Deep Network SAR Target Detection Method Based on Complex Image Information

Abstract

The present invention discloses a deep network SAR target detection method based on complex image information, including: respectively performing amplitude information extraction and complex information extraction on a given training set and a test set to obtain a new training set and a new test set; constructing a SAR target detection network based on the CenterNet target detection algorithm; wherein, the target detection network includes an attention mechanism module, a feature extraction and fusion module, and a predictor module; constructing a coordinated adversarial loss function and training the target detection network with the new training set to obtain a trained target detection network; inputting the new test set into the trained target detection network to obtain the final detection result. The method provided by the present invention simultaneously solves the problems of insufficient information volume of existing algorithms, high algorithm complexity and easy occurrence of low-quality detection frames, and has a certain degree of improvement compared with existing methods in terms of indexes such as precision and recall rate.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar target detection, and particularly relates to a deep network SAR target detection method based on complex image information. Background Art

[0002] Since the synthetic aperture radar came out in the 1950s, the SAR image target detection technology has also been developing rapidly. In 1966, Finn et al. proposed the first Constant False Rate (CFAR) detection method based on the Bayesian theory and applied it to the target detection task of one-dimensional radar signals. Subsequently, many scholars such as Burl and Novak from the Lincoln Laboratory of the United States began to conduct research on SAR image target detection and proposed the famous two-parameter CFAR detection method in the 1990s, successfully extending the CFAR detector to the two-dimensional SAR image target detection field. The core idea of the two-parameter CFAR is based on a sliding window, and the intensity values of the pixel points in the SAR image to be detected are compared with the adaptive threshold one by one, where the adaptive threshold is determined by the given false alarm rate and the background clutter distribution within the sliding window. Since the two-parameter CFAR has a good performance in the SAR image detection task in a simple scenario, it has been deeply studied and has become a widely used SAR image target detection method. After that, in order to further improve the performance of the CFAR detector, many researchers have proposed a large number of improved SAR image target detection algorithms from different angles such as the clutter distribution statistical model and the background clutter sample selection method, such as the SAR image target detection methods for different background clutter distribution statistical models (Gamma distribution, generalized Gamma distribution, K distribution); for the problems of missed alarms and false alarms caused by complex problems such as multi-target interference and clutter edges in SAR images, researchers have also proposed detectors such as minimum selection CFAR, maximum selection CFAR, and sequential statistical CFAR.

[0003] With the development of deep learning, deep network object detection algorithms based on candidate boxes have gradually been applied to SAR object detection tasks, such as SSD, YOLO, Faster-RCNN, EfficientDet, etc. For example, Patent CN202210269829.6 (A CFAR-guided dual-stream SSD SAR image object detection method) introduces a SAR object detection method based on the SSD object detection network. This method introduces the results of the traditional method CFAR as prior information into the SSD network and improves the network structure and loss function, which to a certain extent improves the detection results. However, since SSD is a detection algorithm based on candidate boxes, a large number of candidate boxes need to be laid in the original image, and then subsequent processing operations are performed on the candidate boxes. This approach has a high complexity and is prone to generating low-quality detection boxes. Therefore, there is still much room for improvement in the results of this method.

[0004] Existing deep network object detection algorithms based on candidate boxes still have some defects. On the one hand, such methods only use the original image as input for detection, and the single information cannot fully represent the characteristics of the target, making it easier for the detection performance to reach a bottleneck. On the other hand, detection algorithms based on candidate boxes need to lay a large number of candidate boxes in the original image and then perform subsequent processing operations on the candidate boxes. This approach has a high complexity and is prone to generating low-quality detection boxes. Summary of the Invention

[0005] To solve the above problems existing in the prior art, the present invention provides a deep network SAR object detection method based on complex image information. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0006] A deep network SAR object detection method based on complex image information includes:

[0007] Step 1: Extract amplitude information and complex information from the given training set Φ and test set T respectively to obtain a new training set Φ' and a new test set T';

[0008] Step 2: Build a SAR object detection network ψ based on the CenterNet (a candidate-box-free deep network object detection model) object detection algorithm; wherein, the object detection network ψ includes an attention mechanism module, a feature extraction and fusion module, and a predictor module;

[0009] Step 3: Build a coordinated adversarial loss function and use the new training set Φ' to train the object detection network ψ to obtain a trained object detection network ψ';

[0010] Step 4: Input the new test set T' into the trained object detection network ψ' to obtain the final detection result.

[0011] Advantages of the present invention:

[0012] 1. On the one hand, by introducing prior information such as SAR image complex information as the input of the network, the present invention provides more helpful information for the network, and the amount of information is richer than that of a single CFAR prior information, solving the problem of insufficient information. On the other hand, the CenterNet object detection algorithm is used, which does not require laying a large number of candidate boxes on the original image, and to a certain extent solves the problems of high complexity of the prior art and the appearance of low-quality detection boxes.

[0013] 2. The present invention uses a coordinated adversarial loss function to train the network, and at the same time adds a feature fusion module to the network, achieving a certain degree of improvement in indicators such as precision and recall compared with the original method.

