A remote sensing image rotation target detection method and system

Abstract

The application belongs to the technical field of image processing, and particularly relates to a remote sensing image rotating target detection method and system, which comprises a preprocessing module, a feature extraction module, a multi-scale feature fusion module, a remote sensing image detection module and a spliced original image output module. The method disclosed in the application realizes a new method for remote sensing image rotating target detection and segmentation, innovatively uses adaptive group rotating deformation convolution to extract the direction features of the remote sensing image, simultaneously uses dynamic multi-scale feature fusion to improve the perception of different size targets, and improves the remote sensing image rotating target detection accuracy while ensuring the prediction accuracy of each step.

Description

Technical Field

[0001] This application belongs to the field of image processing technology, and specifically relates to a method and system for detecting rotating targets in remote sensing images. Background Technology

[0002] Remote sensing image detection is a subfield of computer vision, specifically designed to address the challenges of targets in remote sensing scenarios, such as arbitrary angle rotation, large scale differences, and complex backgrounds. Rotated target detection focuses on accurately locating and classifying angled targets, often used in conjunction with other techniques for the refined analysis of typical remote sensing targets such as ships, aircraft, and bridges. Traditional horizontal bounding boxes (HBB) often suffer from incomplete or redundant enclosures when dealing with targets at different angles, especially those arranged consecutively. In contrast, rotated bounding boxes (OBB) can predict angles and accurately enclose targets, making them commonly used for detecting closely packed targets with varying orientations, such as aircraft, ships, and cars, in remote sensing images.

[0003] In recent years, with the rapid development of computer vision, methods to improve rotating target detection have been proposed. However, the accuracy of OBB angle prediction for multi-angle target problems has always been difficult to improve. Feature extraction methods have evolved from convolution to transformer and dynamic convolution. Therefore, designing a good feature extraction method to extract features of different directions and sizes is a worthy research direction. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method and system for detecting rotating targets in remote sensing images. By combining the dynamic feature extraction capabilities of adaptive convolution with an attention mechanism, rotating target detection in remote sensing images is achieved.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: A method for detecting rotating targets in remote sensing images includes the following steps: S1. Receive the image file to be detected and preprocess the image; S2. Feature extraction; S3. Perform dynamic feature fusion on the features output by the feature pyramid to obtain multi-scale fused features; S4. Use fused features to detect rotated targets and output rotated bounding boxes.

[0006] Preferably, the specific method for step S1 is as follows: S11. Upload the remote sensing image to be detected; S12. Select a suitable remote sensing image scene and appropriate model parameters; S13. Perform image detection; S14. Perform data preprocessing on each of the above images: cut or fill to the model input size. .

[0007] Preferably, the feature extraction method in step S2 is as follows: S21. Image preprocessing; S22. Rotational Deformation Convolution: Text location detection is performed using the ResNet-AGRDC model. The input image is processed through a backbone network consisting of five ResNet50 stages replaced with AGRDConv blocks for feature extraction. S23. Attention Adjustment: Each AGRDConv stage uses spatial attention and channel attention mechanisms to adjust the weights of feature maps within different groups based on the obtained feature maps.

[0008] Preferably, image preprocessing S21 includes: calculating the grayscale histogram of the image and calculating the cumulative histogram of the image; The mapping relationship between input gray levels and output gray levels is defined as follows: ; in The input image pixel gray levels, To output the pixel grayscale levels of the image, The grayscale value is The total number of pixels, The height of the input image. Given the width of the input image, the range is enlarged to between 0 and 255 to obtain the pixels of the output image. The gray level of each pixel in the image is output cyclically according to the mapping relationship.

