Method for detecting and positioning rotating target by using KL divergence and horizontal labeling box

Abstract

The invention discloses a method for detecting and positioning a rotating target by using KL divergence and a horizontal labeling box. According to the method, a large number of data sets of an existing HBB labeling method are utilized, a weak supervision branch and a self-supervision branch are used for target prediction, the self-supervision branch uses KL divergence as an evaluation index, and sufficient data is used for training while the performance of a detector is not lost, so that the detector has high precision and high data fitting performance.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and remote sensing image processing, and relates to a rotating target detection method applied to remote sensing target detection, specifically a rotating target detection and localization method using KL divergence and horizontal bounding boxes. Background Technology

[0002] Object detection is one of the fundamental tasks in computer vision. However, targets in remote sensing images exhibit unique characteristics, such as large aspect ratios, dense arrangement, and arbitrary orientation. Particularly regarding arbitrary orientation, OBB-annotated detection methods have been proposed for rotating targets. This annotation method adds an angle parameter compared to HBB-based methods for horizontal target detection. Due to the periodicity of the angle parameter and the diversity of bounding box definitions, regression-based rotation detectors often encounter boundary discontinuities and square-like problems, which can mislead the detector into making highly inaccurate predictions. Furthermore, since the intersection-union ratio (IU) of rotated bounding boxes in remote sensing images is non-differentiable, some classic deep learning-based object detection methods in computer vision cannot be directly applied to the remote sensing field.

[0003] While horizontal object detection technology is relatively mature, in complex scenarios such as remote sensing imagery, finer-grained bounding boxes are needed to improve detector performance. Many object detection datasets are labeled with horizontal bounding boxes, but datasets specifically for remote sensing object detection are very scarce, and manually labeling datasets is extremely expensive. This scarcity of computer vision data in the remote sensing imagery field has led to stagnation and slow progress in detector performance in remote sensing. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a rotating target detection and localization method utilizing KL divergence and horizontal bounding boxes. This method enables the detector to achieve high accuracy and high data fitting performance. The method utilizes a large dataset of existing HBB annotation methods and employs both weakly supervised and self-supervised branches for target prediction. The self-supervised branch uses KL divergence as the evaluation metric. This approach trains the detector with sufficient data without sacrificing its performance, resulting in a detector with high accuracy and high data fitting performance.

[0005] The technical solution adopted by the present invention to solve the aforementioned technical problem is as follows:

[0006] In a first aspect, embodiments of this application provide a method for detecting and locating a rotated target using KL divergence and a horizontal bounding box, comprising the following steps:

[0007] First, acquire remote sensing images to obtain the position information of the bounding rectangle of the target in the remote sensing images, and then normalize the remote sensing images.

[0008] Then, multi-scale features are extracted from the remote sensing images. Finally, the obtained multi-scale features are fused according to the generation rules of the feature pyramid network structure.

[0009] Then, based on the multi-scale feature maps obtained through fusion, a weakly supervised branch generates rotated bounding boxes corresponding to all grid points in each feature map; a self-supervised branch also generates rotated bounding boxes corresponding to all grid points in each feature map. The prediction results of the weakly supervised branch and the self-supervised branch are used to mutually supervise each other to correct the angle.

[0010] Finally, the final prediction result of the weakly supervised branch is decoded to obtain the final rotation target.

[0011] In one possible implementation, a ResNet50 backbone network is used, and a BN layer is employed for normalization to extract feature maps C2, C3, C4, and C5 from the remote sensing images with downsampling factors of 4, 8, 16, and 32.

[0012] In one possible implementation, the multi-scale features are fused as follows:

[0013] According to the generation rules of the feature pyramid network structure, the obtained multi-scale features are fused to obtain multi-scale feature maps P2, P3, P4, and P5. An additional convolutional layer is added after P5 and a ReLU activation function is added before the additional convolutional layer to obtain feature map P6.

[0014] In one possible implementation, the weakly supervised (WS) branch includes a fully connected network and regression and classification subnetworks composed of feedforward neural networks; the regression subnetwork outputs a predicted rotated target box RBox, and the classification subnetwork outputs a predicted class and center-ness.

