SAR (Synthetic Aperture Radar) target orientation discrimination method based on feature region segmentation
By combining feature region segmentation and vector angle calculation with U-Net and deep learning networks, the problem of insufficient target orientation angle information extraction in SAR images is solved, achieving standardized quantization of angles and improved discrimination accuracy.
Patent Information
- Application Number
- CN202511010807.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-22
- Publication Date
- 2025-11-07
AI Technical Summary
Existing technologies struggle to effectively extract target orientation angle information from SAR images, failing to meet the precision requirements of practical applications.
A feature region segmentation method is adopted to calculate the SAR target orientation by the angle between the vector direction and clockwise. The U-Net segmentation network and deep learning network are combined to achieve standardized quantization of the angle and complementary enhancement of deep semantics.
It improves the model's generalization ability, enables standardized quantification of angles in target orientation discrimination, is suitable for scenarios with low contrast between the target and the background, and improves discrimination accuracy.
Smart Images
Figure CN120912671A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the technical field of SAR image intelligent interpretation, and particularly relates to a SAR target orientation discrimination method based on feature region segmentation. BACKGROUND
[0002] As a remote sensing device for realizing detection and reconnaissance tasks by means of radar echo signals, synthetic aperture radar (SAR) has been widely applied in various fields of military and civilian due to its unique advantages of all-weather and all-day and high-resolution SAR image acquisition. In the intelligent interpretation of SAR images, target orientation angle information is of great significance to scenarios such as motion target tracking and target trend judgment. For example, by obtaining the target orientation, the travel trajectory or potential trend of the moving target can be more accurately predicted.
[0003] However, the prior art has the following disadvantages: target detection tasks are mostly focused on target positioning and classification, and there is insufficient extraction of target orientation angle information required for tracking moving targets and judging target trends; although there are research results on SAR target rotation box angle prediction, there is still no method that can effectively discriminate the orientation of SAR targets, which cannot meet the precise demand for target orientation information in actual applications.
[0004] Therefore, there is an urgent need for a new method. SUMMARY
[0005] The purpose of the present application is to provide a SAR target orientation discrimination method based on feature region segmentation, which realizes the standardization and quantification of angles by introducing vector direction and clockwise angle calculation into SAR target orientation labeling, effectively improves the generalization ability of the model, and at the same time, adopts the method of concatenating the channels of the cropped feature map and the up-sampling result to enhance the complementarity of shallow details and deep semantics.
[0006] To achieve the above purpose, the present application provides a SAR target orientation discrimination method based on feature region segmentation, comprising the following steps:
[0007] S1, labeling the head and tail regions of the detected SAR target slice image, calculating the angle between the direction of the tail region pointing to the head region, taking the angle as the angle label, and obtaining the head and tail region labeling data and angle labeling data;
[0008] S2, taking the head and tail region labeling data in S1 as training data to train a U-Net segmentation network model; performing parameter optimization on the constructed U-Net segmentation network, and then outputting the head and tail region segmentation result data;
[0009] S3, taking the segmentation result data in S2 as input, combining the angle labeling data in S1 to train the deep learning network; and taking the angle prediction as a classification task, optimizing the deep learning network to obtain an angle prediction network;
[0010] S4, using the optimized U-Net segmentation network in S2 and the angle prediction network in S3 to process the input SAR target slice; and outputting the orientation angle of the target.
[0011] Preferably, in S1, the included angle of the direction of the tail region pointing to the head of the target is specifically:
[0012] Taking the upper left corner of the SAR target detection slice image as the coordinate origin (0, 0), the horizontal right direction as the positive direction of the x-axis, and the vertical downward direction as the positive direction of the y-axis, the coordinates of the target head center point are:
[0013] p head =(x1,y1);
[0014] Wherein, (x1, y1) is the coordinates of the target head center point;
[0015] Let the coordinates of the tail region center point be:
[0016] p tail =(x2,y2);
[0017] Wherein, (x2, y2) is the coordinates of the tail region center point;
[0018] The vector of the tail region pointing to the head of the target is:
[0019]
[0020] Wherein, is the vector of the tail region pointing to the head of the target;
[0021] Taking the image vertical upward direction, i.e. the unit vector consistent with the negative direction of the y-axis as the reference, the clockwise included angle θ between and is calculated, and the calculation formula is as follows:
[0022]
[0023] Wherein, a is the minimum included angle between the vectors;
[0024] The x component sign of the vector is combined to determine the clockwise direction:
[0025] If x1-x2≥0, then the clockwise included angle θ=a;
[0026] If x1-x2<0, then the clockwise included angle θ=360°-α.
