Directional decoupling target detection method, device, equipment, medium and product for SAR image

Abstract

The present application discloses a method, device, medium and product for SAR image directional decoupling target detection, which relates to the field of synthetic aperture radar target detection. The method includes: obtaining a source data set; the source data set includes: two high-resolution SAR image ship data; preprocessing the source data set; constructing a target detection model; the target detection model includes: a backbone network, an ST-MLP module, an MC-PM module, a neck network and a Gaussian distribution decoding prediction box; constructing a loss function; training the target detection model based on the preprocessed source data and the loss function; inputting the SAR image to be detected into the trained target detection model to obtain a target directional detection result. The above method of the present application can solve the problem that existing detectors are difficult to achieve accurate target rotation positioning due to blurred ship edge contours in SAR images, strong land scattering background interference and mismatched angle regression of ships of different sizes.

Description

Technical Field

[0001] The present application relates to the field of synthetic aperture radar target detection, and in particular to a method, device, equipment, medium and product for SAR image directional decoupling target detection. Background Art

[0002] Synthetic Aperture Radar (SAR) is a passive remote sensing technology that transmits microwave signals and records their reflected waves to obtain high-resolution images of the Earth's surface. Compared to optical remote sensing, SAR offers unique advantages: it enables all-weather and all-day monitoring, boasts high resolution, wide coverage, and the ability to penetrate clouds. These attributes have greatly promoted its application in ship detection.

[0003] Traditional SAR ship detection methods are typically based on image processing and pattern recognition techniques. Their core concept is to extract features from SAR images and identify ships using classifiers or detectors. First, SAR images are preprocessed, including denoising and radiometric correction. Then, features such as texture features and amplitude statistics are extracted to describe the appearance of ships in SAR images. Next, classifiers or detectors (such as Kalman filters and CFAR detectors) are used to identify or detect ships. Finally, the detection results are post-processed and analyzed to improve accuracy and robustness. These traditional methods have played an important role in SAR ship detection. However, traditional methods often require manual adjustment of thresholds or other parameters based on the specific application scenario. Furthermore, in complex scenes, where multiple different types of targets or background signals may exist, traditional methods often struggle to handle this diversity, leading to missed or false detections.

[0004] In recent years, with the development of deep learning, SAR ship target detection has made significant progress. Deep learning-based methods utilize deep neural networks to learn and classify maritime vessels. They can autonomously learn abstract feature representations within images, thereby improving the accuracy and robustness of ship detection. This approach not only identifies different types of ships but also distinguishes ships from other maritime targets or backgrounds, providing more reliable support for maritime surveillance and border security. However, current methods primarily use horizontal boxes for target detection, which can lead to significant redundancy when faced with densely distributed and randomly oriented targets in SAR imagery, thus compromising detection effectiveness.

[0005] Given the limitations of horizontal box object detection, in recent years, more and more research has begun to shift its focus to rotated box object detection. Rotated boxes can more accurately describe the shape and orientation of the target, improving the accuracy and robustness of target detection. For example, the RRPN network generates tilted proposals with text direction angle information, and then adjusts the angle information to bounding box regression so that the proposals are more accurately adapted to the text area in terms of direction. Another example is the R2CNN network, which first uses a region proposal network to generate axis-aligned bounding boxes, then extracts pooling features of different pooling sizes for each axis-aligned text box proposed by the RPN, and uses connected features to simultaneously predict text and non-text scores, axis-aligned boxes, and tilted minimum area boxes. Finally, the detection results are obtained through tilted non-maximum suppression. However, due to the high resolution and special imaging mechanism of SAR images, rotated box object detection faces three major challenges.

[0006] First, how to address the issue of inaccurate rotational positioning caused by blurred target edges in SAR images. Ship detection typically involves two main tasks: positioning and classification. Accurate ship position information is crucial for maritime surveillance and safety management. SAR technology provides high-resolution ocean surface imagery, which helps accurately locate ships at sea. Deep learning methods achieve precise positioning of ships by performing convolution operations on the edge information of ship targets in SAR images. However, because SAR uses microwave signals emitted by radar and echoes reflected from the target surface to create images, this imaging mechanism easily leads to blurred target outlines in the generated image, which in turn affects the effectiveness of target rotational positioning. Addressing this issue has become a key challenge in rotational detection.

[0007] Second, we must effectively address the impact of strong scattering background on land on rotation detection. In SAR images, elements unrelated to the target are considered noise. For example, the scattering intensity of land buildings is similar to that of ships, severely degrading detector performance and thus being considered noise. During rotation detection, since this noise closely resembles real targets, the detector is easily disturbed, resulting in missed detections or false alarms. Enabling the detector to accurately detect targets has become a crucial research issue.