[0014] The following will further elaborate on the present invention in conjunction with the accompanying drawings and embodiments. Description of the Drawings

[0015] Figure 1 is a schematic flowchart of a method for SAR target detection based on a deep network of complex image information provided by an embodiment of the present invention;

[0016] Figure 2 is another schematic flowchart of a method for SAR target detection based on a deep network of complex image information provided by an embodiment of the present invention;

[0017] Figure 3 is a framework diagram of an object detection network constructed based on the CenterNet object detection algorithm provided by an embodiment of the present invention;

[0018] Figure 4 is a framework diagram of an attention mechanism module provided by an embodiment of the present invention;

[0019] Figure 5 is a structural framework diagram of a residual module provided by an embodiment of the present invention;

[0020] Figure 6 is a structural framework diagram of a downsampling convolution module provided by an embodiment of the present invention;

[0021] Figure 7 is a structural framework diagram of a consistency convolution module provided by an embodiment of the present invention;

[0022] Figure 8 is a structural framework diagram of an upsampling feature fusion module provided by an embodiment of the present invention;

[0023] Figure 9 is a structural framework diagram of a downsampling feature fusion module provided by an embodiment of the present invention;

[0024] Figures 10 - 15 is the MiniSAR data image adopted in the embodiment of the present invention; Specific implementation manner

[0025] The present invention will be further described in detail below in conjunction with specific embodiments, but the implementation manners of the present invention are not limited thereto.

[0026] Embodiment 1

[0027] Please refer jointly to Figures 1 - 2 , Figure 1 which is a schematic flow chart of a deep network SAR target detection method based on complex image information provided by the embodiment of the present invention, Figure 2 and which is another schematic flow chart of a deep network SAR target detection method based on complex image information provided by the embodiment of the present invention. The deep network SAR target detection method based on complex image information provided by the present invention includes:

[0028] Step 1: Extract amplitude information and complex information from the given training set Φ and test set T respectively to obtain a new training set Φ' and a new test set T'.

[0029] First, in this embodiment, the acquired original SAR image can be directly divided into a training set Φ and a test set T according to a ratio.

[0030] Then, perform amplitude information extraction and complex information extraction operations on the training set Φ and the test set T respectively to obtain 6 statistical images with the same length and width as the original image. Concatenate the 6 statistical images along the channel dimension, and then perform slice extraction respectively to obtain the training set Φ' and the test set T', specifically as follows:

[0031] 11) Perform modulus operations on the original data of the training set Φ and the test set T respectively to obtain the first-dimensional statistical images Φ D1 , T D1 .

[0032] 12) Calculate the real signal kurtosis RSK and the non-circularity measure Ψ 0 for the training set Φ and the test set T respectively, and obtain the second-dimensional statistical images Φ D2 , T D2 and the third-dimensional statistical images Φ D3 , T D3 .

[0033] In this embodiment, the real signal kurtosis and the non-circularity measure are mainly calculated in the form of traversing a sliding window, specifically as follows:

[0034] 12a) Let the original complex data of any image in the training set Φ and the test set T be G0, and traverse G0 according to a sliding window of a certain size.

[0035] Optionally, as an implementation, in this embodiment, a sliding window of size 13×13 is used to traverse G0.

[0036] 12b) Calculate the real signal kurtosis RSK of all pixel points in the window at each position of the sliding window, and obtain the second-dimensional statistical quantity map Φ of the same scale as the original image D2 , T D2 ; The calculation formula of the real signal kurtosis RSK is:

[0037]

[0038] where N is the total number of pixel points in the sliding window, is the intensity value of the real part of each pixel point in the sliding window, is the average value of the real part of the pixel points in the sliding window, is the intensity value of the imaginary part of each pixel point in the sliding window, is the average value of the imaginary part of the pixel points in the sliding window.

[0039] 12c) Calculate the non-circularity measure ψ of all pixel points in the window at each position of the sliding window 0 , and obtain the third-dimensional statistical quantity map Φ of the same scale as the original image D3 , T D3 ; The non-circularity measure ψ 0 The calculation formula is:

[0040]

[0041] where N is the total number of pixel points in the sliding window, represents the complex data composed of the real part intensity value and the imaginary part intensity value of each pixel point.

[0042] 13) Perform sub-aperture decomposition on the original data of the training set Φ and the test set T respectively, and calculate the real correlation coefficient ρ I and the complex correlation coefficient ρ C , and correspondingly obtain the fourth-dimensional statistical quantity map Φ D4 , T D4 and the fifth-dimensional statistical quantity map Φ D5 , T D5 .

[0043] 13a) Perform azimuth Fourier transform on the original complex data G0 of any image in the training set Φ and the test set T to obtain the azimuth spectrum map H1.

[0044] 13b) Estimate the azimuth weighting function FW(·) using the azimuthal spectrogram H1, and multiply the reciprocal thereof with the azimuthal spectrogram H1 to obtain a weighted spectrogram H2.

[0045] 13c) Divide the weighted spectrogram H2 into a number of sub-bands of the same length in the azimuth direction, shift the center frequency of each sub-band to the 0 frequency, and multiply each sub-band with the weighting function FW(·).

[0046] 13d) Perform zero-padding on each weighted sub-band, and perform inverse azimuthal Fourier transform on each sub-band to obtain a number of sub-aperture images.

[0047] In this embodiment, the weighted spectrogram H2 can be divided into 4 sub-bands of the same length in the azimuth direction, and correspondingly 4 sub-aperture images can be obtained, denoted as Sub1, Sub2, Sub3, and Sub4.

[0048] 13e) Calculate the statistics ρ I and ρ C , and respectively obtain the fourth-dimensional statistic map Φ D4 , T D4 and the fifth-dimensional statistic map Φ D5 , T D5 , and the calculation formula is:

[0049]

[0050] ρ C = log(1 - |R s |)

[0051] where

[0052]

[0053] N represents the number of sub-bands, s i represents the complex-valued data of the pixel at the corresponding position of the i-th sub-band image Sub i , <·> represents the spatial averaging operation, which is implemented by the average value within a 13×13 sliding window.

[0054] 14) Perform Gaussian downsampling, central wrapping, cross-scale addition, and normalization operations on the original data of the training set Φ and the test set T in sequence, and correspondingly obtain the sixth-dimensional statistic map Φ D6 , T D6 .

