A point-supervised rotating target detection method based on placement prior

Abstract

This invention proposes a point-supervised rotating target detection method based on placement priors. The method includes: dividing the original aerial imagery into a training set and a test set for directional target detection; obtaining pixel-level feature maps using the aerial imagery in the training set; obtaining directional bounding boxes corresponding to each pixel based on the pixel-level feature maps; constructing angle prior loss, size prior loss, and Voronoi watershed rotating bounding box loss based on a set of positive sample rotating bounding boxes, thereby constructing a training objective function; optimizing the target detection baseline model through backpropagation based on the training objective function, obtaining an optimized target detection baseline model; and inputting the test set into the optimized target detection baseline model to obtain the target detection results. This invention requires only single-point annotation to train an accurate rotating target detection model, significantly reducing annotation time and manpower costs compared to rotating bounding box annotation, and is suitable for large-scale dataset construction and rapid deployment scenarios.

Description

Technical Field

[0001] This invention relates to the field of processing technology, and in particular to a point-supervised rotating target detection method based on placement prior. Background Technology

[0002] Oriented target detection in aerial imagery is a core task in computer vision and remote sensing image processing. It aims to accurately locate targets facing any orientation within aerial imagery and output oriented bounding boxes. It is widely used in practical scenarios such as aerial scene analysis, UAV inspection, remote sensing monitoring, and autonomous driving. Compared to conventional horizontal target detection, oriented target detection more closely reflects the actual distribution of targets in aerial imagery, significantly improving the accuracy and effectiveness of target localization.

[0003] Traditional aerial imagery-based directional target detection often employs a fully supervised learning paradigm, relying on high-precision annotation of target orientation bounding boxes. This requires annotating multiple parameters for each target, such as center coordinates, width, height, and rotation angle. The annotation process is time-consuming, labor-intensive, and extremely costly. Furthermore, it is difficult to achieve accurate annotation of all large-scale aerial imagery data, severely limiting the training efficiency and practical deployment capabilities of directional target detection models.

[0004] To reduce annotation costs, point-supervised orientation target detection has become a research hotspot in this field. This paradigm only requires labeling each target with a point coordinate and corresponding category label, significantly reducing the annotation workload and providing a feasible path for the large-scale application of aerial image orientation target detection. However, the lack of target size and angle annotation information in the point-supervised mode has become a core bottleneck in improving detection accuracy. To compensate for this deficiency, existing point-supervised orientation target detection methods often estimate the target size and orientation by utilizing various supplementary information, such as pre-training to generate category probability maps to derive pseudo-rotated bounding boxes, constructing multi-view image transformations to learn target parameters, and combining watershed segmentation regularization for size estimation.

[0005] However, existing methods still have many inherent flaws: 1. The model has a complex structure, and operations such as pseudo-label generation and multi-view transformation require additional training stages, which increases the training difficulty and optimization cost of the model. 2. High computational overhead: Repeated image transformation and multi-stage training processes can lead to a significant increase in computational load, which places high demands on hardware resources and reduces the inference efficiency of the model. 3. Some methods rely on online angle and size estimation results from the model to derive supervision information, which is easily misled by predictions from immature models, resulting in insufficient accuracy of supervision information and affecting the final detection performance. These problems make it difficult for existing point-supervised directional target detection methods to achieve a balance between "low annotation cost" and "high detection accuracy and high computational efficiency," and the models have poor versatility and scalability, making it difficult to flexibly migrate to different detection architectures.

[0006] Targets in aerial imagery exhibit significant scene distribution characteristics; they are not randomly distributed but rather situated within a specific ground scene context. Targets within a neighborhood typically show aligned angles, and the size of targets of the same category exhibits a clear proportionality from a perspective viewpoint. These placement prior characteristics provide crucial supplementary information for size and angle estimation in point-supervised models. Therefore, fully leveraging the placement prior knowledge from aerial imagery to design lightweight, efficient, and universal loss supervision strategies, and constructing a point-supervised baseline model for directional target detection that requires no additional training phase and has low computational overhead, thereby reducing annotation costs while improving detection accuracy and efficiency, has become a critical issue urgently needing to be addressed in the field of directional target detection in aerial imagery. Summary of the Invention

[0007] In view of the above, the main objective of this invention is to propose a point-supervised rotating target detection method based on placement priors to solve the aforementioned technical problems.

