A rotating target detection method based on dynamic reference axis

Abstract

The present invention relates to a rotating target detection method based on a dynamic reference axis, which decomposes the angle prediction problem originally based on a single reference axis into an angle prediction problem based on multiple reference axes, and transfers the boundary discontinuity problem to a stable region for processing through region division and angle conversion mechanisms. Through region division, the model can perform angle regression within the stable region, avoiding angle prediction errors near the boundary, significantly improving the accuracy of angle prediction, and the region division mechanism enables the model to better adapt to targets at different angles, enhancing the model's generalization ability in complex scenarios. The ability to flexibly switch between different reference axes further alleviates the boundary discontinuity problem and enhances the flexibility of angle prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision, and in particular to a rotating target detection method based on a dynamic reference axis. Background Art

[0002] Compared with horizontal target detection, directional target detection faces a major challenge: the boundary discontinuity problem caused by angular periodicity. Current solutions to this problem can be roughly divided into two categories:

[0003] The first type of methods focuses on position representation and models position through alternative representations. ,size and angles For example, the CSL method introduces a cyclic smooth label loss to transform the angle regression task into a classification task. DCL accelerates CSL by using dense codes, while GF-CSL stabilizes CSL through the use of dynamic weights. In addition, the sliding vertex method reformulates the rotation box prediction task as four-edge prediction relative to the horizontal box offset.

[0004] The second category of methods uses IoU-like loss to address the limitations of smooth L1 loss in accurately describing the relationship between the true box and the predicted box. For example, GWD and KLD transform the rotated box into a smooth Gaussian distribution and measure the distribution distance to maintain consistency with the IoU. KFIoU also utilizes Gaussian modeling but directly approximates the IoU, eliminating the hyperparameters required to measure the distribution distance. Oblique IoU derives the formula for calculating the IoU between rotated rectangles. The ACM module improves on the IoU-like loss by addressing the inconsistency between the target angle and the annotated box angle.

[0005] Existing angle regression methods still have shortcomings when dealing with boundary discontinuity problems. When dealing with complex scenes and multi-category targets, the stability and accuracy of the model based on position representation still need to be improved. For example, the CSL method may experience unstable predictions near angle boundaries. Although the IoU-like loss function alleviates the boundary discontinuity problem to a certain extent, in practical applications, the boundary discontinuity problem still exists because the regression model needs to predict highly divergent angle values ​​for similar features. In addition, existing position representation-based methods also face challenges in model stability and accuracy when dealing with complex scenes and multi-category targets.

[0006] In the regression head of the Oriented Object Detection (OOD) model, the input is the image features extracted from the proposals, and the output is denoted by a bounding box to represent the location. However, we observed that for objects close to angular boundaries, the regression head receives very similar image features but needs to predict two very different angle values. In contrast, in non-boundary cases, similar image features correspond to similar angle values. This difference causes the regression process to mutate near boundaries, especially near boundaries, where the regression head struggles to produce continuous and accurate angle predictions. As a result, the angle predictions for objects near boundaries become highly unstable during training, which significantly exacerbates the boundary discontinuity problem (BDP), resulting in low prediction accuracy and stability. Summary of the Invention

[0007] In view of the problems existing in the prior art, the purpose of the present invention is to provide a rotating target detection method based on a dynamic reference axis to improve the accuracy and stability of prediction.

[0008] To achieve the above object, the technical solution adopted by the present invention is:

[0009] A rotating target detection method based on a dynamic reference axis is implemented using a rotating target detection model. The rotating target detection model introduces a region partitioning mechanism to divide the angle domain into multiple regions, and classifies the region to which the target belongs through a region predictor, thereby performing angle regression within a stable region. The rotating target detection model also introduces an angle conversion module to map the reference angle based on the x-axis to the reference angle based on the y-axis, allowing the rotating target detection model to flexibly switch between different reference axes.