[0055] 14a) Perform a modulus operation on the original complex data G0 of any image in the training set Φ and the test set T to obtain an intensity image G1.

[0056] 14b) Perform continuous Gaussian filtering and downsampling on the intensity image G1 for m times. Each time of downsampling, the length and width become half of the original, obtaining an m + 1 - layer Gaussian pyramid P i , where i = 0, 1…m, and the 0 - scale is the largest and the m - scale is the smallest.

[0057] Preferably, in this embodiment, continuous Gaussian filtering and downsampling are performed on the intensity image G1 for 8 times, and then a 9 - layer Gaussian pyramid P can be obtained i , where i = 0, 1…8.

[0058] 14c) For the m + 1 - layer Gaussian pyramid obtained in step 14b), respectively select multiple combinations including several layers for cross - scale subtraction to obtain multiple center - surround images. The formula is expressed as:

[0059] I(c, s)=|P c - P s |

[0060] where “ - ” represents cross - scale subtraction, P c represents the large - scale layer, and P s represents the small - scale layer.

[0061] Specifically, first linearly interpolate and upsample the small - scale layer P s to the same scale as P c , and then calculate the difference from P c for each pixel point to obtain multiple center - surround images.

[0062] For example, for a 9 - layer Gaussian pyramid, the 4 combinations of 0 - 4, 1 - 4, 1 - 5, and 2 - 5 can be selected respectively for cross - scale subtraction to obtain the center - surround images C 0-4 , C 1-4 , C 1-5 , C 2-5 .

[0063] 14d) Linearly interpolate and upsample multiple center - surround images to the original image size, add them by pixel points, and then perform normalization operation according to the image gray level to obtain the sixth - dimensional statistic map Φ D6 , T D6 .

[0064] The specific processes of operations such as linear interpolation upsampling and normalization here can refer to the prior art for implementation, and are not introduced in detail in this embodiment.

[0065] 15) Respectively splice the statistic maps Φ D1 - Φ D6 , T D1 - T D6 of the first dimension to the sixth dimension along the channel dimension, correspondingly obtaining a 6 - channel image Φ of the same size as the original image6D and T 6D 。

[0066] 16) Perform sliding window slicing operations on the 6-channel images Φ 6D and T 6D respectively, to obtain a number of slices, denoted as the new training set Φ' and the new test set T'.

[0067] Specifically, sliding window slicing operations with a size of 512×512 and step sizes of 40 and 75 respectively can be performed on Φ 6D and T 6D respectively, to obtain a number of slices with a size of 512×512×6, denoted as the new training set Φ' and the new test set T'.

[0068] Step 2: Construct a SAR target detection network ψ based on the CenterNet target detection algorithm.

[0069] Please refer to Figure 3 , Figure 3 which is the framework diagram of the target detection network constructed based on the CenterNet target detection algorithm provided by the embodiment of the present invention. The target detection specifically includes an attention mechanism module, a feature extraction and fusion module, and a predictor module.

[0070] The following will introduce the above three modules in detail in turn.

[0071] 1. Attention mechanism module

[0072] Please refer to Figure 4 , Figure 4 which is the network framework diagram of the attention mechanism module provided by the embodiment of the present invention. It includes a convolutional neural network A for performing preliminary feature processing on the input image to facilitate better feature extraction by the subsequent feature extraction network; among them, the convolutional neural network A includes Figure 4 the first pooling layer L1, the second splicing layer L2, the third convolutional layer L3, the fourth activation function layer L4, the fifth element multiplication layer L5, the sixth batch normalization layer L6, the seventh pooling layer L7, the eighth fully connected layer L8, the ninth fully connected layer L9, the tenth fully connected layer L 10 、the eleventh element addition layer L 11 、the twelfth activation function layer L 12 、the thirteenth element multiplication layer L 13 。

[0073] Specifically, the input of the convolutional neural network A is j = 1, 2…6, and the input, output, parameter settings and relationships of each layer are as follows:

[0074] The first pooling layer L1, with a pooling dimension of the channel dimension, and pooling methods of max pooling and average pooling, is used to output one max-pooling feature map and one average-pooling feature map This layer serves as the input to the second concatenation layer L2;

[0075] The second concatenation layer L2, with a concatenation dimension of the channel dimension, is used to output one feature Figure X 2 , and this layer serves as the input to the third convolutional layer L3;

[0076] The third convolutional layer L3, with a convolutional kernel K 3 having a window size of 7×7 and a sliding stride S 3 of 1, is used to output one feature Figure X 3 , and this layer serves as the input to the fourth activation function layer L4;

[0077] The fourth activation function layer L4, with a sigmoid function as its activation function, is used to output one weighted feature Figure X 4 , and this layer, together with the input j = 1, 2... 6 of the first pooling layer L1, serves as the input to the fifth element-wise multiplication layer L5;

[0078] The fifth element-wise multiplication layer L5, with its initial features being j = 1, 2... 6 and a weighted feature of X 4 , and with an element-wise multiplication in the spatial dimension as the weighting method, is used to output six feature maps j = 1, 2... 6, and this layer serves as the input to the sixth batch normalization layer L6;

[0079] The sixth batch normalization layer L6, with its normalization method being is used to output six feature maps j = 1, 2... 6, and this layer serves as the input to the seventh pooling layer L7;

[0080] The seventh pooling layer L7, with a pooling dimension of the spatial dimension and pooling methods of max pooling and average pooling, is used to output one max-pooling column vector and one average-pooling column vector This layer serves as the input to the eighth fully-connected layer L8;

[0081] The eighth fully-connected layer L8, which has six neurons, is used to output two 3D column vectors X 8 , and this layer serves as the input to the ninth fully-connected layer L9;