[0009] Preferably, step S22, rotational deformable convolution, includes: Step 1: Initialize a set of Gaussian convolution kernels, totaling... Groups, each group has n independent convolutional kernels, all of which have a size of [missing information]. ; For the width and height of the convolution kernel, and The original input / output channel dimensions. The number of groups is [number], and the input dimension is [dimension]. The features are processed by depthwise convolution to Size, flattened A one-dimensional vector, after passing through a shared fully connected layer, enters a two-branch fully connected layer. Each branch fully connected layer then passes through another two-branch fully connected layer, and the output is processed by either the Softsign activation function or the Sigmoid activation function. The data sets are as follows: Group rotation angle and Group convolution weights , An offset, Each offset weight; Perform group rotation convolution on each group: based on the predicted angle For convolution kernel and offset The matrix rotation operation is performed using the following formula: ; For batches, For width, For height, Number of channels; Represents the rotated convolution kernel. This represents the offset after rotation. Represents the initial convolution kernel. Represents the rotation angle. For matrix rotation functions, Represents the initial offset; Step Two, after obtaining Group rotated convolution kernel { , , ……, After that, the weight parameters obtained through learning The weighted fusion is performed using the following formula: ; in, The learned convolutional kernel weights, , The final convolution kernel after weighted rotation ; Step 3: Perform a convolution operation using the newly calculated convolution kernel, as shown in the following formula: ; in The input feature map has a size of [size missing]. , The output feature map has a size of [size missing]. , For group convolution operations, For the sampling point index of the convolution kernel, This represents the offset of the convolution kernel sampling points. The weights of the sampling points of the convolution kernel. This is the final convolution kernel for step two.

[0010] Preferably, step S23, attention adjustment, is as follows: Step 1: For the output feature map from the previous step Size is Two feature maps are obtained by pooling the width and height separately using global pooling. and The dimensions are respectively and The two feature maps are then concatenated to obtain... Feature map of size ; Step 2: After extracting features through convolution, convert the feature map... Divided into two feature maps and The dimensions are respectively and Multiply their width and height dimensions to get Weighted graph ; Step 3: Weighted Graph Compared with the original feature map Multiplying the corresponding positions together, we get the final result. Feature map ; Step Four, Process After steps S22 and S23, four feature maps are output. .

[0011] Preferably, step S3 is as follows: S31 Feature Routing: The feature map size is adjusted to the same value using 3×3 convolution and deconvolution. Same size, that is Then, spatial attention is used to adjust the weight distribution of the convolutional feature maps; S32 Feature Fusion: First, regarding feature maps By splicing according to the channel dimensions, the resulting size is... Feature map Features are extracted using global average pooling to obtain the weight vector. The formula is: ; in This represents the value of the feature map in the coordinates (width, height, dimension). Represents the weight vector In the index The value on; Then It is a one-dimensional vector with size . After feature extraction through two fully connected layers, the fusion weights are obtained through the Softmax activation function. The formula for this process is as follows: ; Let be the weight vector, with size . , For input / 2 output is The fully connected layer Input 4 The output is / 2 fully connected layer; Finally, the weights Weighted to The final feature map is obtained after summing. The size is .

[0012] Preferably, step S4 is as follows: In the remote sensing image detection stage, the Oriented-RCNN architecture is used to process the image features and output the detection structure. The Oriented-RCNN model includes three parts: RPN, ROI, and bboxhead; RPN stage: Feature map output by S32 Each feature point generates 9 Anchors, totaling... ,for Generate one feature point A score is used to predict whether the anchor belongs to the foreground or background, and The data points are used as Anchor coordinates, then sorted by score, and the top N input ROIs are selected. The loss in this stage includes two types: classification loss and Anchor regression loss, as shown in the following formula: ; ; ; ; ; ; ; ; in , These are the coordinates of the center point of the target bounding box. , The target bounding box width and height, , , , These are the coordinates of the anchor's center point and its width and height, respectively. These are the coordinates and width / height offsets of the horizontal bounding box. For RPN loss, Foreground and background classification loss, The loss function for the horizontal bounding box. For smoothing functions, For feature weight hyperparameters, For predicting categories, Based on the prediction , , , Calculated ; To be based on reality , , , Calculated z is the independent variable of the function.

[0013] ROI Phase: First, predict a rough rotation angle. Mapping the rotated target bounding box to the feature map The above steps involve obtaining features of a fixed size according to a proportional ratio. The loss during this process is the angle prediction loss, calculated using the following formula: ; in, To predict the target box rotation angle, The actual target bounding box rotation angle. Will be regulated in , This refers to ROI loss; Bbox head phase: For each Anchor, there is... Features, flattened A long one-dimensional vector passes through two fully connected layers, then through two branch fully connected layers, one of which passes through... get The first branch outputs a vector representing the classification result; the second branch outputs 3D variables, where 4 variables represent the horizontal bounding box offset and 8 variables represent the rotated bounding box offset. The loss at this stage mainly includes classification loss, horizontal regression loss, and rotational regression loss, as shown in the following formulas: ; ; ; ; ; in For the real category, For predicting categories, This represents the actual horizontal bounding box offset. To predict the horizontal target box offset, It is the offset of the actual rotated bounding box, that is, the difference between the four coordinate points of the actual bounding box and the horizontal target bounding box. To predict the rotation box offset, which is the difference between the four coordinate points of the predicted box and the horizontal target box, To detect the loss, For classifying losses, For bounding box regression loss, For horizontal frame loss, For rotation frame loss, is the feature weight hyperparameter.

