A Deep Learning-Based Method for Aircraft Target Orientation Determination in SAR Images
By employing a deep learning-based approach of semantically ordered vertex encoding, attention enhancement, and mirror penalty loss function, the instability and misjudgment issues in aircraft target orientation discrimination in SAR images are addressed. This approach achieves high-precision 360° all-angle orientation discrimination and is applicable to complex backgrounds and environments with strong speckle noise.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING INSTITUTE OF TECHNOLOGY ANHUI INSTITUTE OF AEROSPACE INFORMATION
- Filing Date
- 2026-05-07
- Publication Date
- 2026-07-31
AI Technical Summary
Existing technologies struggle to effectively distinguish the nose and tail orientations of aircraft targets in SAR images. Furthermore, traditional methods suffer from discrimination errors and gradient abrupt changes at angular boundaries, impacting the stability and accuracy of network training.
A deep learning-based method for aircraft orientation discrimination in SAR images is constructed by employing techniques such as semantically ordered vertex encoding, decoupled vectorized orientation prediction, attention feature enhancement, and mirror penalty loss function. The method provides orientation semantic constraints through semantically ordered vertex encoding, embeds an attention enhancement module, decouples the orientation prediction branch, and constructs a mirror penalty loss function to achieve stable and high-precision 360° all-angle orientation discrimination.
It improves the accuracy and stability of aircraft target orientation determination in SAR images, solves the problem of confusion between the nose and tail of the aircraft, enhances the stability of network training and the accuracy of angle prediction, and is suitable for complex backgrounds and environments with strong speckle noise.
Smart Images
Figure CN122493214A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of SAR image processing and deep learning technology, and in particular to a method for determining the orientation of aircraft targets in SAR images based on deep learning. Background Technology
[0002] Synthetic Aperture Radar (SAR) possesses all-weather, all-day imaging capabilities, unaffected by weather conditions such as illumination, clouds, fog, rain, or snow. It holds irreplaceable application value in fields such as homeland air defense, battlefield reconnaissance, marine monitoring, and civilian remote sensing. With the rapid development of high-resolution radar satellite technology, the spatial resolution of SAR images has significantly improved. The application requirements for target reconnaissance and identification have evolved from traditional "target presence / absence detection" to "refined semantic information extraction from targets." Among these, the orientation information of aircraft targets is the core refined semantic information, providing crucial support for applications such as battlefield situational awareness, target intent analysis, and airport dispatch monitoring.
[0003] Unlike optical images, SAR images, based on coherent imaging mechanisms, suffer from severe speckle noise. The geometry and scattering characteristics of targets exhibit discretized features, resulting in significant differences between target outlines and optical images. Furthermore, SAR employs side-view imaging, leading to perspective distortion between the target's geometry in the image and its actual physical structure, further increasing the complexity of target morphology representation. For aircraft targets, the scattering structures of their nose and tail in SAR images are highly similar. Traditional target detection methods based on rotating bounding boxes rely solely on regressing the major axis parameters of the bounding box to predict angles, failing to effectively distinguish the orientation polarity of the nose and tail, and easily resulting in 180° orientation flip errors. Moreover, traditional methods based on numerical angle regression exhibit numerical jumps at angular boundaries such as 0°, 360°, and 180°, causing abrupt gradient changes during network training, severely limiting the stability of network fitting and the accuracy of angle prediction. Therefore, how to effectively distinguish the nose and tail of aircraft targets in SAR images, overcome the instability of angular boundary regression, and achieve stable and high-precision 360° all-angle orientation discrimination is a core technical problem that urgently needs to be solved in the current field of SAR image processing. Summary of the Invention
[0004] The purpose of this invention is to propose a deep learning-based method for aircraft target orientation discrimination in SAR images. Through core technologies such as semantically ordered vertex encoding, decoupled vectorized orientation prediction, attention feature enhancement, and mirror penalty loss function, it achieves stable rotation detection and high-precision 360° all-angle orientation discrimination of aircraft targets in SAR images.