[0010] Specifically, the method includes a training phase and an inference phase. The training phase is divided into two phases, as follows:

[0011] The first stage: the backbone network and region proposal network of the rotation target detection model remain unchanged, the classification head remains unchanged, and the regression head is optimized: the regression head is used to predict the In addition, the model also predicts , where x, y represent the horizontal and vertical coordinates of the center point of the bounding box predicted by the model, w, h represent the width and height of the bounding box predicted by the model, and θ x Represents the rotation angle of the bounding box of the original prediction of the model based on the x-axis, θ y Represents the rotation angle of the bounding box predicted by the model based on the y-axis; and the regression head introduces an angle conversion module to adjust the angle value;

[0012] Specifically, the training set is input into the rotation target detection model, the backbone network and the region proposal network extract the feature vector, the feature vector is input into the classification head to obtain the category information, and the feature vector is input into the regression head to predict the position information. , the predicted Compare with the corresponding ground truth to calculate the loss of the trained model;

[0013] Phase 2: A region predictor is added to the regression head, and only its parameters are trained to ensure that the angle regression results obtained in the first phase are not affected by the second phase training. The region predictor uses the same structure as the classification head, except that the output is a one-dimensional binary classification, and the region predictor accurately divides the region.

[0014] Specifically, the feature vector is input into the region predictor to obtain the predicted region score , the region encoder proposes the true rotation angle Encode the corresponding area code to get the real area ; and according to the actual area and predicted regional scores Perform binary cross entropy loss calculation;

[0015] In the inference stage, the image of the target to be detected is input into the rotation target detection model. After the backbone network and the region proposal network extract the feature vector, the feature vector is passed through the region predictor to obtain the classification score, and the angle regression value is adjusted according to the score. The rotation target detection model finally outputs the category and location information of the target to be detected. As the test result.

[0016] The region division mechanism is as follows: the angle of the object bounding box can be limited to Within, select The axis is used as the reference axis, and the angle of the object is defined as one side of the y-axis, that is, the range of the rotation angle; the rotation angle range is divided into four areas in clockwise order, namely area 1, area 2, area 3 and area 4, and the angle range of each area is ; Region 2 and Region 3 are stable areas close to the reference axis, and Region 1 and Region 4 are areas where boundary discontinuity problems may occur; construct the y-axis as the second reference axis. At this time, Region 1 and Region 4 are stable areas close to the second reference axis. In order to unify Region 1 and Region 4, Region 4 is made centrally symmetric about the origin to obtain Region 5; at this time, Region 1 and Region 5 are continuous areas.

[0017] The regression head models position regression as a classification task with two regression components: 1) the region where the object appears, and 2) the position and angle regression values ​​of the object based on the x-axis, and the angle regression value based on the y-axis, as shown in the following equation:

[0018]

[0019] in, is the binary cross entropy loss, is the input image feature; Therefore The ground truth angle value with the x-axis as the reference axis; θ is the predicted angle of the bounding box by the original model when the x-axis is the reference axis. is the output of the region predictor, which is used to predict the region where the target is located; is the output of the region encoder, which is used to encode the ground truth angle into region information; , is the output of the regression head, which contains the angle prediction values ​​based on the x-axis and y-axis; is a distance metric used to measure the change between the predicted value and the ground truth value, It is an angle change function used to map the position parameter to the area position corresponding to the y-axis; The angle conversion module is used to convert The reference angle of the axis is mapped to the The reference angle of the axis.

[0020] The calculation formula for mapping the reference angle based on the x-axis to the reference angle based on the y-axis is as follows:

[0021]

[0022] in, Therefore The angle of the reference axis is obtained by changing The axis is the corresponding angle of the reference axis, and the original angle of area four is mapped to area five.

[0023] The specific steps of the reasoning stage are as follows:

[0024] Region classification: The feature vector passes through the region predictor to obtain the classification score of the region to which the target belongs;

[0025] Angle selection: According to the regional classification results, select the angle prediction value corresponding to the reference axis; if the target is classified as axis area, then retain the Axis prediction Otherwise, select Axis prediction , and is output after being processed by the angle conversion module.

[0026] Using the above solution, the present invention introduces a dynamic reference axis mechanism, dynamically adjusting the reference axis based on the approximate orientation of the target. This decomposes the angle prediction problem originally based on a single reference axis into one based on multiple reference axes. This dynamic reference axis mechanism maps regions prone to boundary discontinuities to stable regions along another reference axis, effectively alleviating the boundary discontinuity problem. Furthermore, within any angle region, at least one angle prediction is stable, significantly improving the model's robustness in complex scenarios.