[0008] Third, how to address the mismatch in angular regression for ships of varying sizes. Ships in SAR images often appear at varying sizes. When performing rotation detection, small-scale targets are more sensitive to positional offsets, while large-scale targets prioritize angular optimization. This poses a significant challenge to rotation detection. Because the model typically only uses a single angular regression loss, it often struggles to address targets of varying sizes. A key area of ​​focus is how to dynamically adjust the loss parameters for detectors of varying sizes. Summary of the Invention

[0009] The purpose of this application is to provide a method, device, equipment, medium and product for directional decoupling target detection in SAR images, which can solve the problem that existing detectors are difficult to achieve accurate target rotation positioning due to blurred ship edge contours in SAR images, strong land scattering background interference and angular regression mismatch of ships of different sizes.

[0010] To achieve the above objectives, this application provides the following solutions:

[0011] In a first aspect, the present application provides a method for directional decoupling target detection for SAR images, including:

[0012] Acquire a source data set; the source data set includes: two high-resolution SAR image ship data;

[0013] Preprocessing the source data set;

[0014] Constructing a target detection model; the target detection model includes: a backbone network, an ST-MLP module, an MC-PM module, a neck network, and a Gaussian distribution decoding prediction box;

[0015] Construct loss function;

[0016] Training the target detection model based on the preprocessed source data and the loss function;

[0017] The SAR image to be detected is input into the trained target detection model to obtain the target orientation detection result.

[0018] Optionally, preprocessing the source data set specifically includes the following steps:

[0019] The two high-resolution SAR image ship data were cropped to 640*640 size using linear interpolation method;

[0020] The cropped data set is divided into training set and validation set according to the ratio.

[0021] Optionally, the loss function is expressed as follows:

[0022] Loss all =L cls +L reg +L obj

[0023] Among them, L cls represents the classification loss, L reg represents the regression loss, L obj represents the confidence loss.

[0024] Optionally, inputting the SAR image to be detected into a trained target detection model to obtain the target orientation detection result specifically includes the following steps:

[0025] Inputting the SAR image to be detected into the backbone network to obtain the main features of the image;

[0026] Inputting the main features of the image into the ST-MLP module to extract and integrate complex features, and denoising the features after complex feature extraction and integration;

[0027] The denoised features are input into the MC-PM module for average pooling operation to obtain a multi-branch processing result;

[0028] Fusing the multi-branch processing results from a channel dimension;

[0029] Reduce the dimension of the features after channel dimension fusion;

[0030] Input the reduced dimension features into the horizontal and vertical hard-coded channels to obtain position feature information;

[0031] Perform channel conversion on the position feature information to obtain a final output result;

[0032] Inputting the final output result into the neck network for information fusion;

[0033] The information fused features are input into the Gaussian distribution decoding prediction box to obtain the target orientation detection result.

[0034] Optionally, the denoising of the features after complex feature extraction and integration is specifically performed using the following formula:

[0035]

[0036] Τ=x2-τ

[0037] G=max(Τ,0)

[0038]

[0039] Among them, τ is the threshold, a1 is the residual edge eigenvalue after global average pooling and Faltten operation, a5 is the eigenvalue after multi-layer perceptron feature extraction and integration, T is the difference between the eigenvalue after absolute operation and the threshold, x2 is the eigenvalue after absolute operation, G is the result of comparing T with 0, and returns T when T is greater than 0, and returns 0 when T is less than or equal to 0. x' is the result of multiplying the symbol value x of the input eigenvalue by G term by term, which represents the final eigenvalue after processing by the ST-MLP module.

[0040] Optionally, the denoised features are input into the MC-PM module for average pooling operation using the following formula:

[0041]

[0042] Among them, output is the eigenvalue after single-step average pooling processing, input is the output eigenvalue after ST-MLP module processing, stride is 1, indicating the step size, kernel indicates the pooling kernel, which are 5, 9, and 13 respectively, and padding is the integer part of the kernel value divided by 2.

[0043] In a second aspect, the present application provides a SAR image-oriented decoupling target detection device, the SAR image-oriented decoupling target detection device comprising:

[0044] The source data set acquisition module is used to acquire the source data set; the source data set includes: two high-resolution SAR image ship data;

[0045] A preprocessing module, configured to preprocess the source data set;

[0046] A target detection model construction module is used to construct a target detection model; the target detection model includes: a backbone network, an ST-MLP module, an MC-PM module, a neck network, and a Gaussian distribution decoding prediction box;

[0047] Loss function construction module, used to construct loss function;

[0048] A training module, configured to train the object detection model based on the preprocessed source data and the loss function;

[0049] The target orientation detection module is used to input the SAR image to be detected into the trained target detection model to obtain the target orientation detection result.

