SAR Image Target Detection Method Based on Feature Separation Convolutional Neural Network

Abstract

The present invention discloses a SAR image target detection method based on a feature separation convolutional neural network. The method primarily addresses the problem that prior art techniques have difficulty accurately modeling background clutter in complex scene SAR images, and that false alarms and missed alarms are common when performing target detection on complex scene SAR images. The method comprises the following steps: generating a training set; constructing a feature separation convolutional neural network; training the feature separation convolutional neural network; detecting target positions in the image to be tested; and obtaining a large-scale SAR image after target detection. The method performs feature separation on an input SAR image, utilizing a feature separation module in the network to remove features in the SAR image that are detrimental to target detection. This method eliminates the need for pre-modeling of background clutter and significantly reduces false alarms and missed alarms when performing target detection on complex scene SAR images, thereby improving target detection accuracy.

Description

Technical Field

[0001] This invention belongs to the field of radar image technology, and more specifically, to a method for detecting targets in synthetic aperture radar (SAR) images based on a feature separation convolutional neural network, in the field of automatic target recognition from radar images. The invention can be used to detect targets of interest, such as ground vehicles, from SAR images. Background Art

[0002] Synthetic Aperture Radar (SAR) offers the advantage of providing remote sensing imagery in all weather conditions and at all times of the day, making it widely used in both military and civilian applications. With the rapid development of radar imaging technology, the field of SAR automatic target recognition has grown rapidly. SAR image target detection, as the primary stage of SAR automatic target recognition, has garnered widespread attention. Constant False Alarm Rate (CFAR) is the most widely used and most sophisticated traditional SAR target detection method. This method leverages background information by modeling the statistical distribution of background clutter, deriving an adaptive threshold. Detection results are then compared against the pixel grayscale value using a sliding window. Therefore, determining an appropriate clutter statistical model is crucial for ensuring CFAR detection performance. However, due to the large amount of complex background clutter in field-measured SAR images, selecting a suitable clutter statistical model is difficult, resulting in reduced detection performance. With the advancement of deep learning, numerous convolutional neural network-based methods have been proposed. These methods, thanks to the large amount of labeled training data available for the network to learn from, have achieved significant progress in target detection. However, due to the complex scenes in field-measured SAR images, CNN-based SAR target detection methods still suffer from a high number of false alarms and missed detections, necessitating an urgent need to improve detection accuracy.

[0003] Beijing Institute of Technology has proposed a method for detecting SAR images with a constant false alarm rate (CFAR) based on segmentation and combination in its patent application, "Adaptive CFAR target detection method for SAR images based on segmentation and combination" (patent application number: 201010292987.0, application publication number: CN 101975940 A). The method first utilizes a reference window strategy to obtain parameters for estimating a background clutter model, then utilizes the relationship between the false alarm probability and the clutter model to obtain a detection threshold for the detector, and finally compares the grayscale value of the pixel with a certain threshold through a sliding window to obtain a detection result. The method has good performance in SAR images of some simple scenes. However, the method still has the following disadvantages: since the method requires establishing a background clutter model for the SAR image, the establishment of the model requires very strong professional knowledge and experience, and is difficult to establish. When encountering SAR images with complex scenes, it is difficult to establish a suitable background clutter model, which results in low detection accuracy.

[0004] In its patent application, "A Method for Detecting Tower Targets in SAR Images Based on Deep Learning" (patent application number: 201811100702.1, publication number: CN 109325947 A), the Beijing University of Aeronautics and Astronautics discloses a method for detecting targets in SAR images. This method designs a target detection network based on a convolutional neural network. Labeled training SAR images are then used to train the target detection network. After training converges, the trained model is used to test a test SAR image to obtain detection results for the test SAR image. This method leverages the feature extraction and nonlinear mapping capabilities of convolutional neural networks, achieving good performance in scenes with significant differences between the target and background. However, a drawback of this method is that it processes the entire scene indiscriminately during feature extraction. This can result in a high number of false alarms and missed alarms when the target and background are not clearly differentiated and the scene contains a large amount of complex background clutter, leading to a decrease in detection accuracy. Summary of the Invention

[0005] The present invention addresses the shortcomings of the prior art by proposing a method for SAR image target detection based on a feature separation convolutional neural network. This method aims to address the difficulty in establishing a suitable background clutter model for complex scene SAR images, as well as the high false alarm and missed alarm rates and low detection accuracy associated with complex scene SAR images.