[0008] This invention proposes a point-supervised rotating target detection method based on placement priors, the method comprising the following steps: Step 1: Perform directional target detection on the original aerial imagery and divide it into training and test sets; Step 2: Input the aerial images in the training set into the feature extraction module of the target detection baseline model, and extract features through the backbone network and feature pyramid network in sequence to obtain the pixel-level feature map of layer P3. Step 3: Based on the pixel-level feature map of layer P3, use the angle encoding module and detection head module of the target detection baseline model to obtain the oriented bounding box corresponding to each pixel; Step 4: Use the target point annotation information in the training set to assign positive samples to the oriented bounding boxes to obtain a set of positive sample rotated boxes. Step 5: Based on the set of positive sample rotating boxes, use the angle prior loss module, size prior loss module and Voronoi watershed rotating bounding box loss module of the target detection baseline model to construct the angle prior loss, size prior loss and Voronoi watershed rotating bounding box loss respectively. The training objective function is constructed based on angle prior loss, size prior loss, and Voronoi watershed rotation bounding box loss. Step 6: Based on the training objective function, optimize the target detection baseline model through backpropagation. After reaching the preset training conditions, the optimized target detection baseline model is obtained. Step 7: Input the aerial images from the test set into the optimized target detection baseline model to obtain the target detection results.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention only requires single-point annotation to train an accurate rotating target detection model, which greatly reduces annotation time and manpower costs compared to rotating bounding box annotation, and is suitable for large-scale dataset construction and rapid deployment scenarios.

[0010] 2. This invention simplifies the training process by eliminating complex pseudo-label generation, multi-view transformation, or external model dependence. It achieves single-stage end-to-end training through a simple loss term design, eliminating the need for repeated calculations, lowering the research threshold, and facilitating engineering implementation.

[0011] 3. This invention effectively regularizes model prediction through angle prior loss and size prior loss to compensate for the lack of point supervision information; by vectorizing the angle, the angle difference metric is constrained to the interval [0, π / 4) to adapt to parallel or vertical alignment scenarios.

[0012] 4. This invention provides additional angular supervision through a training-free rotating bounding box derivation method based on the Voronoi watershed region and the priority watershed strategy, and avoids the complex pre-training stage, thus achieving lightweight and efficient supervision signal generation. Attached Figure Description

[0013] Figure 1 This is a flowchart of the point-supervised rotating target detection method based on placement prior proposed in this invention; Figure 2 This is a schematic diagram of the overall framework of a simplified baseline model for the point-supervised rotating target detection method based on placement prior proposed in this invention. Figure 3 This is a schematic diagram of the angle prior loss of the point-supervised rotating target detection method based on placement prior proposed in this invention; Figure 4 This is a schematic diagram of the scale prior loss of the point-supervised rotating target detection method based on placement prior proposed in this invention; Figure 5 This is a schematic diagram of the Voronoi watershed rotating bounding box loss in the point-supervised rotating target detection method based on placement prior proposed in this invention. Detailed Implementation

[0014] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0015] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0016] Please see Figure 1 This invention proposes a point-supervised rotating target detection method based on placement priors, which includes the following steps: Step 1: Perform directional target detection on the original aerial imagery and divide it into training and test sets; In step 1, the original aerial images are subjected to directional target detection and divided into training set and test set. In the training set, each image is labeled with target point annotation information and a unique hot label of the category corresponding to the target point coordinates.

[0017] Step 2: Input the aerial images from the training set into the feature extraction module of the target detection baseline model, and extract features sequentially through the backbone network and the feature pyramid network to obtain the pixel-level feature map of layer P3.

[0018] Furthermore, the feature extraction module described in this step adopts a ResNet50 backbone network combined with a feature pyramid network (FPN) structure to perform multi-scale feature extraction on the input aerial image, and outputs a pixel-level feature map of layer P3 through the ResNet50 backbone network and the feature pyramid network in sequence. The target detection baseline model uses the AdamW optimizer with an initial learning rate of 5×10⁻⁵, a warm-up period of 500 iterations, and a learning rate decay factor of 10 at each scheduling point.