[0015] The predicted RBox is then converted into the corresponding minimum horizontal bounding rectangle HBox, which is used to calculate the total loss between the derived HBox and the GT Hbox (horizontal annotation bounding box) annotations. :

[0016]

[0017] (1)

[0018] in, The centrality of positive samples For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. The regression loss of the weakly supervised branch in equation (1) is... Classification loss Centrality loss They satisfy equations (2), (3), and (4) respectively:

[0019] (2)

[0020] (3)

[0021] (4)

[0022] in, Indicates the number of positive samples. and Let represent the probability distributions of each grid point (x, y) calculated by the Sigmoid function and the target category, respectively. and These represent the RBox and horizontal GT HBox predicted by the weakly supervised (WS) branch, respectively. and This represents the predicted center-ness and the true center-ness. This indicates that the Rbox will be converted to the corresponding Hbox.

[0023] In one possible implementation, the self-supervised branch shares network weights with the weakly supervised branch, but it only contains a regression subnetwork to predict the rotated target box (RBox). * This is used to calculate the loss between the RBox predicted by the weakly supervised branch:

[0024] (5)

[0025] (6)

[0026] in, The centrality of positive samples in the self-supervised branch. For indicator functions, when It is 1 if it is true, otherwise it is 0. For the grid point center of the self-supervised branch prediction, The rotated target box corresponding to the self-supervised branch predicted by the weakly supervised branch. For the rotated target box predicted by the self-supervised branch, , These are the rotation angles of the rotating target boxes for the weakly supervised branch and the self-supervised branch, respectively. , , It is a bounding box representation method designed to solve the problem of rotating target box boundaries. express(- ,-. , , ), express(- ,- , , ), express(- ,- , , The regression loss of the self-supervised branch in equation (5) satisfy:

[0027] (7)

[0028] in, , These are hyperparameters, with default settings of 0.15 and 1. , The center coordinates of the target bounding box are rotated to compensate for the loss of width, height, and angle.

[0029] In formula (7) satisfy:

[0030] (8)

[0031] IoU loss in equation (6) The IoU representation uses the transformation of the RBox into a 2D Gaussian distribution. Then use the KL divergence strategy:

[0032] (9)

[0033] M=(u,v) (10)

[0034] The RBox is converted to a 2D Gaussian distribution using equations (9) and (10), where R represents the rotation matrix and S represents the diagonal matrix of eigenvalues. Let (u, v) be the rotation angle of RBox, and (u, v) be the coordinates of the center point of RBox. The KL divergence formula is shown in equation (11):

[0035] (11)

[0036] The self-supervised branch regression subnetwork and the weakly supervised regression subnetwork share the same weight values. Therefore, the prediction results of the weakly supervised branch and the prediction results of the self-supervised branch are used to supervise each other to correct the angle.

[0037] In one possible implementation, for a selected grid point (x, y), the rotated target box (l, r, t, b) is predicted based on the weakly supervised branch trained under supervised supervision. Based on the prediction criteria of the FCOS network, the length of side AB of the target rotated rectangle ABCD is obtained as w = x + lr + ll, and the length of side BC is obtained as h = y + lt + lb. The coordinates of the center point of the rectangle are (w / 2, h / 2). Finally, according to the feature layer where the grid points are located, the rectangle ABCD is scaled to the size of the original remote sensing image to obtain the rotated rectangle (X, Y, W, H, ...). ).

[0038] Secondly, embodiments of this application provide a rotating target detection and localization system utilizing KL divergence and horizontal bounding boxes, comprising the following modules:

[0039] Image preprocessing module: Obtains the position information of the bounding rectangle of the target in the remote sensing image and performs normalization processing on the remote sensing image.

[0040] Feature extraction module: Used to extract multi-scale features from remote sensing images.

[0041] Feature fusion module: Based on the generation rules of the feature pyramid network structure, the obtained multi-scale features are fused.

[0042] The prediction module includes a weakly supervised branch and a self-supervised branch. Based on the fused multi-scale feature maps, it generates weakly supervised and self-supervised rotated bounding boxes for all grid points in each feature map. The prediction results from the weakly supervised branch and the self-supervised branch are used to mutually supervise each other to correct the angle.

[0043] Decoding module: Used to decode the final prediction result of the weakly supervised branch to obtain the final rotation target.

[0044] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory;

[0045] The memory is used to store computer programs;

[0046] When the processor executes the program stored in the memory, it implements any of the rotating target detection and positioning methods described in this application.

[0047] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the rotating target detection and positioning methods described in this application.

[0048] Fifthly, embodiments of this application provide a computer program product containing instructions that, when run on a computer, cause the computer to execute any of the rotating target detection and positioning methods described in this application.