[0082] The ninth fully-connected layer L9, which has three neurons, is used to output two 6D column vectors X 9, this layer serves as the tenth element addition layer L 10 as the input of;

[0083] The tenth element addition layer L 10 , whose addition method is channel-wise addition, and is used to output a 6-dimensional column vector X 10 , this layer serves as the input of the eleventh activation function layer L 11 as the input of;

[0084] The eleventh activation function layer L 11 , whose activation function is the Sigmoid function, and is used to output a 6-dimensional column vector X 11 , this layer and the input of the seventh pooling layer L7 j = 1, 2…6 together serve as the input of the twelfth element multiplication layer L 12 as the input of;

[0085] The twelfth element multiplication layer L 12 , whose initial feature is j = 1, 2…6, and the weight feature is X 11 , and the weighting method is element-wise multiplication in the channel dimension, and is used to output 6 feature maps j = 1, 2…6, this layer serves as the input of the convolutional neural network B.

[0086] 2. Feature extraction and fusion module

[0087] In this embodiment, the feature extraction and fusion module includes a convolutional neural network B; among them, as shown in the figure, the convolutional neural network B includes a first convolutional layer L'1, a second batch normalization layer L'2, a third activation function layer L'3, a fourth pooling layer L'4, a fifth convolutional residual module M5, a sixth convolutional residual module M6, a seventh convolutional residual module M7, an eighth convolutional residual module M8, a ninth convolutional layer L'9, a tenth upsampling feature fusion module M 10 , an eleventh upsampling feature fusion module M 11 , a twelfth upsampling feature fusion module M 12 , a thirteenth upsampling feature fusion module M 13 , a fourteenth convolutional layer L' 14 , a fifteenth downsampling feature fusion module M 15 , a sixteenth downsampling feature fusion module M 16 , a seventeenth downsampling feature fusion module M 17 , an eighteenth downsampling feature fusion module M 18 , a nineteenth transposed convolutional layer L' 19 , a twentieth transposed convolutional layer L' 20 and a twenty-first transposed convolutional layer L' 21 .

[0088] Please refer toFigures 5 - 9 , Figure 5 is the structural framework diagram of the residual module provided by the embodiment of the present invention; Figure 6 is the structural framework diagram of the downsampling convolution module provided by the embodiment of the present invention; Figure 7 is the structural framework diagram of the consistency convolution module provided by the embodiment of the present invention; Figure 8 is the structural framework diagram of the upsampling feature fusion module provided by the embodiment of the present invention; Figure 9 is the structural framework diagram of the downsampling feature fusion module provided by the embodiment of the present invention.

[0089] Specifically, the input of the convolutional neural network B is j = 1, 2... 6, and the input, output, parameter settings and relationships of each layer are as follows:

[0090] The first convolutional layer L'1, and its convolutional kernel K 1 ' has a window size of 3×3 and a sliding step S 1 ' of 2, and is used to output 64 feature maps j = 1, 2... 64, and this layer serves as the input of the second batch normalization layer L'2;

[0091] The second batch normalization layer L'2, and its normalization method is and is used to output 64 feature maps j = 1, 2... 64, and this layer serves as the input of the third activation function layer L'3;

[0092] The third activation function layer L'3, and its activation function is the Relu function, and is used to output 64 feature maps j = 1, 2... 64, and this layer serves as the input of the fourth pooling function layer L'4;

[0093] The fourth pooling layer L'4, and its pooling dimension is the spatial dimension, and the pooling method is max pooling, and is used to output 64 feature maps j = 1, 2... 64, and this layer serves as the input of the fifth residual module M5;

[0094] The fifth residual module M5 includes a cascade of 3 consecutive consistency convolution modules, where each consistency convolution module contains several convolution operations and a cross-layer addition structure. For the specific structure, please refer to Appendix Figures 5 - 7 , and M5 is used to output 256 feature maps j = 1, 2... 256, and this layer serves as the input of the sixth residual module M6;

[0095] The sixth residual module M6 consists of a cascade of 1 downsampling convolution module and 3 consecutive consistency convolution modules. Each downsampling convolution module and each consistency convolution module contain several convolution operations and cross-layer addition structures. For the specific structure, please refer to Appendices 5-7. M6 is used to output 512 feature maps. For j = 1, 2... 512, this layer serves as the input to the seventh residual module M7.

[0096] The seventh residual module M7 consists of a cascade of 1 downsampling convolution module and 5 consecutive consistency convolution modules. Each downsampling convolution module and each consistency convolution module contain several convolution operations and cross-layer addition structures. For the specific structure, please refer to Appendix Figures 5 - 7 , and M7 is used to output 1024 feature maps. For j = 1, 2... 1024, this layer serves as the input to the eighth residual module M8.

[0097] The eighth residual module M8 consists of a cascade of 1 downsampling convolution module and 2 consecutive consistency convolution modules. Each downsampling convolution module and each consistency convolution module contain several convolution operations and cross-layer addition structures. For the specific structure, please refer to Appendix Figures 5 - 7 , and M8 is used to output 2048 feature maps. For j = 1, 2... 2048, this layer serves as the input to the ninth convolution layer L'9.

[0098] The ninth convolution layer L'9 has a convolution kernel K 9 ' with a window size of 1×1 and a sliding stride S 9 ' of 1, and is used to output 2048 feature maps. For j = 1, 2... 2048, this layer and the output of the seventh convolutional residual module M7 For j = 1, 2... 1024 together serve as the input to the tenth upsampling feature fusion module M 10 .