[0014] Preferred, including: In the post-processing stage, for the bounding boxes corrected by S32, the class with the highest classification score is selected. Those with scores below a set threshold are discarded directly. Different classes are sorted in descending order of score, and the class with the highest score is selected as the baseline class. Those with an intersection-union ratio (IU) exceeding the baseline class threshold are discarded, while those with a IU below the threshold are temporarily retained. Then, the filtering is iterated until all boxes are processed, and only the last remaining detection boxes are retained. In the image stitching stage, the images and detection results are stitched together in reverse according to the cropping method in S14 to obtain the final result, and the different formats of the result are output to the front-end interface.

[0015] A remote sensing image rotation target detection system includes a preprocessing module, a feature extraction module, a multi-scale feature fusion module, a remote sensing image detection module, and a stitched original image output module; Preprocessing module: Performs cropping operations on the image; Feature extraction module: Uses an adaptive group rotation deformable convolutional network to extract features from the image and obtain a feature map pyramid; Multi-scale feature fusion module: performs dynamic feature fusion on the features output by the feature pyramid to obtain multi-scale fused features; Remote sensing image detection module: fuses feature maps to perform rotating target detection and recognition on the image to be detected, and outputs the rotating target bounding box; Original Image Stitching Output Module: Outputs the stitched original image.

[0016] Compared with the prior art, the beneficial effects of this application are as follows: The method disclosed in this invention realizes a new method for detecting and segmenting rotating targets in remote sensing images. It innovatively uses adaptive group rotation deformation convolution to extract the directional features of remote sensing images, and uses dynamic multi-scale feature fusion to improve the perception of targets of different sizes. While ensuring the prediction accuracy of each step, it improves the accuracy of detecting rotating targets in remote sensing images. Attached Figure Description

[0017] Figure 1This is a schematic diagram of a remote sensing image rotation target detection method and system disclosed in an embodiment of the present invention.

[0018] Figure 2 This is a diagram of the AGRDConv block module architecture used in the feature extraction stage of this invention.

[0019] Figure 3 This is an architecture diagram of the Angle prediction module used in the feature extraction stage of this invention.

[0020] Figure 4 This is an architecture diagram of the attention module used in the feature extraction stage of an embodiment of the present invention.

[0021] Figure 5 This is an architecture diagram of the dynamic feature fusion module used in the rotating target detection stage of this invention.

[0022] Figure 6 This is a feature routing architecture diagram of the dynamic feature fusion module used in the rotating target detection stage of this embodiment of the invention.

[0023] Figure 7 This is a feature fusion architecture diagram of the dynamic feature fusion module used in the rotating target detection stage of this embodiment of the invention.

[0024] Figure 8 This is a diagram of the Oriented-RCNN module architecture used in the rotating target detection stage of this invention. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0026] A method for detecting rotating targets in remote sensing images includes the following steps: (1) Preparation stage: (1.1) Upload the remote sensing image to be detected; (1.2) Select the appropriate remote sensing image scene and suitable model parameters through the front-end system; (1.3) The images and model parameters are uploaded to the backend GPU server for image detection; (1.4) Perform a series of data preprocessing tasks for each image uploaded to the server: cut or fill to the model input size, i.e., N*N (N can be specified, and redundant cropping is used for large images). (2) Feature extraction stage: (2.1) Image preprocessing: Due to the inconsistent conditions under which remote sensing images are captured, a series of problems such as blurring may occur, which will affect the accuracy of remote sensing target detection. Therefore, image preprocessing is required, mainly implemented through the Python OpenCV library, including the following steps: Image enhancement: 1) First, calculate the gray-level histogram of the image using the `hist` function in the Matplotlib library, and then calculate the cumulative histogram of the image. 2) Define the mapping relationship between the input gray levels and the output gray levels as follows: ; in For input pixels, To output pixels, the above mapping relationship is equivalent to first obtaining the cumulative probability histogram of the grayscale histogram (i.e., ), and then expand this range to between 0-255 to obtain the pixels of the output image. 3) Circularly output the grayscale level of each pixel in the image according to the mapping relationship. Through the above steps, histogram equalization is achieved, which solves the problem of low contrast caused by the small grayscale range of the image, making the text clearer, thereby further improving the recognition effect.