[0005] To achieve the above objectives, this invention provides a deep learning-based method for determining the orientation of aircraft targets in SAR images, comprising the following steps: Step S1: Preprocess the acquired SAR image dataset, combine the true orientation information of the aircraft target to perform semantically ordered vertex encoding on the target labeled rotation box, and generate training labels embedded with orientation polarity information. Step S2: Construct a SAR target rotation detection network, which includes a cascaded feature extraction network, a feature fusion network, and a decoupled detection head; an attention enhancement module is embedded in the feature extraction network, and a high-resolution feature fusion branch is set in the feature fusion network; Step S3: Set up an independent orientation prediction branch in the decoupled detection head. The orientation prediction branch is completely decoupled from the target classification branch and bounding box regression branch of the detection head. The unit vector of the aircraft target orientation is regressed through the orientation prediction branch to realize the full-angle continuous expression of the target orientation. Step S4: Construct an orientation loss function with a mirror penalty mechanism. Based on the cosine similarity between the predicted orientation unit vector and the true orientation unit vector, apply a loss amplification penalty to the orientation prediction results that have reversed polarity. Step S5: Iteratively train the SAR target rotation detection network using the orientation loss function. Input the SAR image to be detected into the trained network and output the rotation detection result of the aircraft target and the 360° orientation information.
[0006] Preferably, in step S1, the specific implementation process of semantically ordered vertex encoding includes: Step S11: The annotation information of aircraft targets in the SAR image is uniformly represented as a rotated bounding box. The parameterized expression of the rotated bounding box is: ; in, , These are the x-coordinate and y-coordinate of the center of the aircraft target rotation frame, respectively. The width of the rotating frame. The height of the rotating frame. The actual orientation angle of the aircraft target; Step S12: Calculate the coordinates of the four vertices of the rotating frame based on the parameters, determine the side where the nose is located by combining the actual orientation information of the aircraft target, and perform clockwise semantic sorting on the four vertices to generate an ordered vertex set: ; in, , These are the two vertices located on the side of the nose. , These are the two vertices located on the tail side of the aircraft; Step S13: Combine the category information of the aircraft target with the coordinate information of the ordered vertex set to generate training labels embedded with the nose-tail polarity orientation information.
[0007] Preferably, in step S2, the feature extraction network adopts a deep convolutional neural network architecture, which extracts multi-scale features of the input SAR image through multi-layer convolution and downsampling operations; the attention enhancement module is embedded in the convolutional layer of the feature extraction network, and adopts a parameterless SimAM attention module, which enhances the features of the nose scattering center region in the feature map by calculating the neuron energy function, while suppressing the feature weights of background coherent speckle and irrelevant regions, thereby strengthening the network's ability to perceive the differentiated features of the aircraft's nose and tail scattering structure.
[0008] Preferably, in step S2, the feature fusion network adopts an FPN+PAN bidirectional fusion architecture. The high-resolution feature fusion branch adds a high-resolution feature layer with a downsampling rate of 2 times in the feature fusion network. The high-resolution feature layer and the multi-scale feature map in the bidirectional fusion architecture are cross-layer stitched and convolutionally fused to retain the fine-grained scattering structure information of small-scale aircraft targets in SAR images and improve the detection and orientation discrimination accuracy of small targets.
[0009] Preferably, in step S3, the orientation prediction branch adopts a dual-channel convolution output structure, where the output values of the two channels correspond to the two components of the target orientation unit vector, and the expression for the unit vector is: ; in, To predict the orientation of the unit vector, The predicted orientation angle of the aircraft target by the model.
[0010] Preferably, in step S4, the process of constructing the orientation loss function with a mirror penalty mechanism includes: Step S41: Set the true orientation unit vector of the aircraft target as: ; in, For the true orientation unit vector, The actual orientation angle of the aircraft target; Step S42: Calculate the cosine similarity between the predicted orientation unit vector and the true orientation unit vector, expressed as: ; in, To predict the angle between the orientation and the actual orientation, Let cosine similarity be the similarity between two vectors. To predict the orientation of the unit vector, The predicted orientation angle of the aircraft target by the model; Step S43: Construct an orientation loss function with mirror penalty based on cosine similarity, the expression of which is: ; in, Towards the loss value, The mirror penalty coefficient; when The angle between the predicted orientation and the actual orientation hour, Training is performed using the base angle error loss; when The angle between the predicted orientation and the actual orientation hour, , A penalty amplification factor greater than 1 is used to amplify the loss of prediction results that have undergone polarity reversal, thereby increasing the gradient update magnitude of samples with reversed direction.