[0099] The tenth upsampling feature fusion module M 10 concatenates the result of performing a 1×1 convolution on j = 1, 2... 1024 with the result of performing a transposed convolution with a stride of 2 on j = 1, 2... 2048 horizontally, and then performs a convolution with a size of 3×3×1024 and a stride of 1 on the concatenated result to obtain the output of the tenth upsampling feature fusion module M 10 . For j = 1, 2... 1024, the output of this module and the output of the sixth convolutional residual module M6 For j = 1, 2... 512 together serve as the input to the eleventh upsampling feature fusion module M 11 .

[0100] The eleventh upsampling feature fusion module M 11 , the result of performing 1×1 convolution on j = 1, 2... 512 is horizontally concatenated with the result of performing transposed convolution with a stride of 2 on j = 1, 2... 1024, and then the concatenated result is convolved with a kernel size of 3×3×512 and a stride of 1 to obtain the output of the eleventh upsampling feature fusion module M 11 ; j = 1, 2... 512, and the output of this module and the output of the fifth convolutional residual module M5 j = 1, 2... 256 together serve as the input to the twelfth upsampling feature fusion module M 12 ;

[0101] The twelfth upsampling feature fusion module M 12 , the result of performing 1×1 convolution on j = 1, 2... 256 is horizontally concatenated with the result of performing transposed convolution with a stride of 2 on j = 1, 2... 512, and then the concatenated result is convolved with a kernel size of 3×3×256 and a stride of 1 to obtain the output of the twelfth upsampling feature fusion module M 12 ; j = 1, 2... 256, and the output of this module and the output of the third activation function layer L'3 j = 1, 2... 64 together serve as the input to the thirteenth upsampling feature fusion module M 13 ;

[0102] The thirteenth upsampling feature fusion module M 13 , the result of performing 1×1 convolution on j = 1, 2... 64 is horizontally concatenated with the result of performing transposed convolution with a stride of 2 on j = 1, 2... 256, and then the concatenated result is convolved with a kernel size of 3×3×64 and a stride of 1 to obtain the output of the thirteenth upsampling feature fusion module M 13 ; j = 1, 2... 64, and the output of this module serves as the input to the fourteenth convolutional layer L' 14 ;

[0103] The fourteenth convolutional layer L' 14 , its convolution kernel K 14 ' has a window size of 1×1 and a sliding stride S 14 ' of 1, and is used to output 64 feature maps j = 1, 2... 64, and this layer and the output of the twelfth upsampling feature fusion module M 12 together serve as the input to the fifteenth downsampling feature fusion module M ; 15 ​

[0104] The fifteenth downsampling feature fusion module M 15 , the result of performing 1×1 convolution for j = 1, 2... 256 is horizontally concatenated with the result of performing convolution with a stride of 2 for j = 1, 2... 64, and then convolved with a size of 3×3×256 and a stride of 1 to obtain the output of the fifteenth downsampling feature fusion module M 15 ; j = 1, 2... 256, the output of this module and the output of the eleventh upsampling feature fusion module M 11 ; j = 1, 2... 512 are jointly used as the input of the sixteenth downsampling feature fusion module M 16 ;

[0105] The sixteenth downsampling feature fusion module M 16 , the result of performing 1×1 convolution for j = 1, 2... 512 is horizontally concatenated with the result of performing convolution with a stride of 2 for j = 1, 2... 256, and then convolved with a size of 3×3×512 and a stride of 1 to obtain the output of the sixteenth downsampling feature fusion module M 16 ; j = 1, 2... 512, the output of this module and the output of the tenth upsampling feature fusion module M 10 ; j = 1, 2... 1024 are jointly used as the input of the seventeenth downsampling feature fusion module M 17 ;

[0106] The seventeenth downsampling feature fusion module M 17 , the result of performing 1×1 convolution for j = 1, 2... 1024 is horizontally concatenated with the result of performing convolution with a stride of 2 for j = 1, 2... 512, and then convolved with a size of 3×3×1024 and a stride of 1 to obtain the output of the seventeenth downsampling feature fusion module M 17 ; j = 1, 2... 1024, the output of this module and the output of the ninth convolutional layer j = 1, 2... 2048 are jointly used as the input of the eighteenth downsampling feature fusion module M 18 ;

[0107] The eighteenth downsampling feature fusion module M 18 , the result of performing 1×1 convolution for j = 1, 2... 2048 is horizontally concatenated with the result of performing convolution with a stride of 2 for The result of performing convolution with a step size of 2 for j = 1, 2... 1024 is horizontally concatenated, and then convolution with a size of 3×3×2048 and a step size of 1 is performed to obtain the output of the eighteenth downsampling feature fusion module M 18 of For j = 1, 2... 2048, the output of this module serves as the input of the nineteenth transposed convolutional layer L' 19 ;

[0108] The nineteenth transposed convolutional layer L' 19 , whose convolutional kernel K 19 ' has a window size of 3×3 and a sliding step size S 19 ' of 2, and is used to output 256 feature maps For j = 1, 2... 256, this layer serves as the input of the twentieth transposed convolutional layer L' 20 ;

[0109] The twentieth transposed convolutional layer L' 20 , whose convolutional kernel K 20 ' has a window size of 3×3 and a sliding step size S 20 ' of 2, and is used to output 128 feature maps For j = 1, 2... 128, this layer serves as the input of the twenty - first transposed convolutional layer L' 21 ;

[0110] The twenty - first transposed convolutional layer L' 21 , whose convolutional kernel K 21 ' has a window size of 3×3 and a sliding step size S 21 ' of 2, and is used to output 64 feature maps For j = 1, 2... 64, this layer serves as the common input of the convolutional neural networks C1, C2, and C3.

[0111] 3. Predictor module

[0112] Please continue to refer to Figure 3 , the predictor module includes three independent convolutional neural networks, and each convolutional neural network includes two convolutional layers; the predictor module is used to predict the detection result.