[0006] The technical approach to achieving the objectives of the present invention is to construct a feature separation convolutional neural network and utilize the trained feature separation convolutional neural network to directly perform target detection on SAR images, thereby avoiding the problem of prior art requiring the prior empirical modeling of background clutter. The present invention uses the trained network to perform feature separation on the input SAR image, and utilizes the feature separation module within the network to remove features contained in the SAR image that are detrimental to target detection, including complex background clutter that can easily cause false alarms. Ultimately, target detection is performed on the SAR image using detection features that significantly reduce missed alarms. This solves the problem of prior art techniques for target detection in complex scene SAR images, which results in a high number of false alarms and missed alarms, as well as low detection accuracy.

[0007] The specific steps of the present invention are as follows:

[0008] Step 1: Generate training set:

[0009] Step 1.1: Collect at least 7 large-scale synthetic aperture radar (SAR) images;

[0010] In step 1.2, each large SAR image is cropped into multiple sub-images of 300 × 300 pixels each using a window operation with a size of 300 × 300, and each sub-image is labeled.

[0011] Step 1.3, all labeled sub-images are combined into a training set;

[0012] Step 2: Construct a feature separation convolutional neural network:

[0013] Step 2.1: Build a 10-layer backbone module with the following structure: first convolutional layer, second convolutional layer, first pooling layer, third convolutional layer, fourth convolutional layer, second pooling layer, fifth convolutional layer, sixth convolutional layer, seventh convolutional layer, third pooling layer; set the convolution kernel size of the first to seventh convolutional layers to 3×3; set the number of convolution kernels of the first to seventh convolutional layers to 64, 64, 128, 128, 256, 256, 256 respectively; set the pooling area size of the first to third pooling layers to 2×2;

[0014] Step 2.2: Build a feature separation module consisting of two branches connected in parallel, followed by a summation layer, and then connected in series with a decoder. The two branches have the same structure but different parameters, and are responsible for detection feature extraction and unfavorable feature extraction respectively. The two branches are composed of the first convolutional layer, the second convolutional layer, and the third convolutional layer in series; the convolution kernel size of the first to third convolutional layers is set to 3×3; the number of convolution kernels of the first to third convolutional layers is set to 512; the decoder is composed of the first deconvolution layer, the second deconvolution layer, the third deconvolution layer, the fourth deconvolution layer, and the fifth deconvolution layer in series; the convolution kernel size of the first to fifth deconvolution layers is set to 3×3; the number of convolution kernels of the first to fifth deconvolution layers is set to 512, 256, 128, 64, and 3, respectively; the output of the feature separation module is the output of the detection feature extraction branch;

[0015] Step 2.3, build a 15-layer deep semantic feature extraction module, whose structure is as follows: the first pooling layer, the first convolutional layer, the second convolutional layer, the third convolutional layer, the second pooling layer, the fourth convolutional layer, the fifth convolutional layer, the sixth convolutional layer, the seventh convolutional layer, the eighth convolutional layer, the ninth convolutional layer, the tenth convolutional layer, the eleventh convolutional layer, the twelfth convolutional layer, and the thirteenth convolutional layer; set the convolution kernel size of the first to fourth convolutional layers to 3×3; set the convolution kernel size of the fifth to thirteenth convolutional layers to The kernel sizes are set to 1×1, 1×1, 3×3, 1×1, 3×3, 1×1, 3×3, 1×1, 3×3 respectively; the number of convolution kernels of the first to third convolution layers is set to 512; the number of convolution kernels of the fourth to thirteenth convolution layers is set to 1024, 1024, 256, 512, 128, 256, 128, 256, 128, 256 respectively; the pooling area sizes of the first to second pooling layers are set to 2×2 and 3×3 respectively;