[0019] Step 3: Based on the pixel-level feature map of layer P3, use the angle encoding module and detection head module of the target detection baseline model to obtain the oriented bounding box corresponding to each pixel; In step 3, based on the pixel-level feature map of layer P3, the angular encoding module and detection head module of the target detection baseline model are used to obtain the oriented bounding box corresponding to each pixel. The specific steps are as follows: The pixel-level feature map of layer P3 is input into the angle encoding module of the target detection baseline model, and the encoded feature map is obtained by rotation angle encoding. The encoded feature map is input into the detection head module of the object detection baseline model, which predicts the oriented bounding box and class label for each pixel based on the encoded feature map.

[0020] Furthermore, the angle encoding module is connected to the feature extraction module, and a phase offset encoder (PSC) is used to continuously encode the rotation angle to avoid the problem of discontinuity at the angle boundary; The detection head module is connected to the angle encoding module and is used to output pixel-level directional bounding box prediction and category prediction at the P3 layer.

[0021] Step 4: Use the target point annotation information in the training set to assign positive samples to the oriented bounding boxes to obtain a set of positive sample rotated boxes.

[0022] Step 5: Based on the set of positive sample rotating boxes, use the angle prior loss module, size prior loss module and Voronoi watershed rotating bounding box loss module of the target detection baseline model to construct the angle prior loss, size prior loss and Voronoi watershed rotating bounding box loss respectively. The training objective function is constructed based on angle prior loss, size prior loss, and Voronoi watershed rotation bounding box loss. Please see Figure 3 and Figure 4 as well as Figure 5 In step 5, based on the set of positive sample rotated bounding boxes, the angle prior loss module, size prior loss module, and Voronoi watershed rotated bounding box loss module of the target detection baseline model are used respectively to construct the angle prior loss, size prior loss, and Voronoi watershed rotated bounding box loss. The specific steps are as follows: Vectorize the angles of each positive sample rotation box in the positive sample rotation box set to obtain an angle vectorized representation; Calculate the neighborhood weights of each positive sample rotation box using samples from the set of positive sample rotation boxes; Based on the angle vectorization representation and the neighborhood weights of the positive sample rotation box, the neighborhood alignment angle is calculated. An angle prior loss is constructed by vectorizing the angle and aligning it with the neighborhood; Based on the principle of perspective, a linear relationship equation between the width and position of each positive sample rotation box in the positive sample rotation box set is established; Under the condition of satisfying the linear relationship equation, the parameters to be solved are constructed into a parameter matrix; For each positive sample, rotate the bounding box to construct a feature row vector to obtain the feature row vector; stack all the feature row vectors to obtain the input feature matrix; The width values ​​of the rotated frames of each positive sample are stacked to obtain the size label matrix; Using the parameter matrix, input feature matrix, and size label matrix, ridge regression is used to solve for the optimal parameters. The size prior loss is constructed by using the optimal parameters, the input feature matrix, and the size label matrix. Using the target point annotation information in the training set as the foreground marker and the Voronoi Ridge as the background boundary, the target watershed region is calculated using the watershed algorithm. The prior angle is extracted from the bounding box of the target watershed region; Based on a priori perspective, the priority watershed algorithm is used to allocate directional priorities when growing or dividing the watershed region, so as to obtain an optimized watershed region. Using prior angles and target point annotation information, the target size is derived from the optimized enclosing rotating bounding box of the watershed region. By deriving the target size, the Voronoi watershed rotating bounding box loss is calculated using Gauss-Weststein distance loss. Furthermore, in this step, the angle prior loss module is connected to the detection head module and is used to calculate the neighborhood angle alignment loss of the positive sample rotation box and to regularize the angle prediction result of the model; that is, based on the characteristic in the placement prior that "the angles of adjacent target objects in aerial images are usually aligned", the estimated rotation angle of the positive sample rotation box in the neighborhood is subject to regularization constraint. The size prior loss module is connected to the detection head module and is used to calculate the size ratio loss of targets of the same category based on the perspective principle, and to regularize the size prediction results of the model; that is, based on the characteristic in the placement prior that "the size of target objects of the same category is usually proportional", the estimated size of the rotating box of positive samples of the same category is regularized and constrained by the perspective principle. The Voronoi watershed rotating bounding box loss module is connected to the detection head module and is used to derive the prior information of the rotating bounding box through the untrained watershed algorithm and calculate the overall loss of the rotating bounding box; that is, based on the Voronoi diagram and the watershed algorithm, the rotating bounding box target is derived from the point annotations without training, providing the model with supervision signals for size and angle.