[0049] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0050] 1. This invention uses horizontally labeled data from the DOTA dataset to train deep learning, which covers a wide range of datasets and eliminates the problem of scarce datasets for rotating object detection, resulting in good model generalization ability. Furthermore, it uses anchor-free FCOS as the detection head, leading to low model complexity, fast inference speed, and easy deployment.

[0051] 2. This invention uses KL divergence to surrogate the traditional IOU as loss, eliminating the need to consider parameter inconsistency issues, angle definition boundary issues, and square-like issues caused by angles in rotating target detection, thus greatly improving model accuracy. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other embodiments can be obtained based on these drawings.

[0053] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0054] Figure 2 This is a schematic diagram of the overall model structure of an embodiment of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art based on this application are within the scope of protection of this application.

[0056] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0057] like Figure 1 As shown, this application provides a method for rotating target detection and localization using KL divergence and horizontal bounding boxes, including:

[0058] The first step is to acquire remote sensing images, obtain the position information of the bounding rectangle of the target in the remote sensing images, and perform normalization processing on the remote sensing images;

[0059] The second step involves using ResNet50 as the backbone network and employing BN layers for normalization to extract feature maps C2, C3, C4, and C5 from remote sensing images with downsampling factors of 4, 8, 16, and 32.

[0060] The third step is to fuse features C2, C3, C4, and C5 according to the generation rules of the feature pyramid network structure to obtain multi-scale feature maps P2, P3, P4, and P5. An additional convolutional layer is added after P5 and a ReLU activation function is added before the additional convolutional layer to obtain feature map P6.

[0061] Step 4: Based on the multi-scale feature layer {P i |i=2,...,6}, generating rotated bounding boxes (l, r, t, b, ..., 6) for all grid points in each feature map through a weakly supervised branch. ).

[0062] like Figure 2 As shown, the weakly supervised (WS) branch includes a fully connected network and regression and classification subnetworks composed of feedforward neural networks; the regression subnetwork outputs the predicted rotated target box RBox, and the classification subnetwork outputs the predicted class and center-ness.

[0063] The predicted RBox is then converted into the corresponding minimum horizontal bounding rectangle HBox, which is used to calculate the total loss between the derived HBox and the GT Hbox (horizontal annotation bounding box) annotations. :

[0064]

[0065] (1)

[0066] in, The centrality of positive samples For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. The regression loss of the weakly supervised branch in equation (1) is... Classification loss Centrality loss They satisfy equations (2), (3), and (4) respectively:

[0067] (2)

[0068] (3)

[0069] (4)

[0070] in, Indicates the number of positive samples. and Let represent the probability distributions of each grid point (x, y) calculated by the Sigmoid function and the target category, respectively. and These represent the RBox and horizontal GT HBox predicted by the weakly supervised (WS) branch, respectively. and This represents the predicted center-ness and the true center-ness. This indicates that the Rbox will be converted to the corresponding Hbox.

[0071] Step 5: Based on the multi-scale feature layer {P i |i=2,...,6}, generate the rotated bounding box (l) for all grid points of each feature map through a self-supervised branch. * r * , t * b * , ).

[0072] The self-supervised branch shares network weights with the weakly supervised branch, but it only contains a regression subnetwork to predict the rotated target box (RBox). * This is used to calculate the loss between the RBox predicted by the weakly supervised branch:

[0073] (5)

[0074] (6)

[0075] in, The centrality of positive samples in the self-supervised branch. For indicator functions, when It is 1 if it is true, otherwise it is 0. For the grid point center of the self-supervised branch prediction, The rotated target box corresponding to the self-supervised branch predicted by the weakly supervised branch. For the rotated target box predicted by the self-supervised branch, , These are the rotation angles of the rotating target boxes for the weakly supervised branch and the self-supervised branch, respectively. , , It is a bounding box representation method designed to solve the problem of rotating target box boundaries. express(- ,-. , , ), express(- ,- , , ), express(- ,- , , The regression loss of the self-supervised branch in equation (5) satisfy:

[0076] (7)

[0077] in, , These are hyperparameters, with default settings of 0.15 and 1. , The center coordinates of the target bounding box are rotated to compensate for the loss of width, height, and angle.