[0016] Step 2.4: Connect the backbone module, the first output end of the feature separation module, and the deep semantic feature extraction module in series to form a feature extraction branch. Connect the second output end of the feature separation module in series with the first detection head to form the first object detection branch. Connect the outputs of the fifth, seventh, ninth, eleventh, and thirteenth convolutional layers in the deep semantic feature extraction module in series with the five detection heads to form the second to sixth detection branches, respectively. Each detection head contains two parallel convolutional layers with a convolution kernel size of 3×3. Connect the six object detection branches in parallel and then connect them to the feature extraction branch to form a feature separation convolutional neural network.

[0017] Step 3: Train the feature separation convolutional neural network:

[0018] The training set is input into the feature separation convolutional neural network, and the stochastic gradient descent algorithm is used to iteratively update the network weight value and optimize the total loss function of the network until it converges to obtain the trained feature separation convolutional neural network.

[0019] Step 4: Detect the target position in the image to be tested:

[0020] Step 4.1: Using a window operation with a size of 300 × 300, crop each large SAR image to be tested into multiple sub-images with each pixel being 300 × 300.

[0021] Step 4.2: Input each sub-image to be tested into the trained feature separation convolutional neural network, and output the target box position and target box category of the sub-image to be tested;

[0022] Step 5: Get the large SAR image after target detection:

[0023] According to the order of windowing, the target frame position of each sub-image is mapped to the corresponding position of each large-scale SAR image to be tested. The target is framed in the large-scale SAR image according to the mapped target frame position to obtain the large-scale SAR image after target detection.

[0024] Compared with the existing technology, the present invention has the following advantages:

[0025] First, the present invention uses the generated training set to train a feature separation convolutional neural network, and directly performs target detection on SAR images through the trained feature separation convolutional neural network. This makes it unnecessary for the present invention to model complex background clutter based on experience in advance, thereby avoiding the decrease in detection accuracy caused by the difficulty in establishing a suitable clutter statistical model for complex background clutter, and improving the detection accuracy of SAR targets.

[0026] Second, the present invention performs feature separation on the input SAR image through a trained network, and uses the feature separation module in the network to remove features contained in the SAR image that are unfavorable for target detection, including complex background clutter that easily causes false alarms. Finally, the SAR image is used to detect targets using detection features that can significantly reduce missed alarms. This allows the present invention to significantly reduce false alarms and missed alarms when performing target detection on SAR images of complex scenes, thereby improving target detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a flow chart of the present invention;

[0028] Figure 2 It is a structural diagram of the backbone module of the present invention;

[0029] Figure 3 It is a structural diagram of the feature separation module of the present invention;

[0030] Figure 4 It is a schematic diagram of the structure of the deep semantic feature extraction module of the present invention;

[0031] Figure 5 is a simulation diagram of the present invention, wherein, Figure 5 (a) is the detection result of the existing SSD technology on the first test original large image of the miniSAR dataset. Figure 5 (b) is the detection result of the second test original large image of the miniSAR dataset using the existing SSD technology. Figure 5 (c) is the detection result of the first test original large image of the miniSAR data set of the present invention, Figure 5 (d) is the detection result diagram of the second test original large image of the miniSAR data set of the present invention. DETAILED DESCRIPTION

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0033] Reference Figure 1 The specific steps for implementing the present invention are described in detail in the following embodiments.

[0034] Step 1: Generate training and test sets.

[0035] The embodiment of the present invention collects 7 large-scale synthetic aperture radar (SAR) images from a measured MiniSAR data set.