[0050] In a third aspect, the present application provides a computer device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any one of the above-mentioned methods for directional decoupling target detection for SAR images.

[0051] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the above-mentioned methods for directional decoupling target detection for SAR images.

[0052] In a fifth aspect, the present application provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of any one of the above-mentioned methods for directional decoupling target detection for SAR images.

[0053] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0054] The present application provides a method, apparatus, device, medium and product for directional decoupling target detection in SAR images. By combining the multi-layer perceptron (ST-MLP) based on the soft threshold function, the multi-layer perceptron is combined with the soft threshold function, which not only enhances the model's ability to extract and integrate complex features, but also significantly improves the denoising effect. The interference of background information on target positioning is reduced; by setting a multi-channel positioning module (MC-PM), not only can the target edge features be softened by the distribution average pooling operation, but also the horizontal and vertical orientation perception channels can be added to the network to enhance the network's direction perception ability, thereby generating a directional detection frame that is more in line with the target, by setting the Gaussian distribution decoding prediction box (GD-BPB). By using the Kullback-Leibler divergence as a distance metric, the effect of dynamic parameter adjustment can be achieved, thereby alleviating the angle regression mismatch problem of targets of different sizes. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0056] Figure 1 This is a diagram showing the application environment of a method for directional decoupling target detection in SAR images according to an embodiment of the present application;

[0057] Figure 2 A flowchart of a method for directional decoupling target detection in SAR images provided in one embodiment of the present application is provided;

[0058] Figure 3 This is a schematic diagram showing that the SAR image of a ship edge is blurred, resulting in the inability of rotation detection to accurately locate the correct rotation direction of the ship in one embodiment of the present application;

[0059] Figure 4 Schematic diagram of targets of different sizes in SAR images according to an embodiment of the present application, (a) is a small-sized ship target sample, (b) is a medium-sized ship target sample, (c) is a large-sized ship target sample, and (d) is an extra-large-sized ship target sample;

[0060] Figure 5 This is a schematic diagram of the target detection model structure according to an embodiment of the present application;

[0061] Figure 6 Schematic diagram of the ST-MLP module according to an embodiment of the present application, (a) is a schematic diagram of the ST-MLP structure, (b) is a schematic diagram of the multilayer perceptron in the ST-MLP, and (c) is a schematic diagram of the fully connected layer in the multilayer perceptron;

[0062] Figure 7 This is a schematic diagram of the MC-PM structure according to an embodiment of the present application;

[0063] Figure 8 Schematic diagram of the dataset sizes used in one embodiment of this application, (a) is the WR-HRSID dataset, (b) is the R-SSDD dataset;

[0064] Figure 9 Schematic diagram of the detection effect of an embodiment of the present application in two data sets, (a) (b) shows the directional detection performance on the R-SSDD data set, and (c) (d) shows the directional detection performance on the WR-HRSID data set;

[0065] Figure 10 A comparison chart of an embodiment of the present application and other advanced methods on the R-SSDD dataset;

[0066] Figure 11 A comparison chart of an embodiment of the present application with other advanced methods in a complex scenario of the R-SSDD dataset;

[0067] Figure 12 A comparison chart of an embodiment of the present application and other advanced methods on the WR-HRSID dataset;

[0068] Figure 13 A comparison of an embodiment of this application with other advanced methods in a complex scene of the WR-HRSID dataset

[0069] Figure 14 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0070] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0071] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0072] The SAR image directional decoupling target detection method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown, the terminal 102 communicates with the server 104 via a network. The data storage system can store data that the server 104 needs to process. The data storage system can be set up separately, integrated on the server 104, or placed on the cloud or other servers. The terminal 102 can send the SAR image data to be processed to the server 104. After the server 104 receives the SAR image data to be processed, the server 104 obtains a source data set for the SAR image data to be processed; the source data set includes: two high-resolution SAR image ship data; preprocesses the source data set; constructs a target detection model; the target detection model includes: a backbone network, an ST-MLP module, an MC-PM module, a neck network, and a Gaussian distribution decoding prediction box; constructs a loss function; trains the target detection model based on the preprocessed source data and the loss function; inputs the SAR image to be detected into the trained target detection model to obtain a target orientation detection result. The server 104 can feedback the obtained target orientation detection result to the terminal 102. In addition, in some embodiments, the SAR image orientation-decoupled target detection can also be implemented independently by the server 104 or the terminal 102. For example, the terminal 102 can directly perform target orientation detection on the SAR image data to be processed, or the server 104 can obtain the SAR image data to be processed from the data storage system and perform target orientation detection on the SAR image data to be processed.