[0078] In formula (7) satisfy:

[0079] (8)

[0080] IoU loss in equation (6) The IoU representation uses the transformation of the RBox into a 2D Gaussian distribution. Then use the KL divergence strategy:

[0081] (9)

[0082] M=(u,v) (10)

[0083] The RBox is converted to a 2D Gaussian distribution using equations (9) and (10), where R represents the rotation matrix and S represents the diagonal matrix of eigenvalues. Let (u, v) be the rotation angle of RBox, and (u, v) be the coordinates of the center point of RBox. The KL divergence formula is shown in equation (11):

[0084] (11)

[0085] The self-supervised branch regression subnetwork and the weakly supervised regression subnetwork share the same weight values. Therefore, the prediction results of the weakly supervised branch and the prediction results of the self-supervised branch are used to supervise each other to correct the angle.

[0086] Step 6: Decode the prediction results of the weakly supervised branch to obtain the final rotation target:

[0087] For the selected grid point (x, y), the rotated target box (l, r, t, b) is predicted based on the weakly supervised branch trained under supervised supervision. According to the prediction criteria of the FCOS network, the length of side AB of the target rotated rectangle ABCD is w = a + lr + ll, the length of side BC is h = y + lt + lb, and the coordinates of the center point of the rectangle are (w / 2, h / 2). Finally, based on the feature layer where the grid points are located, the rectangle ABCD is scaled to the size of the original remote sensing image to obtain the rotated rectangle (X, Y, W, H, ...). ), to complete the detection of rotating targets.

[0088] This application also provides a rotating target detection and localization system using KL divergence and horizontal bounding boxes, including the following modules:

[0089] Image preprocessing module: Obtains the position information of the bounding rectangle of the target in the remote sensing image and performs normalization processing on the remote sensing image.

[0090] Feature extraction module: Used to extract multi-scale features from remote sensing images.

[0091] Feature fusion module: Based on the generation rules of the feature pyramid network structure, the obtained multi-scale features are fused.

[0092] The prediction module includes a weakly supervised branch and a self-supervised branch. Based on the fused multi-scale feature maps, it generates weakly supervised and self-supervised rotated bounding boxes for all grid points in each feature map. The prediction results from the weakly supervised branch and the self-supervised branch are used to mutually supervise each other to correct the angle.

[0093] Decoding module: Used to decode the final prediction result of the weakly supervised branch to obtain the final rotation target.

[0094] In one possible implementation, the feature extraction module uses ResNet50 as the backbone network and employs BN layers for normalization to extract feature maps C2, C3, C4, and C5 of the remote sensing image with downsampling factors of 4, 8, 16, and 32.

[0095] In one possible implementation, the feature fusion module fuses features C2, C3, C4, and C5 according to the generation rules of the feature pyramid network structure to obtain multi-scale feature maps P2, P3, P4, and P5. An additional convolutional layer is added after P5 and a ReLU activation function is added before the additional convolutional layer to obtain feature map P6.

[0096] In one possible implementation, the weakly supervised branch of the prediction module is based on the multi-scale feature layer {P}. i For each feature map, generate the rotated bounding box (l, r, t, b, ..., 6) corresponding to all grid points. ).

[0097] The weakly supervised (WS) branch includes a fully connected network and regression and classification subnetworks composed of feedforward neural networks; the regression subnetwork outputs the predicted rotated target box RBox, and the classification subnetwork outputs the predicted class and center-ness.

[0098] The predicted RBox is then converted into the corresponding minimum horizontal bounding rectangle HBox, which is used to calculate the total loss between the derived HBox and the GT Hbox (horizontal annotation bounding box) annotations. :

[0099]

[0100] (12)

[0101] in, The centrality of positive samples For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. The regression loss of the weakly supervised branch in equation (12) is... Classification loss Centrality loss Satisfying equations (13), (14), and (15) respectively:

[0102] (13)

[0103] (14)

[0104] (15)

[0105] in, Indicates the number of positive samples. and Let represent the probability distributions of each grid point (x, y) calculated by the Sigmoid function and the target category, respectively. and These represent the RBox and horizontal GT HBox predicted by the weakly supervised (WS) branch, respectively. and This represents the predicted center-ness and the true center-ness. This indicates that the Rbox will be converted to the corresponding Hbox.

[0106] In one possible implementation, the self-supervised branch of the prediction module is based on the multi-scale feature layer {P}. i |i=2,...,6}, generate the rotated bounding box (l) corresponding to all grid points of each feature map. * r * , t * b* , ).