[0027] θ is taken as an angle label.
[0028] Preferably, in S2, the U-shaped architecture of the U-Net segmentation network comprises a downsampling operation and an upsampling operation, the downsampling operation transmits context information, and the upsampling operation fuses deep and shallow features.
[0029] Preferably, in S2, the parameter optimization of the constructed U-Net segmentation network is specifically as follows: a softmax function is defined, and the probability distribution of an activation function of each pixel position of the kth pixel channel is calculated, and the calculation formula is as follows:
[0030] p k (x)=exp(a k (x)) / ∑ k′＝1 exp(a k′ (x));
[0031] Wherein, a k (x) is the activation function of each pixel position of the kth pixel channel, and then the cross entropy is used to supervise the deviation of the softmax of each position to 1.
[0032] The cross entropy loss is calculated for each pixel position, and the calculation formula is as follows:
[0033] E=∑ x∈Ω w(x)log(p l (x));
[0034] Wherein, Ω is all pixel positions, w(x) is a weight introduced to improve the importance of some pixels, and l represents the pixel position where the label is located.
[0035] Preferably, in the training process of S3, the angle prediction is regarded as a classification task, a binary cross entropy loss function is used to calculate the deviation between the predicted angle and the real angle, and the calculation mode is as follows:
[0036] L ang =BCE(q t ,q p );
[0037] Wherein, q t is the angle true value, and q p is the angle prediction value.
[0038] The application also provides a SAR target orientation discrimination system based on feature region segmentation, comprising:
[0039] The data labeling module is configured to label a head region and a tail region of a SAR target slice obtained through detection, calculate an included angle of a direction of the tail region to the head region, and take the included angle as an angle label to obtain head region and tail region labeling data and angle labeling data.
[0040] The segmentation network module is connected with the data labeling module and configured to take the head region and tail region labeling data as training data to train a U-Net segmentation network model, perform parameter optimization on the constructed U-Net segmentation network, and then output segmentation result data of the head region and the tail region.
[0041] The angle prediction network module is connected with the segmentation network module and configured to take the segmentation result data as input and train a deep learning network in combination with the angle labeling data, take angle prediction as a classification task, optimize the deep learning network, and obtain an angle prediction network.
[0042] The orientation discrimination execution module is connected with the angle prediction network module and configured to process the input SAR target slice and output an orientation angle of the target.
[0043] Therefore, the SAR target orientation discrimination method with feature region segmentation has the following beneficial effects compared with the prior art.
[0044] (1) The vector direction and the clockwise included angle calculation are introduced into the SAR target orientation labeling, the standardization and quantization of the angle are realized, and the generalization ability of the model is improved.
[0045] (2) The channels of the cropped feature map and the up-sampling result are connected in series instead of being simply added, the complementarity of the shallow details and the deep semantics is enhanced, and the method is especially suitable for a scene with low contrast between a target and a background in a SAR image.
[0046] The technical solutions of the present application will be further described in detail below with reference to the drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS
[0047] Figure 1 The flowchart of the SAR target orientation discrimination method with feature region segmentation according to the present application is shown. DETAILED DESCRIPTION
[0048] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some but not all of the embodiments of the present application. All other embodiments obtained by a person of ordinary skill in the art without creative work shall fall within the protection scope of the present application. Unless otherwise defined, the technical terms or scientific terms used in the present application shall have the usual meanings understood by those skilled in the art.