[0073] Terminal 102 may include, but is not limited to, various desktop computers, laptops, smartphones, tablet computers, IoT devices, and portable wearable devices. IoT devices may include smart speakers, smart TVs, smart air conditioners, and smart car devices. Portable wearable devices may include smart watches, smart bracelets, and head-mounted devices. Server 104 may be implemented as a standalone server or a server cluster consisting of multiple servers, or may be a cloud server.

[0074] In an exemplary embodiment, Figure 2 As shown, a method for directional decoupling target detection for SAR images is provided. The method is executed by a computer device, specifically a computer device such as a terminal or a server, or a terminal and a server. In the embodiment of the present application, the method is applied to Figure 1The server 104 in the example is used as an example to illustrate the method, including the following steps 1 to 6. Among them:

[0075] Step 1: Obtain source data sets; the source data sets include: two high-resolution SAR image ship data.

[0076] Step 2: Preprocess the source data set.

[0077] The specific steps include:

[0078] S2.1. The two source datasets used in the experiment are the R-SSDD dataset and the WR-HRSID dataset. The R-SSDD dataset includes ships of various sizes (small, medium and large) in various complex background environments. The dataset contains 1,160 SAR images with a resolution of 1-15 meters, including 928 offshore scenes and 232 nearshore scenes. These images contain a total of 2,456 ship targets, with an average of 2.12 ships per image; the WR-HRSID dataset is derived from the HRSID dataset. It contains 136 large-scene SAR images, 5,604 scene slice images, and a total of 16,951 ships, but these are horizontally and vertically labeled and are not suitable for the research content of this application. Therefore, this application re-labeled the dataset; afterwards, a total of 1,704 samples were selected as the dataset of this application, including 1,304 training samples and 400 test samples, named WR-HRSID. This dataset contains large, medium and small ship targets of varying sizes (see Figure 4 ), the dataset size distribution diagram is as follows Figure 8 As shown, Figure 8 Part a is the WR-HRSID dataset, and part b is the R-SSDD dataset.

[0079] S2.2. Cropping the two datasets to 640*640 pixels using linear interpolation. The cropped datasets were then divided into two groups, with a training set and a validation set ratio of 9:1. The number of training sets in the two datasets was 928 and 1304, respectively, and the number of validation sets was 232 and 400, respectively.

[0080] Step 3: Construct a target detection model; the target detection model includes: a backbone network, an ST-MLP module, an MC-PM module, a neck network, and a Gaussian distribution decoding prediction box.

[0081] Figure 5This is a schematic diagram of the target detection model in this application. First, a backbone network is constructed based on the special imaging mechanism of SAR images to extract the main features of the image. Since SAR images usually have interference from strong scattering background from the land, the detector is prone to missed reports or false alarms. The features extracted by the backbone network are passed into a module ST-MLP that can capture and integrate complex features of the image while suppressing noise. The complex features of the image are extracted and integrated through a multi-layer perceptron. The integrated feature information is then combined with the feature information extracted by the original backbone network by point multiplication. After that, it is passed to the denoising layer for comparison operation to obtain the feature information of the denoised image. After a convolution, normalization and activation function operation, the denoised feature map is obtained. Secondly, the edges of ship targets in SAR images are often blurred, making it difficult for the detector to achieve accurate rotational positioning (see Figure 3 ) The SAR image denoising feature map is passed into a module MC-PM that can soften target edge information and enhance the network's directional perception ability. The target edge features are softened through distributed pooling, and the results of distributed pooling are then fused from the channel dimension. A horizontal and vertical perception channel are then added to give the model directional perception capabilities. The number of model channels is then reduced through convolution, normalization, and activation function operations. Finally, the processed feature map is passed into a feature pyramid structure for feature fusion, fusing the network's shallow position information with deep semantic information to obtain the final processed feature map. Finally, a Gaussian distribution decoding prediction box is designed, and a distance measurement method with dynamically adjustable parameters is used to design a regression loss function to address the problem of angle regression mismatch of ships of different sizes. The losses of each part are then fused at a certain ratio to obtain the complete loss function of the detection head. The detection head is used to dynamically adjust the processed feature map to obtain the final rotation detection result.