[0027] Specifically, on the one hand, the present invention introduces a region division mechanism, divides the angle domain into multiple regions, and classifies the region to which the target belongs through a region predictor, so as to perform angle regression within a stable region. Through region division, the model can perform angle regression within a stable region, avoiding angle prediction errors near the boundary, significantly improving the accuracy of angle prediction, and the region division mechanism enables the model to better adapt to targets at different angles, enhancing the generalization ability of the model in complex scenarios. On the other hand, the present invention introduces an angle conversion module, mapping the reference angle based on the x-axis to the reference angle based on the y-axis, so that the model can flexibly switch between different reference axes, further alleviating the boundary discontinuity problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flow chart of the rotating target detection method of the present invention;

[0029] Figure 2 It is the different representations that may appear in angle regression under different circumstances;

[0030] Figure 3 It is a schematic diagram of the structure of training and reasoning of the present invention;

[0031] Figure 4 Schematic diagram of the detection results of an embodiment of the present invention. DETAILED DESCRIPTION

[0032] like Figure 1-3 As shown, Figure 2 (1) indicates that for targets close to the angle boundary, the angles are adjacent but exceed the boundary limit. Figure 2 (2) indicates that crossing the boundary causes a sudden change in angle, forcing the model to regress distinct angles from similar features. Figure 2 (3) It means that the boundary limitation is avoided by transforming the reference axis. The present invention discloses a rotating target detection method based on a dynamic reference axis, which transforms the original method based on a single reference axis (such as The angle prediction problem of the rotating target detection model is decomposed into the angle prediction problem based on multiple reference axes. On the one hand, a region partitioning mechanism is introduced to divide the angle domain into multiple regions, and the region to which the target belongs is classified through a region predictor, so that the angle regression is performed within the stable region. On the other hand, an angle conversion module is introduced to map the reference angle based on the x-axis to the reference angle based on the y-axis, so that the rotating target detection model can flexibly switch between different reference axes.

[0033] The present invention adopts the rotating target detection model to implement, specifically including the training phase and the inference phase. The training phase is divided into two stages, as follows:

[0034] Phase 1: The backbone network and region proposal network of the rotated object detection model remain unchanged, the classification head remains unchanged, and the regression head is optimized: regression head prediction In addition, the model also predicts Since the ground truth angle value is obtained in the preprocessing process The axis is standardized as the reference axis and annotated with a unified angle. Therefore, the regression head introduces an angle transformation module (ATM) to adjust the corresponding angle value. Where x, y represent the horizontal and vertical coordinates of the center point of the bounding box predicted by the model, w, h represent the width and height of the bounding box predicted by the model, and θ x Represents the rotation angle of the bounding box of the original prediction of the model based on the x-axis, θ y Represents the rotation angle of the model's predicted bounding box based on the y-axis.

[0035] The training process of the first stage is as follows: the training set is input into the rotation target detection model, the feature vector is extracted through the backbone network and the region proposal network, the feature vector is input into the classification head to obtain the category information, and the feature vector is input into the regression head to predict the position information. , the predicted The loss function can be any type, such as Smooth L1, KFIOU or GWD. According to the requirements of the selected loss function, the mapped target value can be adjusted accordingly after passing through the angle conversion module. There is a clear mapping relationship between them, and the two can use the same features for regression. In addition, from the perspective of computational overhead, the regression head only needs to add an extra output dimension, so the modification is very small.

[0036] Phase 2: A binary region predictor (RP) is added to the regression head, and only its parameters are trained to ensure that the angle regression results obtained in the first phase are not affected by the second phase. To minimize computational overhead, the region predictor uses the same structure as the classification head, but outputs a one-dimensional binary classification. Since the features extracted after the first phase of training are already sufficient for the classification head to correctly classify proposals, it is reasonable to assume that the same features can be used to accurately classify regions using the similarly structured region predictor.

[0037] Since the angle of the object bounding box can be constrained to So any straight line can be chosen as the reference axis ( Axis), take the reference axis as axis 1, define the angle of the object as one side of the y-axis, that is, the range of the rotation angle. We divide this angle range into four equal parts, then each angle range is , in clockwise order, the four regions are divided into Region 1, Region 2, Region 3 and Region 4. At this time, Region 2 and Region 3 are stable regions close to the reference axis, while Region 1 and Region 4 are regions where boundary discontinuity problems may occur. So the second reference axis is constructed ( Axis), take the reference axis as axis two, axis two is perpendicular to axis one, and is the boundary line of region two and region three. When axis two is taken as the reference axis, region one and region four are stable regions close to the reference axis, and region two and region three are regions where boundary discontinuity problems may occur. In order to unify region one and region four, region four can be made centrally symmetric about the origin to obtain region five. At this time, region one and region five are continuous regions, which is convenient for subsequent calculations. Therefore, under the joint action of the two reference axes, all regions can become stable regions. See the specific diagram. Figure 2 (3).