[0027] (2.2) In the rotational deformation convolution stage, the ResNet-AGRDC model is used for text location detection. The input image is processed by a backbone network consisting of five stages of ResNet50 blocks replaced by ARDConv blocks for feature extraction. The ARDConv blocks are shown in the attached figure. Figure 2 As shown, Gaussian initialization first initializes a set of convolutional kernels, a total of G sets, each set having n independent convolutional kernels, all with a size of [missing information]. The angle prediction module on the right (as shown in the attached image) Figure 3 (As shown) It shares the same input features as the rotation-deformation convolution module, with an input dimension of . ( The features are processed by depthwise convolution. Size, flattened A one-dimensional vector, after passing through a shared fully connected layer, enters a two-branch fully connected layer. Each branch fully connected layer then passes through another two-branch fully connected layer, and the output is processed by either the Softsign activation function or the Sigmoid activation function. The data sets are as follows: Group rotation angle and Group convolution weights , An offset, Each group has an offset weight, and a group rotation convolution is performed on each group. Specifically, this operation is based on the predicted angle. For convolution kernel and offset The matrix rotation operation is performed using the following formula: ; in Represents the rotated convolution kernel. This represents the offset after rotation. Represents the initial convolution kernel. Represents the rotation angle. For matrix rotation functions, This represents the initial offset. In obtaining Group rotated convolution kernel { , , ……, After that, the weight parameters obtained through prediction are... We then perform weighted fusion to obtain the final convolutional kernel specific to this set of features. The formula is as follows: ; in, ( The learned convolutional kernel weights are used to measure the weights of each rotated convolutional kernel. The importance of a rotational convolution kernel in feature extraction is such that the larger the weight value, the greater its contribution to the final feature extraction.

[0028] Finally, the newly calculated convolution kernel is used for convolution operation, where W is the final convolution kernel after weighted rotation: ; in The input feature map has a size of [size missing]. , The output feature map has a size of [size missing]. , For group convolution operations, For convolution kernel, This represents the coordinate offset corresponding to the convolution kernel. These are the weights of the convolution points.

[0029] (2.3) In the attention adjustment stage, each AGRDConv stage uses spatial attention and channel attention mechanisms to adjust the weights of feature maps within different groups based on the obtained feature map, as shown in the appendix. Figure 4 As shown, the feature distribution is first adjusted through a lightweight spatial and channel attention mechanism. Specifically, the input... The size is still the same ( By performing global pooling on the width and height separately, two feature maps are obtained. and The dimensions are respectively and The two feature maps are then concatenated to obtain... Feature map of size Subsequently, features are extracted through convolution, and the feature map is then segmented into... and Multiply their width and height dimensions to get Weighted graph Then the weight map and the original feature map Multiplying the corresponding positions together, we get the final result. Feature map .

[0030] Steps 2.2 and 2.3 above together replace the standard convolutions in the ResNet50 blocks. The feature maps output by stage 1, stage 2, stage 3, and stage 4 are respectively... Input dimensions are hour, for , for , for , for .

[0031] (3) Multi-scale feature fusion stage: (3.1) Feature routing module, the structure of which is shown in the appendix. Figure 5 As shown, the feature map size is adjusted to the same size as the image size using 3×3 convolution and deconvolution. Same size, that is Then, spatial attention is used to adjust the weight distribution of the convolutional feature maps; (3.2) The structure of the feature fusion module is as shown in the attached figure. Figure 6 As shown, firstly for the feature map By splicing according to the channel dimensions, the resulting size is... Feature map ( (512), features are extracted using global average pooling to obtain the weight vector. The formula is: ; ; in This represents the value of the feature map in the coordinates (width, height, dimension). Represents the weight vector In the index The value on.