[0082] Step 4: Construct the loss function.

[0083] Step 5: Train the target detection model based on the preprocessed source data and loss function.

[0084] Step 6: Input the SAR image to be detected into the trained target detection model to obtain the target orientation detection result.

[0085] The specific steps are as follows:

[0086] In step 3, a complex feature integration and noise suppression module ST-MLP, a module MC-PM to enhance the model's direction perception ability, and a Gaussian distribution decoding prediction box are constructed. The specific steps are as follows:

[0087] S6.1. Feature extraction: The backbone network extracts and characterizes features from the original image, which directly affects the detection performance of the model and is the key to the target detection task. The ResNet network can solve the gradient vanishing and gradient exploding problems in deep network training by introducing residual connections, which plays an important role in extracting deep feature information of SAR images. Therefore, this application selects ResNet50 as the backbone network, which downsamples the input data at intensity levels of 4, 8, 16, and 32 to extract the main features of the image. The input feature layer of the ResNet50 network is 640*640*3, and there are four output layers, namely 160*160*256, 80*80*512, 40*40*1024, and 20*20*2048;

[0088] S6.2, Complex Feature Integration and Noise Suppression Module ST-MLP includes a complex feature extraction and integration process and a denoising process. The former strengthens the network's ability to integrate SAR image target feature information and transmits it to the denoising layer. The denoising layer processes the forward information to achieve the denoising effect, such as Figure 6 As shown, Figure 6 Part a is a schematic diagram of the ST-MLP structure, part b is a schematic diagram of the multilayer perceptron in the ST-MLP, and part c is a schematic diagram of the fully connected layer in the multilayer perceptron.

[0089] Strong ground-based scattered background noise can significantly interfere with a detector's rotation detection. Common solutions focus on using filters for better denoising, but overlook the importance of feature extraction, which can also affect the final denoising results. To address this, we designed the ST-MLP module, which uses a multi-layer perceptron to enhance the detector's ability to extract and integrate complex image features. The feature layer size input to the ST-MLP module is 160*160*256. First, absolute operations, global maximum pooling and Flatten processing are performed on the feature layer. At this time, the feature layer size is 1*1*C. Then, a multi-layer perceptron is used to extract complex features. The number of fully connected layers of the multi-layer perceptron is set to 2, and the LeakyReLU activation function is used to connect them to enhance the model's ability to extract and integrate complex features. The specific steps of the fully connected layer are as follows: first, the input features are linearly transformed through a linear layer, and then each feature dimension is normalized through a normalization layer. Then, a nonlinear mapping is introduced using the ReLU activation function to enhance the model's expressive power. This step is followed by another linear transformation and the application of the Sigmoid activation function to compress the output to the range between 0 and 1. The feature layer size becomes 1*1*1. Two fully connected layers are used to act on the input data together, allowing the model to construct more complex feature maps, thereby enhancing learning ability. After the feature map is processed by the multi-layer perceptron, the processed feature information is multiplied with the feature information processed by Flatten, ultimately achieving the extraction and integration of complex features.

[0090] The feature information of the target and background in SAR images is different. Using the soft threshold function, the target and background can be effectively classified. The features processed in the previous step are passed to the denoising layer, which uses the soft threshold function to classify the features. The soft threshold function is one of the commonly used denoising methods in signal processing and image processing. It removes noise while retaining as much effective information as possible. It is based on the assumption of signal sparsity and sets the smaller amplitude values ​​in the signal to zero while retaining the larger amplitude values ​​and treating them as signals. Its function definition is as follows:

[0091]

[0092] Where τ is the threshold, x is the input signal, and y is the output signal.

[0093] The denoising formula of the ST-MLP module is as follows:

[0094]

[0095] Τ=x2-τ

[0096] G=max(Τ,0)

[0097]

[0098] Among them, τ is the threshold, a1 is the residual edge eigenvalue after global average pooling and Faltten operation, a5 is the eigenvalue after multi-layer perceptron feature extraction and integration, T is the difference between the eigenvalue after absolute operation and the threshold, x2 is the eigenvalue after absolute operation, G is the result of comparing T with 0, and returns T when T is greater than 0, and returns 0 when T is less than or equal to 0. x' is the result of multiplying the symbol value x of the input eigenvalue by G term by term, which represents the final eigenvalue after processing by the ST-MLP module.