[0023] Specifically, in the process of vectorizing the angles of each positive sample rotation box in the set of positive sample rotation boxes to obtain the vectorized representation of the angles, the following relationship exists: ; in, Representation of angle vectorization; The predicted rotation angle of the target instance is represented by the coefficient 4. This coefficient is used to adapt to the angle period of θ∈[0,2π) in rotating target detection, compressing the angle period from π to 2π, while eliminating angle ambiguity and aligning the period boundary. In the process of calculating the neighborhood weights of each positive sample rotation box using samples from the set of positive sample rotation boxes, the following relationship exists: ; in; This represents the neighborhood weights of the rotated bounding box for positive samples. Indicates the number of calculations to be performed. The rotating bounding box of a positive sample (target instance) is the main body of the formula. Indicates the first Rotated bounding boxes (target instances) of other positive samples within the neighborhood of a target are used to calculate neighborhood weights; Indicates the first The predicted center coordinates of a positive sample rotation box. Indicates the first The predicted center coordinates of the rotated bounding boxes of the neighboring positive samples. Indicates the first The predicted width of the rotated bounding box for each positive sample. Indicates the first The predicted height of the rotated bounding box for each positive sample; In the process of calculating the neighborhood alignment angle based on the angle vectorization representation and the neighborhood weights of the positive sample rotation box, the following relationship exists: ; in, Indicates the first The neighborhood alignment angle of a positive sample rotation box Indicates the first The angle of the rotated bounding box of each neighboring positive sample is vectorized. Indicates the first The predicted rotation angle of the rotating bounding box of each neighboring positive sample; In the process of constructing the angle prior loss by vectorizing the angle to align with the neighborhood, the following relationship exists: ; in, Indicates the angle prior loss. This represents the total number of positive sample rotation frames. This represents a hyperparameter for preventing division by zero. In the process of establishing a linear relationship between the width and position of each positive sample rotation box in the positive sample rotation box set based on the perspective principle, the following relationship exists: ; in, Represents the ground plane gradient. Indicates the first The size intercept of the target class, Indicates the first In the class of targets, the first The predicted center coordinates of a positive sample rotation box. Indicates the first In the class of targets, the first The predicted width of the rotated bounding box for each positive sample. This represents the linear rate of change of the target size with respect to its position along the x-axis of the image. This represents the linear rate of change of the target size with respect to its position along the y-axis of the image. Under the condition of satisfying the linear relationship equation, the following relationship exists during the process of constructing the parameters to be solved into a parameter matrix: ; in, Represents the parameter matrix, The total number is The target size intercept parameter of the class, Indicates the total number of categories. This represents the size intercept parameter for type 1 targets. This represents the size intercept parameter for type 2 targets. This represents the linear rate of change of the target size with respect to its position along the x-axis of the image. This represents the linear rate of change of the target size with respect to its position along the y-axis of the image. In the process of constructing the feature row vectors for each positive sample rotating bounding box, the following relationship exists: ; in, Indicates the first In the class of targets, the first The feature row vectors of the rotated bounding boxes of each positive sample are the input feature matrix. The row is used for ridge regression to fit a size-position linear relationship; This indicates a category 1 specific indicator, which takes a value of 1 when the target category is category 1, and 0 otherwise. This indicates a category 2 specific indicator, which takes a value of 1 when the target category is category 2, and 0 otherwise. This indicates a category-specific indicator, which takes a value of 1 when the target category is category C, and 0 otherwise. In the process of using the parameter matrix, input feature matrix, and size label matrix to solve for the optimal parameters through ridge regression, the following relationship exists: ; in, Indicates the optimal parameters. Represents the input feature matrix. Represents the size label matrix, This indicates the search for the parameter matrix that minimizes the objective function. , Represents the regularization coefficient. Represents the square of the Euclidean norm; In the process of constructing the size prior loss through optimal parameters, input feature matrix, and size label matrix, the following relationship exists: ; in, Indicates prior loss of dimensions; In the process of allocating directional priorities for watershed growth or segmentation based on prior knowledge using a priority watershed algorithm to obtain an optimized watershed region, the following relationship exists: ; in, This indicates the priority watershed algorithm function processing. This indicates the optimized watershed area. Indicates the first The watershed region after rotating a positive sample bounding box and optimizing it using the priority watershed algorithm. This represents the total number of positive sample rotated boxes in the input image. Represents training images, Indicates the first The coordinates of manually labeled points in the rotated bounding box of a positive sample. Indicates the first The prior angle of the rotating frame for each positive sample. This represents the sequence of point coordinates corresponding to all N target instances in the image. This represents the prior rotation angle sequence corresponding to all N target instances in the image; In the process of deriving the target size from the optimized bounding box of the watershed region using prior angles and target point annotation information, the following relationship exists: ; in, Indicates the first Derivation of the target size of a positive sample rotation frame This indicates the minimum bounding rectangle calculation operator processing. Indicates the first Derivation of the target height of a positive sample rotation frame Indicates the first Derivation of the target width of a positive sample rotation frame; In the process of deriving the target size and calculating the Voronoi watershed rotating bounding box loss using Gauss-Weststein distance loss, the following relationship exists: ; in, This indicates the loss of the rotated bounding box at the Voronoi watershed. Indicates the first The two-dimensional rotation matrix corresponding to the predicted angle of each positive sample rotation box. Indicates the first The two-dimensional rotation matrix corresponding to the prior angle of each positive sample rotation box; This represents the Gauss-Weststein distance loss, used to quantify the spatial difference between two rotated rectangles (model prediction box and prior supervision box), enabling overall supervision of the oriented bounding box; The training objective function is constructed based on angle prior loss, size prior loss, and Voronoi watershed rotation bounding box loss. The corresponding relationship in the process is as follows: ; in, Represents the training objective function. Represents classification loss. Indicates the loss at the center. All of these are loss balance hyperparameters.