[0107] The self-supervised branch shares network weights with the weakly supervised branch, but it only contains a regression subnetwork to predict the rotated target box (RBox). * This is used to calculate the loss between the RBox predicted by the weakly supervised branch:

[0108] (16)

[0109] (17)

[0110] in, The centrality of positive samples in the self-supervised branch. For indicator functions, when It is 1 if it is true, otherwise it is 0. For the grid point center of the self-supervised branch prediction, The rotated target box corresponding to the self-supervised branch predicted by the weakly supervised branch. For the rotated target box predicted by the self-supervised branch, , These are the rotation angles of the rotating target boxes for the weakly supervised branch and the self-supervised branch, respectively. , , It is a bounding box representation method designed to solve the problem of rotating target box boundaries. express(- ,-. , , ), express(- ,- , , ), express(- ,- , , The regression loss of the self-supervised branch in equation (16) satisfy:

[0111] (18)

[0112] in, , These are hyperparameters, with default settings of 0.15 and 1. , The center coordinates of the target bounding box are rotated to compensate for the loss of width, height, and angle.

[0113] In formula (18) satisfy:

[0114] (19)

[0115] IoU loss in equation (17) The IoU representation uses the transformation of the RBox into a 2D Gaussian distribution. Then use the KL divergence strategy:

[0116] (20)

[0117] M=(u,v) (21)

[0118] The RBox is converted to a 2D Gaussian distribution using equations (20) and (21), where R represents the rotation matrix and S represents the diagonal matrix of eigenvalues. Let (u, v) be the rotation angle of RBox, and (u, v) be the coordinates of the center point of RBox. The KL divergence formula is shown in equation (22):

[0119] (twenty two)

[0120] The self-supervised branch regression subnetwork and the weakly supervised regression subnetwork share the same weight values. Therefore, the prediction results of the weakly supervised branch and the prediction results of the self-supervised branch are used to supervise each other to correct the angle.

[0121] In one possible implementation, the decoding module includes:

[0122] For the selected grid point (x, y), the rotated target box (l, r, t, b) is predicted based on the weakly supervised branch trained under supervised supervision. According to the prediction criteria of the FCOS network, the length of side AB of the target rotated rectangle ABCD is w = a + lr + ll, the length of side BC is h = y + lt + lb, and the coordinates of the center point of the rectangle are (w / 2, h / 2). Finally, based on the feature layer where the grid points are located, the rectangle ABCD is scaled to the size of the original remote sensing image to obtain the rotated rectangle (X, Y, W, H, ...). ), to complete the detection of rotating targets.

[0123] This application also provides an electronic device, which includes a processor and a memory;

[0124] The memory is used to store computer programs;

[0125] When the processor executes a program stored in the memory, it implements any of the methods described in this application.

[0126] In one possible implementation, the electronic device of this application embodiment further includes a communication interface and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0127] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not indicate that there is only one bus or one type of bus.

[0128] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0129] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0130] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0131] In another embodiment provided in this application, a computer-readable storage medium is also provided, which stores a computer program that, when executed by a processor, implements any of the methods described in this application.

[0132] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to perform any of the methods described in this application.

[0133] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0134] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0135] The various embodiments in this specification are described in a related manner. Each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other.

[0136] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. A rotation target detection and positioning method using KL divergence and horizontal annotation box, characterized in that, The steps include the following: Firstly, remote sensing images are acquired to obtain position information of a target bounding rectangle in the remote sensing images, and the remote sensing images are normalized; Then, multi-scale features of the remote sensing images are extracted; and the obtained multi-scale features are fused according to a generation rule of a feature pyramid network structure; Then, based on the fused multi-scale feature maps, a weakly supervised branch is used to generate a rotated target box corresponding to all grid points of each feature map, and a self-supervised branch is used to generate a rotated target box corresponding to all grid points of each feature map; and the prediction result of the weakly supervised branch and the prediction result of the self-supervised branch are mutually supervised to correct an angle; Finally, the final prediction result of the weakly supervised branch is decoded to obtain a final rotated target. 2.The method of claim 1, wherein, Resnet50 is used as a backbone network, a BN layer is used for normalization, and feature maps C2, C3, C4 and C5 of the remote sensing images with down-sampling multiples of 4, 8, 16 and 32 are extracted. 3.The method of claim 2, wherein, The multi-scale features are fused as follows: According to the generation rule of the feature pyramid network structure, the obtained multi-scale features are fused to obtain multi-scale feature maps P2, P3, P4 and P5, an additional convolution layer is added after P5, and a Relu activation function is added before the additional convolution layer to obtain a feature map P6.