[0099] After the noise is suppressed by the soft threshold function, the final output is obtained through a convolution, normalization and activation function operation. The convolution kernel and step size are both 1. At this time, the feature layer size is 160*160*256;

[0100] S6.3. Considering that the edges of ship targets in SAR images are often blurred, it is difficult for the detector to achieve accurate rotation positioning. A azimuth perception module MC-PM is added to soften the target edge information and enhance the network's direction perception capability. Figure 7 shown.

[0101] Due to the unique imaging mechanism of SAR images, the edges of targets in the image are often blurred, leading to inaccurate rotational positioning. Therefore, the designed MC-PM position perception module uses distributed average pooling to soften target edge information, allowing the detector to properly perceive the target boundary and reduce the influence of scattered clutter. The feature layer input to the MC-PM module is 160*160*256 in size, and the pooling kernels of distributed average pooling are 5, 9, 13, and a no-op branch. The calculation formula for average pooling is as follows:

[0102]

[0103] The stride is 1, the kernel values ​​are 5, 9, and 13 respectively, and the padding is the integer part of the kernel value divided by 2.

[0104] After that, the processing results of each branch are fused from the channel dimension. At this time, the feature layer size is 160*160*1024. Secondly, the number of channels of the feature layer is reduced by a convolution with a kernel and a stride of 1, normalization and activation function operation. At this time, the feature layer size is 160*160*256.

[0105] The fused feature information is relatively redundant and fails to highlight positional features. Therefore, the MC-PM module adds hard-coded horizontal and vertical input channels to represent the coordinates of feature map pixels, highlighting positional features. These two hard-coded input channels are added to the processed feature map to represent the x and y coordinates of feature map pixels. These channels can learn more about the directional features of the feature map, enabling the MC-PM module to achieve a certain degree of coordinate awareness. Specifically, the x-coordinate channel is an H×W rank-1 matrix, with the first row filled with 0s, the second row with 1s, the third row with 2s, and so on. The y-coordinate channel is similar, but it fills the columns with constants instead of the rows to capture the y-orientation coordinate information. The feature layer size is now 160*160*(256+2). Finally, a convolution with a kernel and stride of 1, followed by normalization and activation, reduces the number of channels in the feature layer to 128, reducing redundant features and yielding the final output.

[0106] S6.4. The shallow features extracted by the backbone network contain a large number of features such as edge shapes, which are very important for target rotation positioning. In order to retain more shallow features, the present invention introduces a feature pyramid structure (PANet) as the neck network. This structure can integrate the shallow position information extracted by the backbone network with the deep semantic information to enhance the feature extraction capability of the model. The structure has four input feature layers with sizes of 160*160*128, 80*80*512, 40*40*1024 and 20*20*2048 respectively, and four output feature layers with the same size as the input feature layer.

[0107] S6.5, Gaussian distribution decoding prediction box adopts a single head prediction form that combines classification and regression. Its design is a 1×1 convolution. Four 1×1 convolutions are performed on the four output layers of the neck network to obtain the final output features. The sizes of the four output feature layers are 160*160*21, 80*80*21, 40*40*21 and 20*20*21 respectively. It is worth noting that we have added an angle output channel to provide more feature information for rotation detection. The number of output channels at all four positions is 21, which can be divided into 3×7 (4+1+1+1), where 4 represents the regression channels of the center point coordinates, target length and width, the first 1 is the angle channel, the second 1 is the confidence channel, and the third 1 is the category channel.

[0108] S6.6, Gaussian distribution decoding prediction box extracts features, performs offset prediction, and obtains (t x p ,t y p ,t h p,t w p ,t θ p ), decode the predicted box, and then convert the predicted box and the real box encoding form (x, y, h, w, θ) into a two-dimensional Gaussian distribution form (μ, Σ):

[0109] μ=(x,y) T

[0110]

[0111] Where x and y are the coordinates of the target center point, h and w represent the length and width of the target respectively, θ is the rotation angle, R is the rotation matrix, and Λ is the eigenvalue diagonal matrix.

[0112] After that, it is passed to the loss part for optimization. The loss part includes classification loss, confidence loss and regression loss. The classification and confidence loss function formulas are as follows:

[0113] Loss=-(y*log(p)+(1-y)*log(1-p))

[0114] Among them, y represents the true label, which takes 0 or 1, 0 represents a negative sample, 1 represents a positive sample, and p represents the probability of predicting the positive class. It is worth noting that the Sigmoid function operation must be performed before passing it to the loss function to map the real input to the (0,1) interval for binary classification.