[0011] Preferably, in step S5, the specific process of model training is as follows: the preprocessed SAR image and the corresponding training label are input into the SAR target rotation detection network, the network parameters are iteratively updated using a stochastic gradient descent optimization algorithm, the loss function is minimized through backpropagation algorithm until the network converges, and the model training is completed; during the training process, multi-scale training and online hard example mining strategies are used to improve the generalization ability and robustness of the model.
[0012] Preferably, in step S5, the specific process of target orientation inference is as follows: the SAR image to be detected is input into the trained SAR target rotation detection network; the decoupled detection head outputs the aircraft target category, rotation box position parameters, and predicted orientation unit vector; the predicted orientation unit vector is solved using the arctangent function to obtain the true orientation angle of the aircraft target, expressed as: ; in, The final output is the aircraft target's orientation value at all angles, with a value range of [value range missing]. or It completes the 360° all-angle orientation determination of the aircraft target.
[0013] Therefore, the present invention employs the above-described deep learning-based SAR image aircraft target orientation determination method, which has the following advantages: (1) This invention implicitly encodes the nose-tail polarity information of the aircraft target into the arrangement order of the vertices of the rotating frame through a semantically ordered vertex encoding method, providing explicit orientation semantic constraints for network training, strengthening the network's ability to learn orientation polarity from the data level, and reducing the probability of nose-tail confusion from the root. (2) The present invention embeds an attention enhancement module in the feature extraction network, which can specifically enhance the feature weight of the scattering center region of the aircraft nose, suppress background coherent speckle clutter interference, strengthen the differential feature expression of the scattering structure of the nose and tail, and solve the discrimination problem caused by the similarity of the scattering features of the nose and tail in SAR images; at the same time, through the high-resolution feature fusion branch, the fine-grained scattering information of small-scale aircraft targets is effectively preserved, and the detection and orientation discrimination accuracy of small targets is improved. (3) The present invention adopts an independent orientation prediction branch that is completely decoupled from the classification and regression branches. By using the vectorized expression of the regression orientation unit vector, the discrete angle regression is converted into a continuous trigonometric function regression, which mathematically guarantees the periodic continuity of the angle prediction. This completely eliminates the numerical jump and gradient mutation problem at the angle boundary in the traditional angle regression method, and greatly improves the stability of network training and the accuracy of angle prediction. (4) The present invention constructs an orientation loss function with a mirror penalty mechanism. Based on cosine similarity, a loss amplification penalty is applied to the prediction results that have polarity reversal, which significantly increases the gradient update amplitude of the orientation reversal samples, effectively suppresses the 180° orientation flip problem that occurs during model training, and further improves the accuracy of orientation discrimination. (5) The overall technical solution of the present invention can realize the rotation detection and 360° orientation discrimination of aircraft targets in SAR images from end to end. It has excellent robustness and discrimination accuracy in complex backgrounds, strong speckle noise, small targets and other scenarios. No additional post-processing steps are required. It is easy to implement in engineering and has extremely high military and civilian application value.
[0014] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0015] Figure 1 This is a schematic diagram of the overall process and network architecture of a deep learning-based SAR image aircraft target orientation discrimination method in an embodiment of the present invention. Detailed Implementation
[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0017] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0018] Example like Figure 1 As shown in the figure, this embodiment proposes a method for aircraft target orientation determination based on deep learning in SAR images. The specific implementation steps are as follows: Step 1: Data preprocessing and semantic vertex encoding label generation: Dataset acquisition: Obtain publicly available SAR image aircraft target datasets, which contain SAR images of different resolutions, scenes, and aircraft types, along with corresponding aircraft target annotation information.