[0038] With commonly used This setting will Axis is used as the reference axis and the rotation angle range is . Angle boundary along Axis, called case 1, similarly, when the reference axis is axis, the angle boundary along Axis, called case 2. Figure 2 (1) as shown, use and As a dividing line, The angular domain is divided into regions and , where the area Equivalent to the area in terms of angle In case 1, the region and There will be no boundary discontinuity problem, which is a stable region. Similarly, in case 2, the region and It is a stable area.

[0039] The second stage of training is: the feature vector is input into the region predictor to obtain the predicted region score , the Region Encoder (RE) proposes the true rotation angle Encode the corresponding area code to get the real area (a 0 / 1 vector); and according to the real area and predicted regional scores Perform binary cross entropy loss calculation.

[0040] In this paper, the regression head models position regression as a classification task with two regression components: 1) the region where the object appears, and 2) the position and angle regression values ​​of the object based on the x-axis, and the angle regression value based on the y-axis, as shown in the following equation:

[0041]

[0042] in, is the input image feature; Therefore The ground truth angle value with the x-axis as the reference axis; θ is the predicted angle of the bounding box by the original model when the x-axis is the reference axis. is the output of the region predictor, which is used to predict the region where the target is located; is the output of the region encoder, which is used to encode the ground truth angle into region information; , is the output of the regression head, which contains the angle prediction values ​​based on the x-axis and y-axis; is a distance metric used to measure the change between the predicted value and the ground truth value, It is an angle change function used to map the position parameter to the area position corresponding to the y-axis. Angle Transfer Module (ATM) is used to convert The reference angle of the axis is mapped to the The reference angle of the axis, the specific formula is as follows:

[0043]

[0044] in, Therefore The angle of the reference axis is obtained by changing The axis is the corresponding angle of the reference axis, and the original angle of area four is mapped to area five.

[0045] The above two-stage training does not directly solve the boundary discontinuity problem; the predicted Still performs poorly near the boundary. However, within their respective regions, i.e. Figure 2 (3) Axis area and Their performance remains stable across all angular regions, which is consistent with previous observations. This suggests that a fully trained ARF-DRA has two key capabilities: 1) ensuring that at least one angle prediction is stable within any angular region; and 2) distinguishing the angular region to which the object belongs.

[0046] In the inference phase, the image of the target to be detected is input into the rotation target detection model. After the backbone network and the region proposal network extract the feature vector, the feature vector is passed through the region predictor to obtain the classification score, and the angle regression value is adjusted according to the score. The specific steps are as follows:

[0047] Region classification: The feature vector passes through the region predictor (RP) to obtain the classification score of the region to which the target belongs.

[0048] Angle selection: According to the regional classification results, select the angle prediction value corresponding to the reference axis. axis area, then retain the Axis prediction ; Otherwise, select based on Axis prediction , and after being processed by the angle conversion module, the rotation target detection model finally outputs the category and location information of the target to be detected As the test result.

[0049] The present invention maps the region prone to boundary discontinuity problems to the stable region of another reference axis by dynamically adjusting the reference axis. An important consideration is that Axis and Is there a boundary problem at the axis region division? From the perspective of the region classifier, this boundary represents a challenging situation because the classification results may be blurred due to the high similarity of features. However, the dividing line of the region boundary does not coincide with the actual angle discontinuity. This means that in this region, whether it is Axis or The regression values ​​of the angles remain stable with the axis as the reference. Therefore, even if a region is misclassified near the boundary, the impact on the model performance is small because and The difference between them is negligible and closely aligned with the ground truth. This mechanism effectively alleviates the boundary discontinuity problem and improves the stability and accuracy of angle prediction.