[0036] Using a window operation with a window size of 300×300, each large SAR image is cropped into sub-images with each pixel being 300×300, resulting in a total of 110 sub-images. Each sub-image is marked according to the target contained in it.

[0037] All labeled sub-images form the training set.

[0038] The remaining two large-scale synthetic aperture radar (SAR) images in the measured MiniSAR dataset are selected.

[0039] Using a window operation with a window size of 300×300, each large SAR image is cropped into sub-images with a pixel size of 300×300, and a total of 33 sub-images are obtained.

[0040] All sub-images are combined into a test set.

[0041] Step 2: Construct a feature separation convolutional neural network.

[0042] Build a 10-layer backbone module with the following structure: first convolution layer, second convolution layer, first pooling layer, third convolution layer, fourth convolution layer, second pooling layer, fifth convolution layer, sixth convolution layer, seventh convolution layer, third pooling layer; set the convolution kernel size of the first to seventh convolution layers to 3×3; set the number of convolution kernels of the first to seventh convolution layers to 64, 64, 128, 128, 256, 256, 256 respectively; set the pooling area size of the first to third pooling layers to 2×2.

[0043] Reference Figure 2 , further describes the backbone modules constructed in the embodiments of the present invention.

[0044] The input of the backbone module is a 300×300 pixel sub-image obtained by cropping a large SAR image using a windowing operation. The first to seventh convolutional layers extract features from the input SAR sub-image layer by layer, and the first to third pooling layers perform dimensionality reduction on the extracted features to reduce the number of parameters in the backbone module and prevent overfitting.

[0045] In an embodiment of the present invention, a feature separation module is constructed, which is composed of two branches connected in parallel, then passing through a summation layer, and then connected in series with a decoder. The two branches have the same structure but different parameters, thereby respectively undertaking detection feature extraction and unfavorable feature extraction. The two branches are composed of a first convolutional layer, a second convolutional layer, and a third convolutional layer in series; the convolution kernel sizes of the first to third convolutional layers are all set to 3×3; the number of convolution kernels of the first to third convolutional layers is set to 512; the decoder is composed of a first deconvolution layer, a second deconvolution layer, a third deconvolution layer, a fourth deconvolution layer, and a fifth deconvolution layer in series; the convolution kernel sizes of the first to fifth deconvolution layers are all set to 3×3; the number of convolution kernels of the first to fifth deconvolution layers is set to 512, 256, 128, 64, and 3, respectively; the output of the feature separation module is the output of the detection feature extraction branch.

[0046] Reference Figure 3 The feature separation module constructed in the embodiment of the present invention is further described.

[0047] The input to the feature separation module is the feature map output by the backbone module. This map is fed into two parallel branches within the feature separation module. Three convolutional layers, each containing three branches, extract features from the backbone module's output feature map. Due to the different parameters of these two branches, they are responsible for extracting detection features and unfavorable features, respectively. The detection features and unfavorable features output by each branch are fed into a summation layer, where they are summed. The output feature map of the summation layer is fed into the subsequent decoder, where it is reconstructed through the first through fifth deconvolutional layers to produce a SAR sub-image. The output feature map of the detection feature extraction branch serves as the final output of the entire feature separation module.

[0048] Build a 15-layer deep semantic feature extraction module, whose structure is as follows: the first pooling layer, the first convolutional layer, the second convolutional layer, the third convolutional layer, the second pooling layer, the fourth convolutional layer, the fifth convolutional layer, the sixth convolutional layer, the seventh convolutional layer, the eighth convolutional layer, the ninth convolutional layer, the tenth convolutional layer, the eleventh convolutional layer, the twelfth convolutional layer, and the thirteenth convolutional layer; set the convolution kernel size of the first to fourth convolutional layers to 3×3; set the convolution kernel size of the fifth to thirteenth convolutional layers to The sizes of the convolution kernels of the first to third convolution layers are set to 512 respectively; the numbers of the convolution kernels of the fourth to thirteenth convolution layers are set to 1024, 1024, 256, 512, 128, 256, 128, 256, 128, 256 respectively; the pooling area sizes of the first to second pooling layers are set to 2×2 and 3×3 respectively.