4. The method of claim 1, wherein, The weakly supervised branch includes a full connection network and a regression and classification sub-network composed of a feedforward neural network; the regression sub-network outputs a predicted rotated target box RBox, and the classification sub-network outputs a predicted class and a center-ness; The predicted RBox is then converted to a corresponding minimum level horizontal bounding box HBox for computing the total loss between the derived HBox and GTHbox annotations : (1) wherein, is the center degree of the positive sample, is an indicator function, which is 1 when and 0 otherwise; the regression loss of the weakly supervised branch in equation (1) , the classification loss , and the center degree loss respectively satisfy equations (2), (3), and (4): (2) (3) (4) where, denotes the number of positive samples, and denote the probability distribution of each grid point (x, y) for each class calculated by the Sigmoid function and the target class, respectively, and denote the RBox and horizontal GT Hbox predicted by the weakly supervised branch, respectively, and denote the predicted center-ness and the true center-ness, denotes the operation of converting the Rbox to the corresponding Hbox.

5. The method of claim 4, wherein, The self-supervised branch shares network weights with the weakly supervised branch, but it only contains a regression subnetwork to predict the rotated bounding box RBox * for computing the loss between the RBox predicted by the weakly supervised branch: (5) (6) wherein, is the center degree of positive samples for the self-supervised branch, is an indicator function, when is 1, otherwise 0, is the grid point center predicted by the self-supervised branch, is the rotated target box predicted by the weakly supervised branch corresponding to the self-supervised branch, is the rotated target box predicted by the self-supervised branch, , are the rotation angles of the rotated target boxes predicted by the weakly supervised branch and the self-supervised branch, respectively; , , is a bounding box representation designed to solve the boundary problem of the rotated target box; represents (- , , , ), represents (- , , , ), represents (- , , , ) ; the regression loss of the self-supervised branch in formula (5) satisfies: (7) wherein, , are hyperparameters, and the default settings are 0.15 and 1, , respectively rotate the center coordinate loss of the target frame and the width, height, and angle loss. In formula (7) satisfies: (8) IoU loss in equation (6) The IoU representation uses the transformation of the RBox into a 2D Gaussian distribution. Then use the KL divergence strategy: (9) M=(u,v) (10) RBox is converted to 2D Gaussian distribution using formula (9) and formula (10), wherein R represents a rotation matrix, S represents a diagonal matrix of eigenvalues, is the rotation angle of RBox, (u, v) is the center point coordinate of RBox; the KL divergence formula is shown in formula (11): (11)。 6. The method of claim 1, wherein, For the selected grid point (x, y), according to the rotation target box (l, r, t, b, a) predicted by the weakly supervised branch of supervised training, ), based on the prediction criterion of the FCOS network, the length of the edge AB of the target rotating rectangular frame ABCD is w = x + lr + ll, the length of the edge BC is h = y + lt + lb, and the center point coordinates of the rectangle are (w / 2, h / 2). Finally, according to the feature layer where the grid point is located, the rectangular frame ABCD is scaled to the original remote sensing image size to obtain the rotating rectangular frame (X, Y, W, H, ).

7. A system for detecting rotating targets in remote sensing images using KL divergence and horizontal bounding boxes, characterized in that, The method comprises the following modules: An image preprocessing module: acquiring position information of a target bounding rectangle in remote sensing images, and normalizing the remote sensing images; A feature extraction module: used for extracting multi-scale features of the remote sensing images; A feature fusion module: according to a generation rule of a feature pyramid network structure, the obtained multi-scale features are fused; A prediction module: including a weakly supervised branch and a self-supervised branch, based on the fused multi-scale feature maps, a weakly supervised rotated target box and a self-supervised rotated target box corresponding to all grid points of each feature map are respectively generated; The prediction result of the weakly supervised branch and the prediction result of the self-supervised branch are mutually supervised to correct an angle; A decoding module: used for decoding the final prediction result of the weakly supervised branch to obtain a final rotated target.

8. An electronic device, comprising: A processor and a memory are included; The memory is used to store a computer program; The processor is used to execute the program stored in the memory to realize the rotated target detection and positioning method in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer program stored in the computer readable storage medium is executed by the processor to realize the rotated target detection and positioning method in any one of claims 1-6.

Images