[0050] In summary, the key to the present invention is:

[0051] 1. Dynamic reference axis mechanism;

[0052] In the prior art, angle regression has discontinuity problems near the boundary, which makes it difficult for the model to stably predict the angle of the target near the boundary during training, thereby affecting the overall detection accuracy. The present invention introduces a dynamic reference axis mechanism to dynamically adjust the reference axis according to the approximate direction of the target, and decomposes the angle prediction problem originally based on a single reference axis into an angle prediction problem based on multiple reference axes. The dynamic reference axis mechanism maps areas prone to boundary discontinuity problems to the stable area of ​​another reference axis, effectively alleviating the boundary discontinuity problem, and at least one angle prediction is stable in any angle area, significantly improving the robustness of the model in complex scenarios.

[0053] 2. Regional division and classification;

[0054] In existing technologies, when processing targets near boundaries, the angle regression head is unstable because the image features are similar but the angle values ​​required to be predicted vary greatly, which in turn affects the model's generalization ability and detection accuracy. This invention introduces a region partitioning mechanism that divides the angle domain into multiple regions. The region predictor (RP) classifies the region to which the target belongs, thereby performing angle regression within a stable region. This region partitioning allows the model to perform angle regression within a stable region, avoiding angle prediction errors near boundaries and significantly improving angle prediction accuracy. Furthermore, the region partitioning mechanism enables the model to better adapt to targets at different angles, enhancing the model's generalization ability in complex scenarios.

[0055] 3. Angle Conversion Module (ATM)

[0056] Existing angle predictions based on a single reference axis suffer from discontinuities near boundaries, making it difficult for the model to stably predict the target angle during training and inference. This paper introduces an Angle Translation Module (ATM) that maps the x-axis reference angle to the y-axis reference angle. This allows the model to flexibly switch between different reference axes, further alleviating the boundary discontinuity issue.

[0057] The ATM module enables the model to flexibly switch between different reference axes, further alleviating the boundary discontinuity problem and enhancing the flexibility of angle prediction. In addition, the introduction of the ATM module only requires adding an additional output dimension to the regression head, which has low computational overhead and is easy to implement.

[0058] The present invention has the following beneficial effects:

[0059] (1) Significantly improve the angle prediction accuracy: Through the dynamic reference axis mechanism and the area division mechanism, the present invention ensures that at least one angle prediction in any angle area is stable, which significantly improves the angle prediction accuracy, especially in target detection near the boundary.

[0060] (2) Enhanced model robustness: The dynamic reference axis mechanism and region division mechanism enable the model to better adapt to targets at different angles in complex scenarios, significantly enhancing the robustness of the model and improving the reliability of the model in practical applications.

[0061] (3) Improve the generalization ability of the model: The region division mechanism enables the model to better adapt to targets at different angles, enhances the generalization ability of the model in complex scenarios, and enables the model to maintain high detection accuracy in different environments and scenarios.

[0062] (4) Simplifying the calculation process: The introduction of the angle transformation module (ATM) only requires adding an additional output dimension to the regression head, which has low computational overhead and is easy to implement, making the present invention highly efficient and scalable in practical applications.

[0063] (5) Strong compatibility: The framework of the present invention is theoretically compatible with most existing target detection models, has good scalability, and can be easily integrated into existing target detection systems, providing new technical means for research and application in related fields.

[0064] In order to better illustrate the effects achieved by the present invention, a simulation experiment will be used to further illustrate the present invention.

[0065] 1. Simulation conditions

[0066] The present invention was developed on the Ubuntu platform, and the deep learning framework developed was based on Pytorch. The main language used in the present invention was Python.

[0067] 2. Simulation Content

[0068] Take the DOTA dataset, a commonly used rotation target dataset in the field of remote sensing images, train the network according to the above steps and test it using the test set. Figure 4The test results of the present invention and other methods on the DOTA dataset show that the present invention consistently improves the performance of the baseline algorithm and achieves the best results compared to other methods. Since the names of various algorithms contain specific technical details, it is common practice in the industry to use English model names, which is not convenient to expand here. The detection performance of this method on the DOTA dataset reached 66.53%, which is higher than the detection performance of other methods, proving that the present invention has better results in rotating target detection.