[0032] Then Expand into a one-dimensional vector with size ( (512), after feature extraction through two fully connected layers, the fusion weights are obtained through the Softmax activation function, as shown in the following formula: ; Let be the weight vector, with size . , It is a fully connected layer with input of C / 2 and output of 4. It is a fully connected layer with 4C inputs and C / 2 outputs.

[0033] Finally, the weights Weighted to The final feature map is obtained by summing the results.

[0034] (4) Rotating target detection stage: (4.1) In the remote sensing image detection stage, the Oriented-RCNN architecture is used to process image features and output the detection structure. The Oriented-RCNN model mainly consists of three parts: RPN, ROI, and bounding box head (as shown in the appendix). Figure 8 (As shown).

[0035] 1) Feature map output by S32 Each feature point generates 9 Anchors, totaling... ,for Generate one feature point A score is used to predict whether the anchor belongs to the foreground or background, and The data points are used as Anchor coordinates, then sorted by score, and the top N input ROIs are selected. The loss in this stage includes two types: classification loss and Anchor regression loss, as shown in the following formula: ; ; ; ; ; ; ; ; in These are the coordinates and width / height offsets of the horizontal bounding box. For RPN loss, Foreground and background classification loss, The loss function for the horizontal bounding box. It is a smoothing function.

[0036] 2) ROI stage: For 1), first predict a rough rotation angle. The rotated bounding box is mapped onto the feature map, and features of a fixed size are obtained proportionally (for example, if the original image has anchors of a certain size). The feature map size is Then the Anchor is mapped to the feature map as follows: Assuming a fixed mapping to Size, then in Upper division One equal area is For each region, the feature value of the center point is taken, and bilinear interpolation is used.

[0037] In this process, the loss is the angle prediction loss, and the formula is as follows, where... Will be regulated in .

[0038] .

[0039] 3) Bbox head phase: For each Anchor, there is Features, flattened A long one-dimensional vector passes through two fully connected layers (FC), then through two branch fully connected layers, one of which passes through... get One branch outputs a vector representing the classification result; the other branch outputs a (12×1) dimensional variable, where 4 variables represent the horizontal bounding box offset and 8 variables represent the rotated bounding box offset. The loss at this stage mainly includes classification loss, horizontal regression loss, and rotational regression loss, with the specific formulas as follows: ; ; ; ; ; in For the real category, For predicting categories, This is the horizontal bounding box offset, the calculation method of which has been given above. It is the rotation box offset (i.e., the difference between the four coordinates of the predicted box and the actual values). To detect the loss, For classifying losses, For bounding box regression loss, For horizontal frame loss, For rotation frame loss, is the feature weight hyperparameter.

[0040] (4.2) In the post-processing stage, for the bounding boxes corrected in (3.2), the class with the highest classification score is selected. Those with scores below the set threshold are directly discarded. Different classes are sorted in descending order of score. The class with the highest score is selected as the benchmark class. Those with an intersection-union ratio (IU) exceeding the threshold with the benchmark class are discarded, while those with a IU below the threshold are temporarily retained. Then, the filtering is iterated until all boxes are processed, and only the last remaining detection boxes are retained.

[0041] (4.3) In the image stitching stage, the images and detection results are stitched in reverse according to the cropping method in (1.4) to obtain the final result, and the different formats of the result (png, jpg, json) are output to the front-end interface.

[0042] This method extracts image features by using group rotation convolution, effectively reducing the number of parameters and improving feature extraction capability. It also generates multi-scale feature maps through dynamic feature fusion, ultimately achieving the task of rotating target detection in remote sensing images.

[0043] A remote sensing image rotation target detection system includes a preprocessing module, a feature extraction module, a multi-scale feature fusion module, a remote sensing image detection module, and a stitched original image output module; Preprocessing module: Performs cropping operations on the image; Feature extraction module: Uses an adaptive group rotation deformable convolutional network to extract features from the image and obtain a feature map pyramid; Multi-scale feature fusion module: performs dynamic feature fusion on the features output by the feature pyramid to obtain multi-scale fused features; Remote sensing image detection module: fuses feature maps to perform rotating target detection and recognition on the image to be detected, and outputs the rotating target bounding box; Original Image Stitching Output Module: Outputs the stitched original image.