[0049] Embodiment one
[0050] As shown in the figure, a feature region segmentation SAR target orientation discrimination method of the present application comprises the following steps: Figure 1
[0051] S1, for the detected SAR target slice graph, the head and tail regions of the target are marked by a rectangular frame (not a single point marking). This design can avoid the problem of slow convergence of small initial value loss in training by single point marking, and provide more rich spatial feature information for the segmentation network;
[0052] The included angle of the direction of the tail region pointing to the target head is specifically:
[0053] Taking the upper left corner of the SAR target detection slice graph as the coordinate origin (0, 0), the horizontal right direction as the positive direction of the x-axis, and the vertical downward direction as the positive direction of the y-axis, the target head center point coordinates are:
[0054] p head =(x1,y1);
[0055] Wherein, (x1, y1) is the target head center point coordinates;
[0056] Let the tail region center point coordinates be:
[0057] p tail =(x2,y2);
[0058] Wherein, (x2, y2) is the tail region center point coordinates;
[0059] The vector of the tail region pointing to the target head is:
[0060]
[0061] Wherein, is the vector of the tail region pointing to the target head;
[0062] Taking the image vertical upward direction, i.e. the unit vector consistent with the negative direction of the y-axis as the reference, the included angle between and the clockwise included angle θ of the two vectors, and the calculation formula is as follows:
[0063]
[0064] wherein α is the minimum included angle between the vectors;
[0065] Combining the x component of the vector to determine the clockwise direction:
[0066] If x1-x2≥0, then the clockwise included angle θ=α;
[0067] If x1-x2<0, then the clockwise included angle θ=360°-α.
[0068] Take θ as the angle label, which is used for the training of the subsequent angle prediction network. This method converts abstract orientation information into quantifiable angle values, ensuring the objectivity and consistency of the labeling;
[0069] S2, taking the target head and tail region labeling data in S1 as training data, training the U-Net segmentation network model;
[0070] The U-shaped architecture includes down-sampling and up-sampling operations. Down-sampling passes context information to higher resolution layers, and up-sampling passes deep high-resolution information back to shallow layers, realizing the fusion of deep and shallow features, thereby accurately segmenting the target head and tail region and providing structured features for subsequent angle prediction.
[0071] The region segmentation is regarded as a classification task, and a classification loss function is used to optimize the parameters of the U-Net network. Specifically, a softmax function is defined to calculate the activation function probability distribution of each pixel position in the kth pixel channel, and the calculation formula is as follows:
[0072] p k (x)=exp(a k (x)) / ∑ k′＝1 exp(a k′ (x));
[0073] wherein a k (x) is the activation function of each pixel position in the kth pixel channel, and then cross entropy is used to supervise the deviation of softmax for 1 at each position;
[0074] The cross entropy loss is calculated for each pixel position, and the calculation formula is as follows:
[0075] E=∑ x∈Ω w(x)log(p l (x));
[0076] Wherein, Omega is all pixel positions, w(x) is a weight introduced to improve the importance of some pixels, and I represents the pixel position where the label is located.
[0077] The target head and tail region annotation data in S1 are used as a supervision signal to adjust the network parameters through gradient descent, so that the segmentation result (mask) output by the network is consistent with the label, and the target head region and tail region segmentation result data are output.
[0078] S3, a deep learning network for converting the segmentation result into an orientation angle is built, the deep learning network takes the segmentation result data in S2 as input and is trained in combination with the angle label annotated in S1.
[0079] During the training process, the angle prediction is regarded as a classification task, and a binary cross-entropy loss function is used to calculate the deviation between the predicted angle and the true angle, and the calculation method is as follows:
[0080] L ang =BCE(q t ,q p );
[0081] Wherein, q t is the true value of the angle, and q p is the predicted value of the angle.
[0082] The network parameters are optimized through gradient descent to reduce the loss value until the model can stably output a high-precision angle prediction result, the angle prediction network is constructed and trained, and the angle prediction network is obtained.
[0083] S4, input the SAR airplane detection slice image to be judged into the trained U-Net segmentation network to obtain the nose and tail regions, and then input the segmentation result into the angle prediction network to finally output the included angle between the nose orientation and the picture directly above in clockwise direction, and complete the orientation discrimination.
[0084] Therefore, the SAR target orientation discrimination method of the above-mentioned feature region segmentation is adopted, the vector direction and the clockwise included angle calculation are introduced into the SAR target orientation annotation, the standardization and quantization of the angle are realized, and the generalization ability of the model is effectively improved. At the same time, the channels of the cropped feature map and the up-sampling result are connected in series, and the complementarity of the shallow details and the deep semantics is enhanced.
[0085] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application but not to limit them, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can still be modified or replaced by equivalents, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.