[0115] The regression loss uses KL divergence as a distance measurement method to achieve the effect of dynamically adjusting parameters, thereby reducing the problem of regression mismatch between angles of targets of different sizes. The specific formula is as follows:

[0116]

[0117] Expand the three sub-terms of the above equation into:

[0118]

[0119] D kl Each term in is composed of a number of coupled parameters, forming a chain-like coupling relationship between all parameters. During detector optimization, each parameter can influence and optimize each other collaboratively, thus forming a self-modulating optimization mechanism for the model. This unique optimization mechanism enables GD-BPB to dynamically adjust parameters when processing targets of varying sizes, alleviating the problem of angular regression mismatches for targets of varying sizes.

[0120] Finally, the classification loss, confidence loss, and regression loss are fused proportionally to obtain the final loss value:

[0121] Lossall =L cls +L reg +L obj

[0122] Among them L cls Represents classification loss, the ratio is 1 / 160(0.5*(1 / 80)), L reg represents the regression loss, the ratio is 1 / 20 (0.05), L obj Represents the confidence loss, the ratio is 9 / 409600(1*(3*3) / (640*640)).

[0123] Figure 9 、 Figure 10 、 Figure 11 、 Figure 12 as well as Figure 13 The detection effect of an embodiment of the present application in two datasets, R-SSDD dataset and WR-HRSID dataset, is demonstrated. Figure 9 It can be seen that the rotation frame generated by an embodiment of the present application can fit the target well and the detection effect is significant. Figure 10 and Figure 12 This is a comparison chart of an embodiment of the present application with different methods in two data sets, where the yellow box indicates a correct prediction, the blue box indicates no detection, and the red box indicates a false alarm. (d) (h) of each figure represents the detection result of an embodiment of the present application, (a) (e) represents the detection result of the CSL method, (b) (f) represents the detection result of the Rotated Faster R-CNN method, and (c) (g) represents the detection result of the Rotated-ATSS method. It can be seen from the figure that an embodiment of the present application can generate a rotating box that fits the detection target better, and the fitting effect is more obvious than other networks, highlighting the superiority of the present application. Figure 11 and Figure 13 This is a comparative experiment of an embodiment of the present application with different methods in complex scenes of two data sets. It can be seen from the figure that the detection effect of an embodiment of the present application in complex scenes is still better than other detection methods. The number of correct detections is large, and the generated prediction box is more in line with the target, which verifies the practicality and generalization of the present application.

[0124] Based on the same inventive concept, embodiments of the present application also provide a device for SAR image-based directionally decoupled target detection, which is used to implement the aforementioned method for SAR image-based directionally decoupled target detection. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more embodiments of the device for SAR image-based directionally decoupled target detection provided below can be found in the aforementioned limitations of the method for SAR image-based directionally decoupled target detection, and will not be further elaborated here.

[0125] In an exemplary embodiment, a SAR image-oriented decoupling target detection device is provided, comprising:

[0126] The source data set acquisition module is used to acquire the source data set; the source data set includes: two high-resolution SAR image ship data;

[0127] A preprocessing module, configured to preprocess the source data set;

[0128] A target detection model construction module is used to construct a target detection model; the target detection model includes: a backbone network, an ST-MLP module, an MC-PM module, a neck network, and a Gaussian distribution decoding prediction box;

[0129] Loss function construction module, used to construct loss function;

[0130] A training module, configured to train the object detection model based on the preprocessed source data and the loss function;

[0131] The target orientation detection module is used to input the SAR image to be detected into the trained target detection model to obtain the target orientation detection result.

[0132] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 14 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store SAR image directional decoupling target detection data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a SAR image directional decoupling target detection method is implemented.

[0133] Those skilled in the art will understand that Figure 14The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0134] In an exemplary embodiment, a computer device is further provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0135] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0136] In an exemplary embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0137] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0138] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0139] The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may include, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic units, data processing logic units based on quantum computing, and the like.