[0113] Specifically, the three independent convolutional neural networks are C1, C2, and C3 respectively, and their corresponding two convolutional layers are L” C1-1 , L” C2-1 , L” C3-1 and L” C1-2 , L” C2-2 , L” C3-2 .

[0114] In this embodiment, the common input of the convolutional neural networks C1, C2, and C3 is j = 1, 2…64, and the input, output, parameter settings of each layer and the way of network prediction target are as follows:

[0115] The first convolutional layer L” C1-1 、L” C2-1 、L” C3-1 , whose convolutional kernels K C1-1 ”、K C2-1 ”、K C3-1 ” all have a window size of 3×3, and the sliding strides S C1-1 ”、S C2-1 ”、S C3-1 ” are all 1, and are respectively used to output 64 feature maps j = 1, 2…64、 j = 1, 2…64 and j = 1, 2…64, and the three are respectively used as the inputs of the second convolutional layer L” C1-2 、L” C2-2 、L” C3-2 .

[0116] The second convolutional layer L” C1-2 、L” C2-2 、L” C3-2 , whose convolutional kernels K C1-2 ”、K C2-2 ”、K C3-2 ” all have a window size of 1×1, and the sliding strides S C1-2 ”、S C2-2 ”、S C3-2 ” are all 1, and are respectively used to output 1 feature map with a size of 128×128 Figure X C1-2 ”、2 feature maps with a size of 128×128 j = 1, 2 and 2 feature maps with a size of 128×128 j = 1, 2.

[0117] Heat map prediction feature Figure X C1-2 ”, the physical meaning of each pixel point is the confidence prediction value of the existence of the target centered on this point. Pixel points with a confidence greater than 0.5 are set to 1, indicating that this point is the target, and pixel points less than 0.5 are set to 0, indicating that this point is not the target, to obtain a new heat Figure X C1-2 ” -bin , the 4 times of the coordinate value of each pixel point corresponds to the center position of the target in the original input image, and subsequent scale prediction and offset prediction are only performed on pixel points with a value of 1 in the heat map;

[0118] Scale prediction feature map and The physical meaning of each pixel point is the predicted values of the width w and height h of the target centered at that point in the feature map. Four times the predicted value of each pixel point corresponds to the center position of the target in the original input image. Only when the Figure X C1-2 ” -bin position of this point on the heat map is 1, and the predicted values at the corresponding points are meaningful;

[0119] The offset prediction feature map and The physical meaning of each pixel point is the predicted values of the horizontal offset Δx and vertical offset Δy of the target centered at that point in the feature map relative to the center point. Four times the predicted value of each pixel point corresponds to the center position of the target in the original input image. Only when the Figure X C1-2 ” -bin position of this point on the heat map is 1, and the predicted values at the corresponding points are meaningful.

[0120] Step 3: Construct a coordinated adversarial loss function and use the new training set Φ' to train the target detection network ψ to obtain a trained target detection network ψ'.

[0121] First, construct a coordinated adversarial loss function.

[0122] In this embodiment, a dynamic coordination loss function is constructed based on the heat map loss, scale loss, and offset loss.

[0123] Specifically, the function expression of the heat map loss is:

[0124]

[0125] where, is the ground truth, is the predicted value, i represents the i-th point on the heat map, α = 2 and β = 4 are artificially set hyperparameters, and the predicted value represents the predicted value at each position in the feature map, and the ground truth map G ab uses soft labels:

[0126]

[0127] where, the center point coordinates of the real target on the original image are p ∈ R 2 , which can be specifically expressed as p = (X, Y). After downsampling with a downsampling rate R = 4, the coordinates on the heat map become which can be specifically expressed as a and b represent the coordinates at any position on the heatmap, and σ is the standard deviation that adaptively varies according to the target size.

[0128] The functional expression of the scale loss is:

[0129]

[0130] Among them, ||·||2 represents the L2 norm, i represents the i-th real target. Assuming the size of the original image is W×H, in the original image, assuming the target size is Wi×Hi, the real size S p =(Wi, Hi), and the predicted target size is The calculation method of the loss function is The position of the target center point on the feature map is still

[0131] The functional expression of the offset loss is:

[0132]

[0133] Among them, ||·||2 represents the L2 norm, i represents the i-th real target. In the feature map, represents the real offset, including the real horizontal offset and the vertical offset That is, the real position offset of the target on the feature map; represents the predicted offset, including the predicted horizontal offset and the vertical offset That is, the position offset on the predicted feature map; The specific calculation method of the loss function is The predicted position of the target on the feature map is The real position on the feature map is (x + Δx, y + Δy). During testing, the offset is restored to the original image according to the downsampling rate to obtain the real position of the predicted target center on the original image, denoted as

[0134] Then the functional expression of the total loss function constructed based on the heatmap loss, scale loss, and offset loss is:

[0135]

[0136] Among them, N is the total number of real targets, is the coordination loss, is the adversarial loss. The sum of the two is the coordination and adversarial loss, which is calculated only at the true value G ab = 1; is calculated only at the true value G ab ≠ 1; IOU i is the IOU value of the predicted bounding box at this position and the real bounding box with the largest intersection over union ratio with it. prob i is the heatmap prediction value (probability prediction value) at the heatmap position corresponding to i. λ size λ, off and margin are hyperparameters set manually, and in this embodiment, they can be set to 0.1, 1, and 0.2 respectively.

[0137] Then, based on the above loss function, the target detection network ψ is trained using the new training set Φ', and the trained target detection network ψ' is obtained.

[0138] Step 4: Input the new test set T' into the trained target detection network ψ' to obtain the final detection result.