[0024] Furthermore, to alleviate the uncertainty of point supervision signals, a relaxation mechanism is introduced in the early stages of training (the first 10% of loss values) to ignore extreme outlier loss values.

[0025] Step 6: Based on the training objective function, optimize the target detection baseline model through backpropagation. After reaching the preset training conditions, the optimized target detection baseline model is obtained.

[0026] Furthermore, this step optimizes the system parameters through backpropagation, where a relaxation factor is added to the top 10% of the loss terms.

[0027] Step 7: Input the aerial images from the test set into the optimized target detection baseline model to obtain the target detection results.

[0028] For further details, please refer to Figure 2 The point-supervised directional target detection baseline model constructed in this invention is a single-stage end-to-end deep learning system, mainly composed of the following core components: Feature extraction module: As the visual front end of the model, it is responsible for extracting multi-scale features with rich semantics from the input aerial images; this invention uses ResNet-50 as the backbone network, combined with a feature pyramid network structure; Angle encoding module: used to solve the problem of discontinuous angle boundaries in rotation detection. This invention uses a PSC angle encoder to map the angle to a continuous representation space; Detection Head Module: Based on the extracted features and encoded angle information, it outputs pixel-level rotated bounding box (RBox) predictions at the P3 layer. Each RBox contains center coordinates, width, height, rotation angle, and class probability. Angle Prior Loss Module: Based on the angle alignment characteristics in "placement prior", regularize the estimated angles of positive sample RBoxes in the neighborhood, making them tend to be parallel or perpendicular to each other; Size Prior Loss Module: Based on the size ratio characteristics in the "placement prior", the estimated size of the RBox of positive samples of the same class is regularized by the perspective principle so that it conforms to the linear variation law of ground projection.

[0029] The Voronoi Watershed Rotated Bounding Box Loss module (VWRBox Loss) generates a Voronoi map using point annotations, derives the exclusive region of the target object through a water-filling algorithm, and extracts a rotated bounding box from it as a strong supervision signal without the need for training.