[0140] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0141] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for directional decoupling target detection in SAR images, characterized in that: The SAR image-oriented decoupling target detection method includes: Acquire a source data set; the source data set includes: two high-resolution SAR image ship data; Preprocessing the source data set; Constructing a target detection model; the target detection model includes: a backbone network, an ST-MLP module, an MC-PM module, a neck network, and a Gaussian distribution decoding prediction box; The ST-MLP module is used to extract and integrate the main features of the image into complex features, and to denoise the features after complex feature extraction and integration; the main features of the image are obtained by inputting the SAR image to be detected into the backbone network; The following formula is used to denoise the features after complex feature extraction and integration: Τ=x2-τ G=max(Τ,0) Where τ is the threshold, a1 is the residual edge eigenvalue after global average pooling and Faltten operation, a5 is the eigenvalue after multi-layer perceptron feature extraction and integration, T is the difference between the eigenvalue after absolute operation and the threshold, x2 is the eigenvalue after absolute operation, G is the result of comparing T with 0, T is returned when T is greater than 0, and 0 is returned when it is less than or equal to 0, x' is the result of multiplying the symbol value x of the input eigenvalue by G term by term, which represents the final eigenvalue after processing by the ST-MLP module; The MC-PM module is used to perform average pooling operations on the denoised features to obtain multi-branch processing results; The denoised features are input into the MC-PM module for average pooling operation using the following formula: Where output is the eigenvalue after single-step average pooling, input is the output eigenvalue after ST-MLP module processing, stride is 1, indicating the step size, kernel is the pooling kernel, which are 5, 9, and 13 respectively, and padding is the integer part of the kernel value divided by 2; Construct loss function; Training the object detection model based on the preprocessed source data and the loss function; The SAR image to be detected is input into the trained target detection model to obtain the target orientation detection result.

2. The method for directional decoupling target detection based on SAR images according to claim 1, characterized in that: Preprocessing the source data set specifically includes the following steps: The two high-resolution SAR image ship data were cropped to 640*640 size using linear interpolation method; The cropped data set is divided into training set and validation set according to the ratio.

3. The method for directional decoupling target detection based on SAR images according to claim 1, characterized in that: The expression of the loss function is as follows: Loss all =L cls +L reg +L obj Among them, L cls represents the classification loss, L reg represents the regression loss, L obj represents the confidence loss.

4. The method for directional decoupling target detection based on SAR images according to claim 1, characterized in that: Inputting the SAR image to be detected into the trained target detection model to obtain the target orientation detection result specifically includes the following steps: Inputting the SAR image to be detected into the backbone network to obtain the main features of the image; Inputting the main features of the image into the ST-MLP module to extract and integrate complex features, and denoising the features after complex feature extraction and integration; The denoised features are input into the MC-PM module for average pooling operation to obtain a multi-branch processing result; Fusing the multi-branch processing results from a channel dimension; Reduce the dimension of the features after channel dimension fusion; Input the reduced dimension features into the horizontal and vertical hard-coded channels to obtain position feature information; Perform channel conversion on the position feature information to obtain a final output result; Inputting the final output result into the neck network for information fusion; The information fused features are input into the Gaussian distribution decoding prediction box to obtain the target orientation detection result.

5. A SAR image directional decoupling target detection device, characterized in that: The SAR image oriented decoupling target detection device comprises: The source data set acquisition module is used to acquire the source data set; the source data set includes: two high-resolution SAR image ship data; A preprocessing module, configured to preprocess the source data set; A target detection model construction module is used to construct a target detection model; the target detection model includes: a backbone network, an ST-MLP module, an MC-PM module, a neck network, and a Gaussian distribution decoding prediction box; The ST-MLP module is used to extract and integrate the main features of the image into complex features, and to denoise the features after complex feature extraction and integration; the main features of the image are obtained by inputting the SAR image to be detected into the backbone network; The following formula is used to denoise the features after complex feature extraction and integration: Τ=x2-τ G=max(Τ,0) Where τ is the threshold, a1 is the residual edge eigenvalue after global average pooling and Faltten operation, a5 is the eigenvalue after multi-layer perceptron feature extraction and integration, T is the difference between the eigenvalue after absolute operation and the threshold, x2 is the eigenvalue after absolute operation, G is the result of comparing T with 0, T is returned when T is greater than 0, and 0 is returned when it is less than or equal to 0, x' is the result of multiplying the symbol value x of the input eigenvalue by G term by term, which represents the final eigenvalue after processing by the ST-MLP module; The MC-PM module is used to perform average pooling operations on the denoised features to obtain multi-branch processing results; The denoised features are input into the MC-PM module for average pooling operation using the following formula: Where output is the eigenvalue after single-step average pooling, input is the output eigenvalue after ST-MLP module processing, stride is 1, indicating the step size, kernel is the pooling kernel, which are 5, 9, and 13 respectively, and padding is the integer part of the kernel value divided by 2; Loss function construction module, used to construct loss function; A training module, configured to train the object detection model based on the preprocessed source data and a loss function; The target orientation detection module is used to input the SAR image to be detected into the trained target detection model to obtain the target orientation detection result.

6. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the SAR image directional decoupling target detection method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the SAR image directional decoupling target detection method according to any one of claims 1 to 4 are implemented.

8. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the SAR image directional decoupling target detection method according to any one of claims 1 to 4 are implemented.

Images