[0049] Reference Figure 4 The deep semantic feature extraction module constructed in the embodiment of the present invention is further described.

[0050] The deep semantic feature extraction module takes as input the output feature map of the feature separation module, i.e., the detection feature map. This feature map is then subjected to dimensionality reduction through the first and second pooling layers, reducing the number of parameters in the module and preventing overfitting. The first through thirteenth convolutional layers extract increasingly abstract semantic information from the input detection feature map, preparing the final multi-scale object detection.

[0051] The backbone module, the first output end of the feature separation module, and the deep semantic feature extraction module are connected in series in sequence to form a feature extraction branch. The second output end of the feature separation module is connected in series with the first detection head to form the first target detection branch. The outputs of the fifth, seventh, ninth, eleventh, and thirteenth convolutional layers in the deep semantic feature extraction module are connected in series with the five detection heads to form the second to sixth detection branches. Each detection head contains two parallel convolutional layers, and the convolution kernel size is 3×3. The six target detection branches are connected in parallel and then connected to the feature extraction branch to form a feature separation convolutional neural network.

[0052] Step 3: Train the feature separation convolutional neural network.

[0053] The training set is input into the feature separation convolutional neural network, and the stochastic gradient descent algorithm is used to iteratively update the weight value of the network, and the total loss function of the network is optimized until it converges to obtain a trained feature separation convolutional neural network.

[0054] The total loss function is as follows:

[0055]

[0056] in, represents the total loss function of the feature separation convolutional neural network, Represents the position loss between the target box output by the feature separation convolutional neural network and the marked real box, Represents the category loss of the target box output by the feature separation convolutional neural network, It represents the orthogonal loss between the detection features and unfavorable features extracted by the feature separation module in the feature separation convolutional neural network. Represents the reconstruction loss between the reconstructed image output by the decoder in the feature separation module of the feature separation convolutional neural network and the original image input to the network. α represents the weight of the orthogonal loss function, and β represents the weight of the reconstruction loss function. Both weight values ​​are based on the dimensions of each loss function and are set in the range of α∈[0,1] and β∈[0,1], and α≠β.

[0057] The position loss is calculated by the following formula:

[0058]

[0059]

[0060] Among them, I represents the total number of target frames output by the network, Σ represents the sum operation, i represents the serial number of the target frame output by the network, J represents the total number of manually labeled real frames, j represents the serial number of the manually labeled real frames, and x ij Indicates the state where the i-th target box output by the network matches the j-th real box. If xij A value of 0 indicates no match, 1 indicates a match, i represents the i-th target box output by the network, g j represents the jth ground-truth box manually labeled;

[0061] The class loss is:

[0062]

[0063] Among them, z i Indicates the true category label corresponding to the i-th target box output by the network, log represents the logarithmic operation with base 2, and s i Represents the network's predicted category probability for the i-th target box;

[0064] The orthogonal loss is:

[0065]

[0066] Where N is the total number of samples in the training set, n is the sequence number of the sample in the training set, and f n1 It represents the vector of all feature maps output by the nth sample in the training set after the detection feature extraction branch of the feature separation module, f n2 It represents the vector of all feature maps output by the nth sample in the training set after passing through the unfavorable feature extraction branch of the feature separation module, and T represents the transpose of the vector;

[0067] The reconstruction loss is:

[0068]

[0069] Among them, ||·||2 represents the 2-norm operation, x n represents the nth sample in the training set, x n Represents the reconstructed image output by the decoder in the feature separation module after the nth sample in the training set passes through.

[0070] Step 4: Detect the target position in the SAR image.

[0071] Each test sub-image in the test set is input into the trained feature separation convolutional neural network, and the target box position and target box category of the test sub-image are output.

[0072] Step 5: Obtain a large SAR image after target detection.