[0019] Standardize annotation format: Convert the annotation information of all aircraft targets in the dataset into a parametric form of rotated bounding boxes, with the following expression: ;in, , These are the x-coordinate and y-coordinate of the center of the aircraft target's rotating bounding box in the image coordinate system, respectively, in pixels; The width of the rotating frame corresponds to the minor axis length of the aircraft target fuselage, in pixels; The height of the rotating frame corresponds to the major axis length of the aircraft target fuselage, in pixels; The true orientation angle of the aircraft target is defined as the angle between the major axis of the fuselage and the positive x-axis of the image coordinate system, with a value range of [value missing]. .
[0020] Semantic ordered vertex encoding: The original coordinates of the four vertices of the rotated bounding box are calculated based on the parameters of the rotated bounding box. Combined with the actual orientation information of the aircraft target, the short side where the nose is located is determined. The four vertices are then semantically sorted clockwise to generate an ordered vertex set. ;in, , These are the two vertices located on the shorter side of the nose of the aircraft. , These are two vertices located on the shorter side of the tail. The order of the vertices strictly follows a clockwise rule, making the order of the vertices strongly correlated with the orientation polarity of the aircraft target. This implicitly encodes the polarity information of the nose-tail axis into the vertex coordinate sequence.
[0021] Training label generation: The category labels of aircraft targets (such as civil airliners, fighter jets, helicopters, etc.) are combined with the coordinate information of an ordered set of vertices to generate the final training labels. At the same time, SAR images are preprocessed by normalization, denoising and other operations to complete the construction of the training dataset.
[0022] Step 2: Constructing the SAR Target Rotation Detection Network: The SAR target rotation detection network constructed in this embodiment is an end-to-end fully convolutional network architecture, including three core parts: a feature extraction network, a feature fusion network, and a decoupled detection head, which are cascaded in sequence. The specific structure is as follows: Feature extraction network: ResNet-50 is used as the backbone network. Through five convolutional stages and downsampling operations, multi-scale features of the input SAR image are extracted, outputting five feature maps of different scales (C1-C5) with downsampling ratios of 2, 4, 8, 16, and 32, respectively. SimAM attention enhancement modules are embedded in the 3rd, 4th, and 5th convolutional stages of ResNet-50. SimAM attention modules are parameter-free attention modules that construct neuron energy functions to evaluate the importance of each neuron, enhancing the high-response region of the aircraft nose scattering center in the feature map while suppressing the low-response region of background coherent speckle, thus strengthening the network's perception of the differentiated features of the nose and tail scattering structures.
[0023] Feature Fusion Network: A bidirectional feature fusion architecture of FPN+PAN is adopted. Based on the traditional FPN architecture, a high-resolution feature layer F2 with a 2x downsampling rate is added, corresponding to the C2 feature layer output by the backbone network, to construct a high-resolution feature fusion branch. In the top-down FPN fusion path, high-level semantic features are upsampled and then concatenated with the high-resolution feature layer F2. In the bottom-up PAN fusion path, fine-grained features of the high-resolution feature layer are downsampled and then fused with high-level features. Finally, four fused feature maps of different scales, P2-P5, are output, corresponding to downsampling rates of 8, 16, 32, and 64, respectively. The high-resolution feature fusion branch effectively preserves the fine-grained scattering structure information of small-scale aircraft targets.
[0024] Decoupled Detection Head: For the fused feature map at each scale, a completely decoupled detection head structure is set up. The detection head is divided into three independent branches: object classification branch, bounding box regression branch, and orientation prediction branch. Each of the three branches uses an independent convolutional layer for parameter learning, avoiding learning conflicts between different tasks and improving the fitting accuracy of each branch.
[0025] Step 3: Decoupled Orientation Prediction Branch Design: An independent orientation prediction branch is set up in the decoupled detection head. The orientation prediction branch uses two 3×3 convolutional layers to extract orientation features, and finally outputs the dual-channel prediction results through a 1×1 convolutional layer. The output values of the two channels correspond to the sine and cosine components of the target orientation unit vector, respectively, that is, the predicted orientation unit vector: ;in, This represents the aircraft target orientation angle predicted by the model. This vectorized representation transforms the traditional 360° discrete angle numerical regression task into a continuous trigonometric function regression task. Since both sine and cosine functions are continuous periodic functions in the real number domain, the angle prediction maintains complete periodic continuity in the mathematical space. This completely solves the problem of numerical jumps and gradient mutations at angle boundaries such as 0°, 180°, and 360° in traditional angle regression methods, improving the stability of network training and the accuracy of angle prediction.
