforward neural network layer. After Layer Norm, a 197*768 data Output1 is obtained. Output1 is then resized and deconvolved to obtain a 260*400*1 data Output. 11 ;

[0069] Step 5, transfer radar data. d The size was adjusted to 224*224*3, and then fed into a pyramid convolutional structure for feature extraction. Table 1 shows some parameters of the pyramid convolutional structure.

[0070] Table 1. Network parameters of pyramidal convolutional structures

[0071] Layer name Input size Output size kernel size Number of convolution kernels Convolutional layer 1 300x 300x 3 300x 300x 64 3x 3 64 Convolutional layer 2 19x 19x 1024 19x 19x 256 1x 1 256 Convolutional layer 3 19x 19x 256 19x 19x 512 3x 3 512 Convolutional layer 4 19x 19x 512 19x 19x 128 1x 1 128 Convolutional layer 5 19x 19x 128 19x 19x 256 3x 3 256 Convolutional layer 6 19x 19x 256 19x 19x 128 1x 1 128 Convolutional layer 7 19x 19x 128 19x 19x 256 3x 3 256 8 convolutional layers 19x 19x 256 19x 19x 128 1x 1 128 Convolutional layer 9 19x 19x 128 19x 19x 256 3x 3 256 Convolutional layer 10 19x 19x 256 19x 19x 128 1x 1 128 Convolutional layer 11 19x 19x 128 19x 19x 256 3x 3 256 Convolutional layer 12 19x 19x 256 19x 19x 128 1x 1 128 Convolutional layer 13 19x 19x 128 19x 19x 256 3x 3 256 Convolutional layer 14 19x 19x 256 19x 19x 256 3x 3 256 Convolutional layer 15 19x 19x 256 19x 19x 425 1x 1 425

[0072] Reference Figure 5 The pyramid convolutional structure is a five-layer feature extraction network, consisting of five, four, three, two, and one layers from top to bottom. This pyramid convolutional structure obtains local detail features of the radar target through upsampling and downsampling methods, resulting in feature data Output2 of size 260*400*3.

[0073] Step 6: Using the `cat` function in the PyTorch library, connect the data output based on the third channel of the data. 11 Output2 and the data output together produce a fused data set that combines the features of the two branches. This fused data set has a size of 260*400*4. After convolution, the response map data Output3 is obtained, which has a size of 260*400*1.

[0074] Step 7: Process each element of the response map data Output3. For elements whose response map data is greater than the model's extreme value threshold K, set the value of that element to 1, indicating that the location is a small radar target; otherwise, set the value of that element to 0, indicating that the location is not a small radar target. After the above processing, data Output4 is obtained.

[0075] Step 8: Repeat steps 3 through 7, processing each of the 10 radar data points. b Following the relevant steps, 10 data points (Output4) are obtained. These 10 Output4 data points are then concatenated to obtain the final output. This final output can be mapped back to the original image to mark the target location.

[0076] Figure 4 To be Figure 3 The test data is input into the model to obtain the detection results. Figure 3 Analysis shows that this radar small target detection method not only ensures the accuracy of radar small target detection, but also has a good effect on suppressing interference.