[0073] According to the order of windowing, the target frame position of each sub-image is mapped to the corresponding position of each large-scale SAR image in the test set. The target is framed in the large-scale SAR image according to the mapped target frame position to obtain the large-scale SAR image after target detection.

[0074] The effects of the present invention are further described below in conjunction with simulation experiments.

[0075] 1. Simulation experiment conditions:

[0076] The hardware platform of the simulation experiment of the present invention is: the processor is Intel Xeon Silver 4114CPU, the processor main frequency is 2.20GHz, the memory is 128GB, and the graphics card is NVIDIA GTX 2080Ti.

[0077] The software platform for the simulation experiment of the present invention is: Ubuntu 16.04LTS operating system, Pytorch, Python 2.6.

[0078] The data set used in the simulation experiment of the present invention is the MiniSAR measured data set, which is the measured SAR image data based on complex scenes released by Sandia National Laboratory in the United States in 2006. The MiniSAR data image contains vehicle targets and complex backgrounds, such as buildings, grass, trees, etc. In this experiment, the vehicle targets are used as detection targets. The data set contains 9 original large images, the image size is 1638×2510 pixels, and the image format is png. In this experiment, 7 of the original 9 SAR images are selected as training images, and the remaining 2 are selected as test images. The original training and test SAR images are cropped respectively to obtain sub-images with a size of 300×300, and finally 110 training sub-images and 33 test sub-images are obtained.

[0079] 2. Simulation experiment content and result analysis:

[0080] The simulation experiment of the present invention uses the present invention and a prior art SSD to perform target detection on the test images in the input MiniSAR data set, and obtains the detection results as follows: Figure 5 shown.

[0081] In the simulation experiment, the prior art SSD refers to the target detection model proposed by W. Liu et al. in “SSD: Single Shot MultiBox Detector”.

[0082] The following combination Figure 5 The simulation diagram of the present invention is further described.

[0083] Figure 5 The detection results of the present invention and the prior art on two original large test images in the miniSAR data set are shown in FIG. Figure 5The green rectangular box indicates the correct detection result, the red rectangular box indicates the wrong detection result, and the blue rectangular box indicates the missed vehicle target. Figure 5 (a) is the detection result of the existing SSD technology on the first test original large image of the miniSAR dataset. Figure 5 (b) is the detection result of the second test original large image of the miniSAR dataset using the existing SSD technology. Figure 5 (c) is the detection result of the first test original large image of the miniSAR data set of the present invention, Figure 5 (d) is the detection result diagram of the second test original large image of the miniSAR data set of the present invention.

[0084] Depend on Figure 5 (a) and Figure 5 (b) It can be seen that the detection result of the prior art SSD contains a large number of red rectangles, i.e., incorrect detection results (false alarms), and a large number of blue rectangles, i.e., missed vehicle targets (missed alarms). Compared with this, Figure 5 (c) and Figure 5 The number of red and blue rectangles in (d) is greatly reduced, that is, the false alarms and missed alarms are greatly reduced. Figure 5 (c) There are only two missed detections in the upper left corner. This is because the two vehicles in the upper left corner have only a small part of their bodies in the SAR image, making them difficult to detect accurately. In addition, Figure 5 (c) There are only three false alarms at the building in the center of the image. This is because some structures of the building and the vehicle target show very similar scattering results, which is very easy to cause false alarms. Figure 5 (d) There are only a few missed detections in the vehicle area in the upper right corner. This is because the vehicle targets in this area are densely packed and the scattering intensity of some vehicle targets is low, which makes detection difficult. Figure 5 There are only a few false alarms in (d). These false alarms are caused by objects with scattering results that are very similar to vehicle targets, making them difficult to distinguish correctly during detection.

[0085] contrast Figure 5 (a) and Figure 5 (c) Figure 5 (b) and Figure 5 As can be seen from the detection result diagram (d), the method of the present invention can effectively reduce the number of false alarms and missed alarms in target detection, thereby improving the accuracy of SAR target detection.