[0026] Step 4: Loss Function Construction: In this embodiment, a multi-task joint loss function is used to train the network end-to-end. The total loss function includes three parts: target classification loss, bounding box regression loss, and orientation loss with mirror penalty mechanism.
[0027] Orientation loss function with mirror penalty mechanism The construction process is as follows: First, let the true orientation unit vector of the aircraft target be: ; in, The actual orientation angle of the aircraft target; Secondly, calculate the cosine similarity between the predicted orientation unit vector and the actual orientation unit vector: ; in, To predict the angle between the orientation and the actual orientation, the value range is: ; Finally, construct the orientation loss function with mirror penalty: ; in, Let be the mirror penalty coefficient, when (Right now )hour, The basic angle error loss is used; when (Right now When this occurs, it is determined that the prediction result carries a risk of polarity reversal. That is, the penalty amplification factor Setting the value to 10 amplifies the loss by 10 times for samples with reversed orientation, significantly increasing the gradient update magnitude for such samples and reducing the probability of the model flipping 180°.
[0028] Step 5: Model Training and Orientation Inference Model Training: Preprocessed SAR images and training labels were input into the constructed SAR target rotation detection network. Iterative training was performed using the AdamW optimization algorithm, with an initial learning rate of 1e-4, weight decay of 1e-5, batch size of 8, and a total of 100 training epochs. Cosine annealing was used to adjust the learning rate. During training, data augmentation strategies such as random flipping, random rotation, and multi-scale scaling, as well as online hard example mining, were employed to improve the model's generalization ability. Training was terminated early, and the optimal model weights were preserved, when the total loss on the validation set no longer decreased and the orientation discrimination accuracy no longer improved.
[0029] Orientation inference: After preprocessing the SAR image to be detected in the same way as during the training phase, it is input into the optimal model after training. The decoupled detection head outputs the aircraft target's class confidence, bounding box position parameters, and predicted orientation unit vector. ,pass The function solves for the unit vector to obtain the final aircraft target orientation angle: Map the calculated angle values to The interval represents the final 360° orientation value of the aircraft target, enabling rotation detection and 360° orientation discrimination of the aircraft target in the SAR image.
[0030] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A method for determining the orientation of an aircraft target in a SAR image based on deep learning, characterized in that, Includes the following steps: Step S1: Preprocess the acquired SAR image dataset, combine the true orientation information of the aircraft target to perform semantically ordered vertex encoding on the target labeled rotation box, and generate training labels embedded with orientation polarity information. Step S2: Construct a SAR target rotation detection network, which includes a cascaded feature extraction network, a feature fusion network, and a decoupled detection head; an attention enhancement module is embedded in the feature extraction network, and a high-resolution feature fusion branch is set in the feature fusion network; Step S3: Set up an independent orientation prediction branch in the decoupled detection head. The orientation prediction branch is completely decoupled from the target classification branch and bounding box regression branch of the detection head. The unit vector of the aircraft target orientation is regressed through the orientation prediction branch to realize the full-angle continuous expression of the target orientation. Step S4: Construct an orientation loss function with a mirror penalty mechanism. Based on the cosine similarity between the predicted orientation unit vector and the true orientation unit vector, apply a loss amplification penalty to the orientation prediction results that have reversed polarity. Step S5: Iteratively train the SAR target rotation detection network using the orientation loss function. Input the SAR image to be detected into the trained network and output the rotation detection result of the aircraft target and the 360° orientation information.
2. The method for determining the orientation of aircraft targets in SAR images based on deep learning according to claim 1, characterized in that: In step S1, the specific implementation process of semantically ordered vertex encoding includes: Step S11: The annotation information of aircraft targets in the SAR image is uniformly represented as a rotated bounding box. The parameterized expression of the rotated bounding box is: ； in, , These are the x-coordinate and y-coordinate of the center of the aircraft target rotation frame, respectively. The width of the rotating frame. The height of the rotating frame. The actual orientation angle of the aircraft target; Step S12: Calculate the coordinates of the four vertices of the rotating frame based on the parameters, determine the side where the nose is located by combining the actual orientation information of the aircraft target, and perform clockwise semantic sorting on the four vertices to generate an ordered vertex set: ； in, , These are the two vertices located on the side of the nose. , These are the two vertices located on the tail side of the aircraft; Step S13: Combine the category information of the aircraft target with the coordinate information of the ordered vertex set to generate training labels embedded with the nose-tail polarity orientation information.