[0044] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting rotating targets in remote sensing images, characterized in that, Includes the following steps: S1. Receive the image file to be detected and preprocess the image; S2. Feature extraction; S21. Image preprocessing; S22. Rotational Deformation Convolution: Text location detection is performed using the ResNet-AGRDC model. The input image is processed through a backbone network consisting of five stages of ResNet50 blocks replaced with AGRDConv blocks for feature extraction. S23. Attention Adjustment: Each AGRDConv stage uses spatial attention and channel attention mechanisms to adjust the weights of feature maps within different groups based on the obtained feature maps; S3. Perform dynamic feature fusion on the features output by the feature pyramid to obtain multi-scale fused features; S31 Feature Routing: The feature map size is adjusted to the same value using 3×3 convolution and deconvolution. Same size, that is Then, spatial attention is used to adjust the weight distribution of the convolutional feature maps; S32 Feature Fusion: First, regarding feature maps By splicing according to the channel dimensions, the resulting size is... Feature map Features are extracted using global average pooling to obtain the weight vector. The formula is: ； in This represents the value of the feature map on the coordinate system. Represents the weight vector In the index The value on; Then It is a one-dimensional vector with size . After feature extraction through two fully connected layers, the fusion weights are obtained through the Softmax activation function. The formula for this process is as follows: ； Let be the weight vector, with size . , For input / 2 output is The fully connected layer Input 4 The output is / 2 fully connected layer; Finally, the weights Weighted to The final feature map is obtained after summing. The size is ; S4. Use fused features to detect rotated targets and output rotated bounding boxes.

2. The remote sensing image rotation target detection method according to claim 1, characterized in that, The specific method for step S1 is as follows: S11. Upload the remote sensing image to be detected; S12. Select a suitable remote sensing image scene and appropriate model parameters; S13. Perform image detection; S14. Perform data preprocessing on each of the above images: cut or fill to the model input size. .

3. The method for detecting rotating targets in remote sensing images according to claim 1, characterized in that, Image preprocessing S21 includes: calculating the grayscale histogram of the image and calculating the cumulative histogram of the image; The mapping relationship between input gray levels and output gray levels is defined as follows: ； in The input image pixel gray levels, To output the pixel grayscale levels of the image, The grayscale value is The total number of pixels, The height of the input image. Given the width of the input image, the range is enlarged to between 0 and 255 to obtain the pixels of the output image; the gray level of each pixel in the image is output cyclically according to the mapping relationship.

4. The method for detecting rotating targets in remote sensing images according to claim 1, characterized in that, Step S22, rotational deformable convolution, includes: Step 1: Initialize a set of Gaussian convolution kernels, totaling... Groups, each group has n independent convolutional kernels, all of which have a size of [missing information]. ; For the width and height of the convolution kernel, and The original input / output channel dimensions are [dimensions], and the input dimension is [dimension]. The features are processed by depthwise convolution to Size, flattened A one-dimensional vector, after passing through a shared fully connected layer, enters a two-branch fully connected layer. Each branch fully connected layer then passes through another two-branch fully connected layer, and the output is processed by either the Softsign activation function or the Sigmoid activation function. The data sets are respectively Group rotation angle and Group convolution weights , An offset, Each offset weight; Perform group rotation convolution on each group: based on the predicted angle For convolution kernel and offset The matrix rotation operation is performed using the following formula: ； For batches, For width, For height, Number of channels; Represents the rotated convolution kernel. This represents the offset after rotation. Represents the initial convolution kernel. Represents the rotation angle. Let be a matrix rotation function. Represents the initial offset; Step Two, after obtaining Group of rotated convolution kernels Then, through the learned weight parameters The weighted fusion is performed using the following formula: ； in, The learned convolutional kernel weights, , The final convolution kernel after weighted rotation; Step 3: Perform a convolution operation using the newly calculated convolution kernel, as shown in the following formula: ； in The input feature map has a size of [size missing]. , The output feature map has a size of [size missing]. , For group convolution operations, For the sampling point index of the convolution kernel, This represents the offset of the convolution kernel sampling points. The weights are the sampling points of the convolution kernel.