[0086] In order to verify the simulation effect of the present invention, the detection results of the two methods are evaluated as follows using the F1-score formula, and all the calculation results are plotted in Table 1:

[0087]

[0088]

[0089]

[0090] Table 1. Quantitative analysis of the test results of the present invention and the prior art in the simulation experiment

[0091]

[0092] From Table 1, it can be seen that the F1-score of the present invention is improved by 6.07% compared with the existing technology SSD, which shows that the present invention has better detection performance than the existing technology.

[0093] The above simulation experiments demonstrate that the present invention proposes a method for SAR image target detection based on a feature separation convolutional neural network. This method eliminates the need for empirical modeling of complex background clutter, thus avoiding the problem of decreased detection accuracy caused by the difficulty in establishing a suitable clutter statistical model for complex background clutter. The present invention performs feature separation on the input SAR image using a trained network, utilizes a feature separation module within the network to remove features in the SAR image that are detrimental to target detection, and ultimately uses detection features that significantly reduce missed detections to perform target detection on the SAR image. This method significantly reduces false alarms and missed detections when performing target detection on SAR images of complex scenes, thereby improving target detection accuracy and possessing significant practical application value.

Claims

1. A SAR image target detection method based on feature separation convolutional neural network, characterized in that: The backbone module is used to extract features of the SAR image, and the feature separation module is used to separate the features of the SAR image. The steps of this method include the following: Step 1: Generate training set: Step 1.1: Collect at least 7 large-scale synthetic aperture radar (SAR) images; In step 1.2, each large SAR image is cropped into multiple sub-images of 300 × 300 pixels each using a window operation with a size of 300 × 300, and each sub-image is labeled. Step 1.3, all labeled sub-images are combined into a training set; Step 2: Construct a feature separation convolutional neural network: Step 2.1: Build a 10-layer backbone module with the following structure: first convolutional layer, second convolutional layer, first pooling layer, third convolutional layer, fourth convolutional layer, second pooling layer, fifth convolutional layer, sixth convolutional layer, seventh convolutional layer, third pooling layer; set the convolution kernel size of the first to seventh convolutional layers to 3×3; set the number of convolution kernels of the first to seventh convolutional layers to 64, 64, 128, 128, 256, 256, 256 respectively; set the pooling area size of the first to third pooling layers to 2×2; In step 2.2, a feature separation module is constructed, which consists of two branches connected in parallel, followed by a summation layer and then connected in series with a decoder. The two branches have the same structure but different parameters, and are responsible for detecting feature extraction and unfavorable feature extraction, respectively. Both branches are composed of the first convolutional layer, the second convolutional layer, and the third convolutional layer in series. The convolution kernel size of the first to third convolutional layers is set to 3×3, and the number of convolution kernels in the first to third convolutional layers is set to 512. The decoder consists of the first deconvolution layer, the second deconvolution layer, the third deconvolution layer, the fourth deconvolution layer, and the fifth deconvolution layer connected in series. The convolution kernel size of the first to fifth deconvolution layers is set to 3×3. The number of convolution kernels of the first to fifth deconvolution layers is set to 512, 256, 128, 64, and 3 respectively. The output of the feature separation module is the output of the detection feature extraction branch. Step 2.3, build a 15-layer deep semantic feature extraction module, whose structure is as follows: the first pooling layer, the first convolutional layer, the second convolutional layer, the third convolutional layer, the second pooling layer, the fourth convolutional layer, the fifth convolutional layer, the sixth convolutional layer, the seventh convolutional layer, the eighth convolutional layer, the ninth convolutional layer, the tenth convolutional layer, the eleventh convolutional layer, the twelfth convolutional layer, and the thirteenth convolutional layer; set the convolution kernel size of the first to fourth convolutional layers to 3×3; set the convolution kernel size of the fifth to thirteenth convolutional layers to The kernel sizes are set to 1×1, 1×1, 3×3, 1×1, 3×3, 1×1, 3×3, 1×1, 3×3 respectively; the number of convolution kernels of the first to third convolution layers is set to 512; the number of convolution kernels of the fourth to thirteenth convolution layers is set to 1024, 1024, 256, 512, 128, 256, 128, 256, 128, 256 respectively; the pooling area sizes of the first to second pooling layers are set to 2×2 and 3×3 respectively; Step 2.4: Connect the backbone module, the first output end of the feature separation module, and the deep semantic feature extraction module in series to form a feature extraction branch. Connect the second output end of the feature separation module in series with the first detection head to form the first object detection branch. Connect the outputs of the fifth, seventh, ninth, eleventh, and thirteenth convolutional layers in the deep semantic feature extraction module in series with the five detection heads to form the second to sixth detection branches, respectively. Each detection head contains two parallel convolutional layers with a convolution kernel size of 3×3. Connect the six object detection branches in parallel and then connect them to the feature extraction branch to form a feature separation convolutional neural network. Step 3: Train the feature separation convolutional neural network: The training set is input into the feature separation convolutional neural network, and the stochastic gradient descent algorithm is used to iteratively update the network weight value and optimize the total loss function of the network until it converges to obtain the trained feature separation convolutional neural network. Step 4: Detect the target position in the image to be tested: Step 4.1: Using a window operation with a size of 300 × 300, crop each large SAR image to be tested into multiple sub-images with each pixel being 300 × 300. Step 4.2: Input each sub-image to be tested into the trained feature separation convolutional neural network, and output the target box position and target box category of the sub-image to be tested; Step 5: Get the large SAR image after target detection: According to the order of windowing, the target frame position of each sub-image is mapped to the corresponding position of each large-scale SAR image to be tested. The target is framed in the large-scale SAR image according to the mapped target frame position to obtain the large-scale SAR image after target detection.