[0069] The above description is merely an embodiment of the present invention and does not limit the technical scope of the present invention. Therefore, any minor modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A rotating target detection method based on a dynamic reference axis, characterized by: The rotating target detection model is implemented by introducing a region partitioning mechanism, which divides the angle domain into multiple regions and classifies the region to which the target belongs through a region predictor, thereby performing angle regression within a stable region. The rotating object detection model introduces an angle conversion module to map the reference angle based on the x-axis to the reference angle based on the y-axis, allowing the rotating object detection model to flexibly switch between different reference axes; Specifically, the method includes a training phase and an inference phase. The training phase is divided into two phases, as follows: The first stage: the backbone network and region proposal network of the rotation target detection model remain unchanged, the classification head remains unchanged, and the regression head is optimized: the regression head is used to predict the In addition, the model also predicts , where x, y represent the horizontal and vertical coordinates of the center point of the bounding box predicted by the model, w, h represent the width and height of the bounding box predicted by the model, and θ x Represents the rotation angle of the bounding box of the original prediction of the model based on the x-axis, θ y Represents the rotation angle of the bounding box predicted by the model based on the y-axis; and the regression head introduces an angle conversion module to adjust the angle value; Specifically, the training set is input into the rotation target detection model, the backbone network and the region proposal network extract the feature vector, the feature vector is input into the classification head to obtain the category information, and the feature vector is input into the regression head to predict the position information. , the predicted Compare with the corresponding ground truth to calculate the loss of the trained model; Phase 2: A region predictor is added to the regression head, and only its parameters are trained to ensure that the angle regression results obtained in the first phase are not affected by the second phase training. The region predictor uses the same structure as the classification head, except that the output is a one-dimensional binary classification, and the region predictor accurately divides the region. Specifically, the feature vector is input into the region predictor to obtain the predicted region score , the region encoder proposes the true rotation angle Encode the corresponding area code to get the real area ; and according to the actual area and predicted regional scores Perform binary cross entropy loss calculation; In the inference stage, the image of the target to be detected is input into the rotation target detection model. After the backbone network and the region proposal network extract the feature vector, the feature vector is passed through the region predictor to obtain the classification score, and the angle regression value is adjusted according to the score. The rotation target detection model finally outputs the category and location information of the target to be detected. As the test result.

2. The rotating target detection method based on a dynamic reference axis according to claim 1, characterized in that: The region division mechanism is as follows: the angle of the object bounding box can be limited to Within, select The axis is used as the reference axis, and the angle of the object is defined as one side of the y-axis, that is, the range of the rotation angle; the rotation angle range is divided into four areas in clockwise order, namely area 1, area 2, area 3 and area 4, and the angle range of each area is ; Region 2 and Region 3 are stable areas close to the reference axis, and Region 1 and Region 4 are areas where boundary discontinuity problems may occur; construct the y-axis as the second reference axis. At this time, Region 1 and Region 4 are stable areas close to the second reference axis. In order to unify Region 1 and Region 4, Region 4 is made centrally symmetric about the origin to obtain Region 5; at this time, Region 1 and Region 5 are continuous areas.

3. The rotating target detection method based on a dynamic reference axis according to claim 1, characterized in that: The regression head models position regression as a classification task with two regression components: 1) the region where the object appears, and 2) the position and angle regression values ​​of the object based on the x-axis, and the angle regression value based on the y-axis, as shown in the following equation: in, is the binary cross entropy loss, is the input image feature; Therefore The ground truth angle value with the x-axis as the reference axis; θ is the predicted angle of the original model for the bounding box when the x-axis is the reference axis. is the output of the region predictor, which is used to predict the region where the target is located; is the output of the region encoder, which is used to encode the ground truth angle into region information; , is the output of the regression head, which contains the angle prediction values ​​based on the x-axis and y-axis; is a distance metric used to measure the change between the predicted value and the ground truth value, It is an angle change function used to map the position parameter to the area position corresponding to the y-axis; The angle conversion module is used to convert The reference angle of the axis is mapped to the The reference angle of the axis.

4. The method for detecting a rotating target based on a dynamic reference axis according to claim 3, wherein: The calculation formula for mapping the reference angle based on the x-axis to the reference angle based on the y-axis is as follows: in, Therefore The angle of the reference axis is obtained by changing The axis is the corresponding angle of the reference axis, and the original angle of area four is mapped to area five.

5. The rotating target detection method based on a dynamic reference axis according to claim 1, characterized in that: The specific steps of the reasoning stage are as follows: Region classification: The feature vector passes through the region predictor to obtain the classification score of the region to which the target belongs; Angle selection: According to the regional classification results, select the angle prediction value corresponding to the reference axis; If the target is classified as belonging to axis area, then retain the Axis prediction ; Otherwise, select based on Axis prediction , and is output after being processed by the angle conversion module.

Images