5. The method for detecting rotating targets in remote sensing images according to claim 1, characterized in that, Step S23, attention adjustment, is as follows: Step 1: For the output feature map from the previous step Size is Two feature maps are obtained by pooling the width and height separately using global pooling. and The dimensions are respectively and The two feature maps are then concatenated to obtain... Feature map of size ; Step 2: After extracting features through convolution, convert the feature map... Divided into two feature maps and The dimensions are respectively and Multiply their width and height dimensions to get Weighted graph ; Step 3: Weighted Graph Compared with the original feature map Multiplying the corresponding positions together, we get the final result. Feature map ; Step 4: Steps S22 and S23 together replace the standard convolutions in the ResNet50 blocks, outputting four feature maps. .

6. The remote sensing image rotation target detection method according to claim 1, characterized in that, Step S4 is as follows: In the remote sensing image detection stage, the Oriented-RCNN architecture is used to process image features and output the detection structure. The Oriented-RCNN model consists of three parts: RPN, ROI, and bounding box head. RPN stage: Feature map output by S32 Each feature point generates 9 Anchors, totaling... ,for Generate one feature point A score is used to predict whether the anchor belongs to the foreground or background, and The data points are used as Anchor coordinates, then sorted by score, and the top N input ROIs are selected. The loss in this stage includes two types: classification loss and Anchor regression loss, as shown in the following formula: ； ； ； ； ； ； ； ； in These are the coordinates of the center point of the target bounding box. The target bounding box width and height, , , , These are the coordinates of the anchor's center point and its width and height, respectively. These are the coordinates and width / height offsets of the horizontal bounding box. For RPN loss, Foreground and background classification loss, The loss function for the horizontal bounding box. For smoothing functions, For feature weight hyperparameters, For predicting categories, Based on the prediction , , , Calculated ; To be based on reality , , , Calculated z is the independent variable of the function; ROI Phase: First, predict a rough rotation angle. Mapping the rotated target bounding box to the feature map The above steps involve obtaining features of a fixed size according to a proportional ratio. The loss during this process is the angle prediction loss, calculated using the following formula: ; in, To predict the target box rotation angle, Rotate the actual bounding box by the angle. Standards in , Refers to ROI loss; Bbox head phase: For each Anchor, there is... Features, flattened A long one-dimensional vector passes through two fully connected layers, then through two branched fully connected layers, one of which passes through... get The first branch outputs a vector representing the classification result; the second branch outputs 3D variables, where 4 variables represent the horizontal bounding box offset and 8 variables represent the rotated bounding box offset. The loss at this stage mainly includes classification loss, horizontal regression loss, and rotational regression loss, as shown in the following formulas: ； ； ； ； ； in For the real category, For predicting categories, This represents the actual horizontal bounding box offset. To predict the horizontal target box offset, It is the offset of the actual rotated bounding box, that is, the difference between the four coordinate points of the actual bounding box and the horizontal target bounding box. To predict the rotation box offset, which is the difference between the four coordinate points of the predicted box and the horizontal target box, To detect the loss, For classifying losses, For bounding box regression loss, For horizontal frame loss, For rotation frame loss, is the feature weight hyperparameter.

7. The remote sensing image rotation target detection method according to claim 6, characterized in that, include: In the post-processing stage, for the bounding boxes corrected by S32, the class with the highest classification score is selected. Those with scores below a set threshold are discarded directly. Different classes are sorted in descending order of score, and the class with the highest score is selected as the baseline class. Those with an intersection-union ratio (IU) exceeding the baseline class threshold are discarded, while those with a IU below the threshold are temporarily retained. Then, the filtering is iterated until all boxes are processed, and only the last remaining detection boxes are retained. In the image stitching stage, the images and detection results are stitched together in reverse according to the cropping method in S14 to obtain the final result, and the different formats of the result are output to the front-end interface.

8. A remote sensing image rotation target detection system, applicable to the method described in any one of claims 1-7, characterized in that, It includes a preprocessing module, a feature extraction module, a multi-scale feature fusion module, a remote sensing image detection module, and a stitched original image output module; Preprocessing module: Performs cropping operations on the image; Feature extraction module: Uses an adaptive group rotation deformable convolutional network to extract features from the image and obtain a feature map pyramid; Multi-scale feature fusion module: Dynamic feature fusion is performed on the features output by the feature pyramid to obtain multi-scale fused features; Remote sensing image detection module: fuses feature maps to perform rotating target detection and recognition on the image to be detected, and outputs the rotating target bounding box; Original Image Stitching Output Module: Outputs the stitched original image.

Images