[0030] Furthermore, during the inference phase of this invention, when aerial imagery of the test set is input, the model only performs forward propagation. The feature extraction module, angle encoding module, and detection head module work together to directly output the rotated bounding box parameters and class labels for each target; at this time, the angle prior loss module, size prior loss module, and Voronoi watershed rotated bounding box loss module are not enabled, thus not adding any additional computational overhead.

[0031] Specifically, this invention uses representative UAV datasets and remote sensing datasets for experimental analysis to verify detection performance in aerial imagery scenarios. These include the DroneVehicle dataset, the CO-Drone dataset, the DOTA-v1.0 dataset, and the DOTA-v1.5 dataset. Among them, DroneVehicle is a large-scale visible light-thermal imaging bimodal dataset, using only its visible light subset, containing 28,439 image pairs and 5 vehicle categories; the CO-Drone dataset contains 10,004 high-resolution aerial images, with boundary polygons labeled for 12 urban object categories, covering complex urban scenes; DOTA-v1.0 contains 2,806 high-resolution aerial images and 15 common object categories; DOTA-v1.5 adds one more object category on top of v1.0.

[0032] As shown in Table 1, the proposed method of this invention is compared with two types of existing technologies: one type is the fully supervised method that relies on expensive rotated bounding box annotations (such as YOLOv5s), representing the upper limit of performance; the other type is the point-supervised method that relies only on point annotations (such as the PointOBB series and Point2RBox series), which is a direct competitor of this invention. In Table 1, "Yes" indicates end-to-end training and testing, and "No" indicates that pseudo-labels need to be generated for two-stage training. It can be seen that the baseline of this invention achieves competitive results without the need for a complex pseudo-label generation strategy. More importantly, when the three loss terms proposed in this invention (angle prior, size prior, and Voronoi watershed rotated bounding box) are applied to the current state-of-the-art point-supervised methods (such as Point2RBox-v2), the performance is significantly improved.

[0033] Table 1. Quantitative comparison results of different methods on four aerial image datasets.

[0034] As shown in Table 1, the simple baseline of this invention achieves a mAP of 40.75% on the DroneVehicle dataset, outperforming PointOBB-v2 and approaching the performance of PointOBB-v3, demonstrating that effective point-supervised learning can be achieved solely through loss term constraints. Integrating the loss term of this invention into state-of-the-art methods improves the mAP from 27.53% to 29.12% on the CODrone dataset and from 54.06% to 55.47% on the OTA-v1.5 dataset. This indicates that the proposed prior loss term has broad applicability and can effectively supplement the shortcomings of existing methods.

[0035] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0036] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0037] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A point-supervised rotating target detection method based on placement prior, characterized in that, The method includes the following steps: Step 1: Perform directional target detection on the original aerial imagery and divide it into training and test sets; Step 2: Input the aerial images in the training set into the feature extraction module of the target detection baseline model, and extract features through the backbone network and feature pyramid network in sequence to obtain the pixel-level feature map of layer P3. Step 3: Based on the pixel-level feature map of layer P3, use the angle encoding module and detection head module of the target detection baseline model to obtain the oriented bounding box corresponding to each pixel; Step 4: Use the target point annotation information in the training set to assign positive samples to the oriented bounding boxes to obtain a set of positive sample rotated boxes. Step 5: Based on the set of positive sample rotating boxes, use the angle prior loss module, size prior loss module and Voronoi watershed rotating bounding box loss module of the target detection baseline model to construct the angle prior loss, size prior loss and Voronoi watershed rotating bounding box loss respectively. The training objective function is constructed based on angle prior loss, size prior loss, and Voronoi watershed rotation bounding box loss. Step 6: Based on the training objective function, optimize the target detection baseline model through backpropagation. After reaching the preset training conditions, the optimized target detection baseline model is obtained. Step 7: Input the aerial images from the test set into the optimized target detection baseline model to obtain the target detection results.

2. The point-supervised rotating target detection method based on placement prior as described in claim 1, characterized in that, In step 1, the original aerial imagery is subjected to directional target detection and divided into a training set and a test set. In the training set, each image is labeled with target point annotation information and a unique hot label for the category corresponding to the target point coordinates.