2. The SAR image target detection method based on feature separation convolutional neural network according to claim 1, characterized in that: The total loss function of the network described in step 3 is as follows: in, represents the total loss function of the feature separation convolutional neural network, Represents the position loss between the target box output by the feature separation convolutional neural network and the marked real box, Represents the category loss of the target box output by the feature separation convolutional neural network, It represents the orthogonal loss between the detection features and unfavorable features extracted by the feature separation module in the feature separation convolutional neural network. Represents the reconstruction loss between the reconstructed image output by the decoder in the feature separation module of the feature separation convolutional neural network and the original image input to the network. α represents the weight of the orthogonal loss function, and β represents the weight of the reconstruction loss function. Both weight values ​​are based on the dimensions of each loss function and are set in the range of α∈[0,1] and β∈[0,1], and α≠β. The position loss is calculated by the following formula: Among them, I represents the total number of target frames output by the network, Σ represents the sum operation, i represents the serial number of the target frame output by the network, J represents the total number of manually labeled real frames, j represents the serial number of the manually labeled real frames, and x ij Indicates the state where the i-th target box output by the network matches the j-th real box. If x ij A value of 0 indicates no match, 1 indicates a match, i represents the i-th target box output by the network, g j represents the jth ground-truth box manually labeled; The class loss is: Among them, z i Indicates the true category label corresponding to the i-th target box output by the network, log represents the logarithmic operation with base 2, and s i Represents the network's predicted category probability for the i-th target box; The orthogonal loss is: Where N is the total number of samples in the training set, n is the sequence number of the sample in the training set, and f n1 It represents the vector of all feature maps output by the nth sample in the training set after the detection feature extraction branch of the feature separation module, f n2 It represents the vector of all feature maps output by the nth sample in the training set after passing through the unfavorable feature extraction branch of the feature separation module, and T represents the transpose of the vector; The reconstruction loss is: Among them, ||·||2 represents the 2-norm operation, x n represents the nth sample in the training set, x n Represents the reconstructed image output by the decoder in the feature separation module after the nth sample in the training set passes through.

Images