Claims
1. A method for SAR target orientation discrimination by feature region segmentation, characterized in that, The method comprises the following steps: S1, labeling the target head and tail region of the detected SAR target slice image, calculating the included angle of the direction of the tail region pointing to the target head, taking the included angle as the angle label, and obtaining the target head and tail region labeling data and the angle labeling data; S2, taking the target head and tail region labeling data in S1 as training data to train a U-Net segmentation network model; performing parameter optimization on the constructed U-Net segmentation network, and then outputting the target head region and tail region segmentation result data; S3, taking the segmentation result data in S2 as input, and training a deep learning network in combination with the angle labeling data in S1; and taking angle prediction as a classification task, optimizing the deep learning network, and obtaining an angle prediction network; S4, using the optimized U-Net segmentation network in S2 and the angle prediction network in S3 to process the input SAR target slice; and outputting the orientation angle of the target.
2. The method according to claim 1, wherein the method is characterized by, In S1, the calculation of the included angle of the direction of the tail region pointing to the target head is as follows: Taking the upper left corner of the SAR target detection slice image as the coordinate origin (0, 0), the horizontal right direction as the positive direction of the x-axis, and the vertical downward direction as the positive direction of the y-axis, the coordinates of the target head center point are: p head = (x1, y1); Where (x1, y1) is the coordinates of the target head center point. The coordinates of the tail region center point are: p tail = (x2, y2); Where (x2, y2) is the coordinates of the tail region center point. The vector of the tail region pointing to the target head is: wherein, is a vector pointing from the tail region to the target head; with the y-axis negative direction consistent unit vector As a reference, calculate The clockwise angle θ with The calculation formula is as follows: Where a is the smallest included angle between the vectors. Combining vectors The x-component sign determines the clockwise direction: If x1-x2≥0, the clockwise included angle θ=a; If x1-x2<0, the clockwise included angle θ=360°-a; Taking θ as the angle label.
3. The method of claim 1, wherein the feature region segmentation is performed by using a neural network. In S2, the U-shaped architecture of the U-Net segmentation network includes down-sampling operation and up-sampling operation, the down-sampling operation transmits context information, and the up-sampling operation fuses deep and shallow features.
4. The method of claim 1, wherein the feature region segmentation is performed by using a neural network. In S2, the parameter optimization of the constructed U-Net segmentation network is as follows: defining a softmax function, calculating the activation function probability distribution of each pixel position in the kth pixel channel, and the calculation formula is: p k (x) = exp(a k (x)) / ∑ k′＝1 exp(a k′ (x)) where a k (x) is the activation function for each pixel location of the kth pixel channel, followed by a cross-entropy to supervise the softmax for each location to be 1 ; Calculating the cross-entropy loss of each pixel position, and the calculation formula is: E = ∑ x∈Ω w(x) log(p l (x)); Where Ω is all pixel positions, w(x) is a weight introduced to improve the importance of certain pixels, and l represents the pixel position of the label.
5. The method of claim 1, wherein the feature region segmentation is performed by using a neural network. In the training process of S3, the angle prediction is regarded as a classification task, a binary cross-entropy loss function is used to calculate the deviation between the predicted angle and the true angle, and the calculation method is as follows: L ang = B CE(q t , q p ); where q t is the angle true value, q p is the angle predicted value.
6. A feature region segmentation based SAR target orientation discrimination system, applied to the feature region segmentation based SAR target orientation discrimination method of any one of claims 1-5, characterized in that, It comprises: A data labeling module for labeling the target head and tail region of the detected SAR target slice image, calculating the included angle of the direction of the tail region pointing to the target head, taking the included angle as the angle label, and obtaining the target head and tail region labeling data and the angle labeling data; A segmentation network module connected with the data labeling module, configured to take the target head and tail region labeling data as training data to train a U-Net segmentation network model; perform parameter optimization on the constructed U-Net segmentation network, and then output the target head region and tail region segmentation result data; The angle prediction network module is connected with the segmentation network module, is configured to take the segmentation result data as input, train a deep learning network in combination with angle labeling data, and regard angle prediction as a classification task, optimize the deep learning network, and obtain an angle prediction network. The orientation discrimination execution module is connected with the angle prediction network module, is configured to process the input SAR target slice, and output an orientation angle of the target.