3. The point-supervised rotating target detection method based on placement prior as described in claim 2, characterized in that, In step 3, based on the pixel-level feature map of layer P3, the angular encoding module and detection head module of the target detection baseline model are used to obtain the oriented bounding box corresponding to each pixel. The specific steps are as follows: The pixel-level feature map of layer P3 is input into the angle encoding module of the target detection baseline model, and the encoded feature map is obtained by rotation angle encoding. The encoded feature map is input into the detection head module of the object detection baseline model, which predicts the oriented bounding box and class label for each pixel based on the encoded feature map.

4. The point-supervised rotating target detection method based on placement prior as described in claim 3, characterized in that, In step 5, based on the set of positive sample rotated bounding boxes, the angle prior loss module, size prior loss module, and Voronoi watershed rotated bounding box loss module of the target detection baseline model are used respectively to construct the angle prior loss, size prior loss, and Voronoi watershed rotated bounding box loss. The specific steps are as follows: Vectorize the angles of each positive sample rotation box in the positive sample rotation box set to obtain an angle vectorized representation; Calculate the neighborhood weights of each positive sample rotation box using samples from the set of positive sample rotation boxes; Based on the angle vectorization representation and the neighborhood weights of the positive sample rotation box, the neighborhood alignment angle is calculated. An angle prior loss is constructed by vectorizing the angle and aligning it with the neighborhood; Based on the principle of perspective, a linear relationship equation between the width and position of each positive sample rotation box in the positive sample rotation box set is established; Under the condition of satisfying the linear relationship equation, the parameters to be solved are constructed into a parameter matrix; Feature row vectors are constructed for each positive sample rotated bounding box to obtain the feature row vectors; Stack all feature row vectors to obtain the input feature matrix; The width values ​​of the rotated frames of each positive sample are stacked to obtain the size label matrix; Using the parameter matrix, input feature matrix, and size label matrix, ridge regression is used to solve for the optimal parameters. The size prior loss is constructed by using the optimal parameters, the input feature matrix, and the size label matrix. Using the target point annotation information in the training set as the foreground marker and the Voronoi Ridge as the background boundary, the target watershed region is calculated using the watershed algorithm. The prior angle is extracted from the bounding box of the target watershed region; Based on a priori perspective, the priority watershed algorithm is used to allocate directional priorities when growing or dividing the watershed region, so as to obtain an optimized watershed region. Using prior angles and target point annotation information, the target size is derived from the optimized enclosing rotating bounding box of the watershed region. By deriving the target size, the Voronoi watershed rotating bounding box loss is calculated using Gauss-Weststein distance loss.

5. The point-supervised rotating target detection method based on placement prior as described in claim 4, characterized in that, In the process of vectorizing the angles of each positive sample rotation box in the set of positive sample rotation boxes to obtain the vectorized representation of the angles, the following relationship exists: ； in, This represents the vectorized representation of angles. Represents the predicted rotation angle of the target instance; In the process of calculating the neighborhood weights of each positive sample rotation box using samples from the set of positive sample rotation boxes, the following relationship exists: ； in, This represents the neighborhood weights of the rotated bounding box for positive samples. Indicates the first The predicted center coordinates of a positive sample rotation box. Indicates the first The predicted center coordinates of the rotated bounding boxes of the neighboring positive samples. Indicates the first The predicted width of the rotated bounding box for each positive sample. Indicates the first The predicted height of the rotating frame for each positive sample.

6. The point-supervised rotating target detection method based on placement prior as described in claim 5, characterized in that, In the process of calculating the neighborhood alignment angle based on the angle vectorization representation and the neighborhood weights of the positive sample rotation box, the following relationship exists: ； in, Indicates the first The neighborhood alignment angle of a positive sample rotation box Indicates the first The angle of the rotated bounding box of each neighboring positive sample is vectorized. Indicates the first The predicted rotation angle of the rotating bounding box of each neighboring positive sample; In the process of constructing the angle prior loss by vectorizing the angle to align with the neighborhood, the following relationship exists: ； in, Indicates the angle prior loss. This represents the total number of positive sample rotation frames. This represents the hyperparameter for preventing division by zero.