[0139] 41) Denote the total number of test slices in the new test set T' as M, and each test slice as T i ', where i = 1, 2... M, and record the horizontal offset and vertical offset ΔY i global of the center point of each test slice relative to the upper left corner point of the original large image.

[0140] 42) Send the test slices into the trained network ψ' for testing to obtain the initial test result Result i .

[0141] 43) Based on the sizes of all targets in the new training set Φ', the initial test result Result i is screened to eliminate false alarm detection boxes.

[0142] Specifically, count the maximum and minimum values of the length (width), area, and the maximum aspect ratio of all targets in the training set, and multiply the maximum value of the length (width) and area by 120%, the minimum value by 80%, and the maximum aspect ratio by 110% as the scale screening range to identify Result i , where i = 1, 2... M, and the detection boxes outside the above range are regarded as false alarms and eliminated.

[0143] 44) Offset the screened initial test result Result i to obtain the new test result Result global .

[0144] Specifically, add the corresponding horizontal offset i and vertical offset ΔY to the positions of all predicted targets in the result Result i global, obtain the positions of the predicted targets of each test slice in the original large image, and denote the set of all slice prediction results as the new result Result global 。

[0145] 45) Perform non-maximum suppression on the said new test result Result global to obtain the final test result Result nms 。

[0146] On the one hand, the present invention provides more helpful information for the network by introducing prior information such as SAR image complex information as the input of the network. The amount of information is richer than that of a single CFAR prior information, solving the problem of insufficient information. On the other hand, the CenterNet object detection algorithm is used, which does not require laying a large number of candidate boxes on the original image, and to a certain extent solves the problems of high complexity of the prior art and the appearance of low-quality detection boxes;

[0147] In addition, the present invention uses a coordinated adversarial loss function to train the network and adds a feature fusion module to the network, obtaining a certain degree of improvement in indicators such as precision and recall compared with the original method.

[0148] To further verify the deep network SAR object detection method based on complex image information proposed by the present invention, this embodiment also detects it on the MiniSAR data image.

[0149] Please refer to Figures 10 - 15 , Figures 10 - 15 which is the MiniSAR data image adopted in the embodiment of the present invention; Table 1 below gives the detection results of the method of the present invention on the Figures 10 - 15 shown MiniSAR data image, and compares it with the existing detection methods CFAR-Guided-EfficientDet SAR image object detection method (abbreviation: CFAR-Guided-EfficientDet, from the doctoral thesis of Xidian University, Wang Ning, 2021, "Research on SAR Image Object Detection and Identification Methods") and the CFAR-guided dual-stream SSD SAR image object detection method (abbreviation: ICSAF-CFAR-SSD, from the master's thesis of Xidian University, Tang Tiangu, 2022, "Research on SAR Object Detection Combining Constant False Alarm Detection and Deep Network"). In Table 1, Pre represents precision, that is, the percentage of real targets in the detected target boxes; Rec represents recall, that is, the percentage of targets correctly detected; F1-score represents the harmonic mean, and AP represents the average precision rate. They are systematic indicators that unify precision Pre and recall Rec; Images 1-Image 6 in Table 1 correspond to Figures 10 - 15The SAR image shown. It can be seen from the results that the method proposed by the present invention has a certain degree of improvement compared with the existing methods in most detection indexes.

[0150] Table 1 Comparison of detection performance between the method of the present invention and some existing methods

[0151]

[0152] It can be known from the above experiments that the method proposed by the present invention shows good performance in the vehicle target detection task and has a certain effectiveness and feasibility.

[0153] The above content is a further detailed description of the present invention in combination with specific preferred embodiments. It cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention belongs, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, which should all be regarded as belonging to the protection scope of the present invention.

Claims

1. A deep network SAR target detection method based on complex image information, characterized in that, Including: Step 1: Extract amplitude information and complex information from the given training set Φ and test set T respectively to obtain a new training set Φ' and a new test set T'; Specifically including: 11) Perform modulo operations on the original data of the training set Φ and the test set T respectively, and correspondingly obtain the first-dimensional statistic graphs Φ D1 , T D1 ; 12) Calculate the real signal kurtosis RSK and the non-circularity measure Ψ for the training set Φ and the test set T respectively 0 to obtain the second-dimensional statistic graph Φ D2 and T D2 and the third-dimensional statistic graph Φ D3 and T D3 ; 13) Sub-aperture decomposition is performed on the original data of the training set Φ and the test set T respectively, and the real correlation coefficient ρ is calculated using the obtained sub-band images I and the complex correlation coefficient ρ C , and the fourth-dimensional statistic maps Φ D4 , T D4 and the fifth-dimensional statistic maps Φ D5 , T D5 ; 14) Gaussian downsampling, central surround, cross-scale addition, and normalization operations are sequentially performed on the original data of the training set Φ and the test set T, respectively, and the sixth-dimensional statistic map Φ D6 and T D6 ; 15) Respectively splice the statistical quantity diagrams Φ in the first to sixth dimensions D1 -Φ D6 、T D1 -T D6 along the channel dimension, and correspondingly obtain 6-channel images Φ 6D and T 6D ; 16) Perform sliding window slicing operations on the 6-channel image Φ 6D and T 6D respectively, to obtain a number of slices correspondingly, denoted as the new training set Φ' and the new test set T'; Step 2: Construct a SAR target detection network ψ based on the CenterNet object detection algorithm; wherein, the target detection network ψ includes an attention mechanism module, a feature extraction and fusion module, and a predictor module; wherein, the attention mechanism module includes a convolutional neural network A for performing preliminary feature processing on the input image to facilitate better feature extraction by the subsequent feature extraction network; the feature extraction and fusion module includes a convolutional neural network B; the predictor module is used to predict the detection result, including three independent convolutional neural networks, and each of the convolutional neural networks includes two convolutional layers; Step 3: Construct a coordinated adversarial loss function and use the new training set Φ' to train the target detection network ψ to obtain a trained target detection network ψ'; Step 4: Input the new test set T' into the trained target detection network ψ' to obtain the final detection result.