3. The method for determining the orientation of aircraft targets in SAR images based on deep learning according to claim 1, characterized in that: In step S2, the feature extraction network adopts a deep convolutional neural network architecture, which extracts multi-scale features of the input SAR image through multi-layer convolution and downsampling operations; The attention enhancement module is embedded in the convolutional layer of the feature extraction network. It adopts the parameterless SimAM attention module, which enhances the features of the nose scattering center region in the feature map by calculating the neuron energy function. At the same time, it suppresses the feature weights of background coherent speckle and irrelevant regions, thereby strengthening the network's ability to perceive the differentiated features of the aircraft's nose and tail scattering structure.
4. The method for determining the orientation of aircraft targets in SAR images based on deep learning according to claim 1, characterized in that: In step S2, the feature fusion network adopts the FPN+PAN bidirectional fusion architecture. The high-resolution feature fusion branch adds a high-resolution feature layer with a downsampling rate of 2 times in the feature fusion network. The high-resolution feature layer and the multi-scale feature map in the bidirectional fusion architecture are cross-layer stitched and convolutionally fused to retain the fine-grained scattering structure information of small-scale aircraft targets in SAR images and improve the detection and orientation discrimination accuracy of small targets.
5. The method for determining the orientation of aircraft targets in SAR images based on deep learning according to claim 1, characterized in that: In step S3, the orientation prediction branch adopts a dual-channel convolution output structure. The output values of the two channels correspond to the two components of the target orientation unit vector, and the expression of the unit vector is: ； in, To predict the orientation of the unit vector, This refers to the aircraft target orientation angle predicted by the model.
6. The method for determining the orientation of aircraft targets in SAR images based on deep learning according to claim 1, characterized in that: In step S4, the construction process of the orientation loss function with mirror penalty mechanism includes: Step S41: Set the true orientation unit vector of the aircraft target as: ； in, For the true orientation unit vector, The actual orientation angle of the aircraft target; Step S42: Calculate the cosine similarity between the predicted orientation unit vector and the true orientation unit vector, expressed as: ； in, To predict the angle between the orientation and the actual orientation, Let cosine similarity be the similarity between two vectors. To predict the orientation of the unit vector, The predicted orientation angle of the aircraft target by the model; Step S43: Construct an orientation loss function with mirror penalty based on cosine similarity, the expression of which is: ； in, Towards the loss value, The mirror penalty coefficient; when The angle between the predicted orientation and the actual orientation hour, Training is performed using the base angle error loss; when The angle between the predicted orientation and the actual orientation hour, , A penalty amplification factor greater than 1 is used to amplify the loss of prediction results that have undergone polarity reversal, thereby increasing the gradient update magnitude of samples with reversed direction.
7. The method for determining the orientation of aircraft targets in SAR images based on deep learning according to claim 1, characterized in that: In step S5, the specific process of model training is as follows: the preprocessed SAR image and the corresponding training label are input into the SAR target rotation detection network, the network parameters are iteratively updated using a stochastic gradient descent optimization algorithm, and the loss function is minimized through backpropagation until the network converges, thus completing the model training; during the training process, multi-scale training and online hard example mining strategies are used to improve the generalization ability and robustness of the model.
8. The method for determining the orientation of aircraft targets in SAR images based on deep learning according to claim 1, characterized in that: In step S5, the specific process of target orientation inference is as follows: The SAR image to be detected is input into the trained SAR target rotation detection network. The decoupled detection head outputs the aircraft target category, rotation box position parameters, and predicted orientation unit vector. The predicted orientation unit vector is solved using the arctangent function to obtain the true orientation angle of the aircraft target, expressed as: ； in, The final output is the aircraft target's orientation value at all angles, with a value range of [value range missing]. or It completes the 360° all-angle orientation determination of the aircraft target.