7. The point-supervised rotating target detection method based on placement prior as described in claim 6, characterized in that, In the process of establishing a linear relationship between the width and position of each positive sample rotation box in the positive sample rotation box set based on the perspective principle, the following relationship exists: ； in, Represents the ground plane gradient. Indicates the first The size intercept of the target class, Indicates the first In the class of targets, the first The predicted center coordinates of a positive sample rotation box. Indicates the first In the class of targets, the first The predicted width of the rotated bounding box for each positive sample. This represents the linear rate of change of the target size with respect to its position along the x-axis of the image. This represents the linear rate of change of the target size with respect to its position along the y-axis of the image. Under the condition of satisfying the linear relationship equation, the following relationship exists during the process of constructing the parameters to be solved into a parameter matrix: ； in, Represents the parameter matrix, The total number is The target size intercept parameter of the class, Indicates the total number of categories. This represents the size intercept parameter for type 1 targets. This represents the size intercept parameter for type 2 targets. This represents the linear rate of change of the target size with respect to its position along the x-axis of the image. This represents the linear rate of change of the target size with respect to its position along the y-axis of the image. In the process of constructing the feature row vectors for each positive sample rotating bounding box, the following relationship exists: ； in, Indicates the first In the class of targets, the first Feature row vectors of a positive sample rotated bounding box This indicates a category 1 exclusive indicator. This indicates a category 2 specific indicator. This indicates a category-specific indicator.

8. The point-supervised rotating target detection method based on placement prior as described in claim 7, characterized in that, In the process of using the parameter matrix, input feature matrix, and size label matrix to solve for the optimal parameters through ridge regression, the following relationship exists: ； in, Indicates the optimal parameters. Represents the input feature matrix. Represents the size label matrix, This indicates the search for the parameter matrix that minimizes the objective function. , Represents the regularization coefficient. Represents the square of the Euclidean norm; In the process of constructing the size prior loss through optimal parameters, input feature matrix, and size label matrix, the following relationship exists: ； in, This represents the prior loss due to size.

9. The point-supervised rotating target detection method based on placement prior as described in claim 8, characterized in that, In the process of allocating directional priorities for watershed growth or segmentation based on prior knowledge using a priority watershed algorithm to obtain an optimized watershed region, the following relationship exists: ； in, This indicates the priority watershed algorithm function processing. This indicates the optimized watershed area. Indicates the first The watershed region after rotating a positive sample bounding box and optimizing it using the priority watershed algorithm. This represents the total number of positive sample rotated boxes in the input image. Represents training images, Indicates the first The coordinates of manually labeled points in the rotated bounding box of a positive sample. Indicates the first The prior angle of the rotating frame for each positive sample. This represents the sequence of point coordinates corresponding to all N target instances in the image. This represents the prior rotation angle sequence corresponding to all N target instances in the image; In the process of deriving the target size from the optimized bounding box of the watershed region using prior angles and target point annotation information, the following relationship exists: ； in, Indicates the first Derivation of the target size of a positive sample rotation frame This indicates the minimum bounding rectangle calculation operator processing. Indicates the first Derivation of the target height of a positive sample rotation frame Indicates the first Derivation of the target width of a positive sample rotation frame; In the process of deriving the target size and calculating the Voronoi watershed rotating bounding box loss using Gauss-Weststein distance loss, the following relationship exists: ； in, This indicates the loss of the rotated bounding box at the Voronoi watershed. Indicates the first The two-dimensional rotation matrix corresponding to the predicted angle of each positive sample rotation box. Indicates the first The two-dimensional rotation matrix corresponding to the prior angle of each positive sample rotation box. This represents the Gauss-Weststein distance loss.

10. The point-supervised rotating target detection method based on placement prior as described in claim 9, characterized in that, In step 5, the training objective function is constructed based on the angle prior loss, size prior loss, and Voronoi watershed rotation bounding box loss. The corresponding relationship in this process is as follows: ； in, Represents the training objective function. Represents classification loss. Indicates the loss at the center. All of these are loss balance hyperparameters.

Images