2. The deep network SAR target detection method based on complex image information according to claim 1, wherein Step 12) includes: 12a) Let the original complex data of any image in the training set Φ and the test set T be G0, and traverse G0 according to a sliding window of a certain size; 12b) Calculate the real signal kurtosis (RSK) of all pixel points within the window for each position of the sliding window, and obtain the second-dimensional statistic map Φ with the same scale as the original image. D2 , T D2 ; The calculation formula for the real signal kurtosis (RSK) is: Where N is the total number of pixel points within the sliding window, is the intensity value of the real part of each pixel point within the sliding window, is the average value of the real part of the pixel points within the sliding window, is the intensity value of the imaginary part of each pixel point within the sliding window, is the average value of the imaginary part of the pixel points within the sliding window; 12c) Calculate the non-circularity measure ψ of all pixel points within the window for each position of the sliding window 0 , obtaining the third-dimensional statistical quantity map Φ with the same scale as the original image D3 , T D3 ; The calculation formula for the non-circularity measure ψ 0 is as follows: where N is the total number of pixel points within the sliding window, represents the complex data formed by the real part intensity value and the imaginary part intensity value of each pixel point.

3. The deep network SAR target detection method based on complex image information according to claim 1, wherein Step 13) includes: 13a) Perform azimuthal Fourier transform on the original complex data G0 of any image in the training set Φ and the test set T to obtain an azimuthal spectrogram H1; 13b) Estimate the azimuth weighting function FW(·) using the azimuthal spectrogram H1, and multiply its reciprocal by the azimuthal spectrogram H1 to obtain a weighted spectrogram H2; 13c) Divide the weighted spectrogram H2 into several subbands of the same length in the azimuth direction, move the center frequency of each subband to the 0 frequency, and multiply each subband by the weighting function FW(·); 13d) Perform zero-padding on each weighted subband, and perform inverse azimuthal Fourier transform on each subband to obtain a number of sub-aperture images; 13e) Calculate the statistics ρ using the respective plurality of sub-aperture images I and ρ C to obtain the fourth-dimensional statistic map Φ D4 and T D4 and the fifth-dimensional statistic map Φ D5 and T D5 , and the calculation formula is: ρ C = log(1 - |R s |) Wherein, N represents the number of subbands, s i represents the complex-valued data of the pixel at the corresponding position of the i-th subband image, and <·> represents the spatial averaging operation.

4. The deep network SAR target detection method based on complex image information according to claim 1, characterized in that, Step 14) includes: 14a) Perform a modulus operation on the original complex data G0 of any image in the training set Φ and the test set T to obtain an intensity image G1; 14b) Continuously perform Gaussian filtering and downsampling on the intensity image G1 for m times. Each time of downsampling, the length and width become half of the original, obtaining an m + 1 - layer Gaussian pyramid P i , where i = 0, 1…m, with the 0 - scale being the largest and the m - scale being the smallest; 14c) For the m + 1-layer Gaussian pyramid obtained in step 14b), respectively select multiple combinations including several layers for cross-scale subtraction to obtain multiple center-surrounded images, and the formula is expressed as: I(c,s) = |P c -P s | where "-" represents cross-scale subtraction, and P c represents the large-scale layer, and P s represents the small-scale layer; 14d) Upsample the multiple center-surrounded images to the original image size by linear interpolation, and perform addition and normalization operations pixel by pixel to obtain the sixth-dimensional statistic map Φ D6 , T D6 .

5. The deep network SAR target detection method based on complex image information according to claim 1, characterized in that In step 2, the convolutional neural network A includes one convolutional layer, two fully connected layers, two pooling layers, two activation function layers, two splicing layers, one element addition layer, two element multiplication layers, and one batch normalization layer.

6. The deep network SAR target detection method based on complex image information according to claim 1, characterized in that In step 2, the convolutional neural network B includes three convolutional layers, one batch normalization layer, one activation function layer, one pooling layer, 4 convolutional residual modules, 4 upsampling feature fusion modules, 4 downsampling feature fusion modules, and three deconvolutional layers.

7. The deep network SAR target detection method based on complex image information according to claim 1, characterized in that, In step 3, constructing a coordinated adversarial loss function includes: Construct a total loss function based on heatmap loss, scale loss, and offset loss, and the expression of the total loss function is where N is the total number of true targets, represent the heatmap loss, scale loss, and offset loss respectively, is the coordination loss, is the adversarial loss, and the sum of the two is the coordination adversarial loss, which is calculated only when the true value G ab = 1; is calculated only when the true value G ab ≠ 1; IOU i is the IOU value of the intersection over union of the predicted box at this location and the true box with the largest intersection over union with it, prob i is the heatmap prediction value at the heatmap position corresponding to i, λ size 、λ off and margin are hyperparameters set artificially.

8. The deep network SAR target detection method based on complex image information according to claim 1, characterized in that, Step 4 includes: 41) Denote the total number of test slices in the new test set T' as M, and each test slice as T i ', where i = 1, 2... M, and record the horizontal offset and vertical offset ΔY of the center point of each test slice relative to the upper left corner point of the original large image i global ; 42) Send the test slice into the trained network ψ' for testing to obtain the initial test result Result i ; 43) Filter the initial test result Result based on the sizes of all the targets in the new training set Φ'. i to eliminate false alarm detection boxes; 44) Offset the initially screened test result Result i to obtain a new test result Result global ; 45) Perform non-maximum suppression on the new test result Result global to obtain the final test result Result nms .

Images