Claims

1. A radar target detection method based on multi-branch combined deep network, characterized in that, The following steps are implemented: Step 1, collect the radar echo signal, normalize the radar echo signal, and obtain radar data Data a ; Step 2, data slicing operation is performed on the radar data Data a to obtain 10 radar data Data b , and a model extreme threshold K is set. Step 3, take a radar data Data a , increase the channel number of the radar data Data a to 3 by copying the channel number of the radar data Data b twice, obtain the radar data Data c , and perform layer normalization processing on the radar data Data c again to obtain the radar data Data d ; Step 4, the radar data Data d Resized, sent into the Transformer decoder branch through the Flatten operation; and then Layer Norm is performed to obtain data Output1, and the data Output1 is subjected to resize and deconvolution operations to obtain data Output 11 ; Step 5, radar data Data d Resizing, sent into the pyramid convolution structure to extract features, get data Output2; Step 6, connecting data Output 11 and data Output2, through convolution operation to obtain response map data Output3; Step 7, process each element of the response map data Output3 to obtain data Output4; Step 8, repeat steps 3 to 7 to 10 radar data Data b According to the related steps, 10 data Output4 are obtained, the 10 data Output4 are spliced to obtain a final result output, and the final result output is mapped back to the original image to label the target position on the original image. 2.The multi-branch combined based deep network radar target detection method of claim 1, wherein, In step 1, the specific process is: After the radar receiver receives the echo signal, the signal is filtered, demodulated and digitized to obtain the radar RD matrix data information Data ori , the RD matrix data information Data ori is processed by layer normalization to obtain radar data Data a , The specific operation of layer normalization is as follows: 1.1) For a sample, calculate the mean and variance of the sample, expressed as: wherein d represents Data ori The amount of data contained in one channel, here Data ori The amount of data in one channel is 260*4000, x i represents Data ori The value size of each data in one channel; 1.2) Compute Data ori Each data x in a channel i Normalized result y i The expression is: where i = 1, 2,..., 260*4000, y i is x i normalized corresponding results, γ and β are scaling factor and offset for learning, is a small constant; Data ori After the above processing, normalized radar data Data a is obtained. 3.The multi-branch combined based deep network radar target detection method of claim 1, wherein, In step 2, the specific process is: Keep the number of rows of radar data Data a unchanged, make equidistant cutting with length of 400 on the number of columns, get 10 radar data Data with size of 260*400*1 b ; At the same time, set the model extreme threshold K when training the model, which is a hyperparameter. When the last output data element of the model is the hyperparameter K, it is considered as the target. 4.The multi-branch combined based deep network radar target detection method of claim 1, wherein, In step 4, the specific process is, 4.1) Radar data Data d The size is adjusted to 224*224*3 by the reshape function of the PIL library, and the resize() function generates new pixel values by interpolating the image and rearranging the pixels according to the specified size to generate a new image. 4.2) After resizing, the data is sliced into 16*16 image segments and then flattened into a data sequence for input into the decoder branch of the Transformer, The position encoding layer of the Transformer: adds position information to each word in the input sequence so that the model can learn the sequence position relationship; Multi-head self-attention layer: uses self-attention mechanism to interact with the input sequence, thereby obtaining the importance weight of each sequence in the whole sequence; Encoder-decoder attention layer: uses the output of the encoder and the output of the previous layer of the decoder to perform attention interaction to obtain the relevant information between the input sequence and the output sequence; Feedforward neural network layer: contains two linear layers and an activation function, which converts the embedding vector of each position into a higher-dimensional hidden representation for further processing; 4.3) After Layer Norm, the data Output1 with the size of 197*768 is obtained, and the data Output1 is subjected to resize and deconvolution operation to obtain data Output with the size of 260*400*1 11 .

5. The multi-branch combination-based deep network radar target detection method of claim 1, wherein, In step 5, the specific process is: The radar data Data d The size is adjusted to 224*224*3, and the pyramid convolution structure is used to extract features; the pyramid convolution structure is a five-layer feature extraction network, and from top to bottom, it is five layers, four layers, three layers, two layers, and one layer; the pyramid convolution structure obtains feature data Output2 with a size of 260*400*3 through upsampling and downsampling. 6.The multi-branch combined based deep network radar target detection method of claim 1, wherein, In step 6, the specific process is: Using the `cat` function in the PyTorch library, the data is connected to the Output channel based on the third channel of the data. 11 Output2 is used to obtain fused data that combines the features of the two branches. This data has a size of 260*400*4. Then, after convolution, the response map data Output3 is obtained, which has a size of 260*400*1.

7. The multi-branch combination based deep network radar target detection method of claim 1, wherein, In step 7, the specific process of processing the response map data is: if the element of the response map data is greater than the model extreme threshold K, set the value of the element to 1, indicating that this position is a radar small target position; otherwise, set the value of the element to 0, indicating that this position is not a radar small target position.

Images