A remote sensing image ship target detection method and system based on a graph semantic redirection module
By introducing a graph semantic retargeting module into the remote sensing image detection head, and utilizing graph structure and multilayer perceptron for feature decoupling and retargeting, the problem of insufficient extraction of rotation-invariant features of ship targets in remote sensing images is solved, and high-precision ship target detection is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 耕宇牧星(北京)空间科技有限公司
- Filing Date
- 2026-02-03
- Publication Date
- 2026-05-29
AI Technical Summary
Existing methods for detecting ship targets in remote sensing images suffer from poor decoupling capabilities and insufficient extraction of rotation-invariant features in the detection head network, leading to decreased detection accuracy, especially in complex backgrounds where it is difficult to accurately identify and locate ship targets.
The detection head structure is optimized by using a graph semantic redirection module. Feature decoupling and redirection are performed through graph structure topology modeling and multilayer perceptron. Pixel-level semantic similarity is used to construct the graph structure, and global semantic consistency is integrated by combining the exponential moving average mechanism to explicitly extract rotation-invariant features.
It significantly improves the accuracy and robustness of ship target detection, solves the problems of feature misalignment and missed detection caused by rotation and complex backgrounds, and achieves high-precision ship target detection.
Smart Images
Figure CN122116128A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of remote sensing image processing technology, and more specifically to a method and system for detecting ship targets in remote sensing images based on a graph semantic retargeting module. Background Technology
[0002] Remote sensing images have broad application prospects in fields such as national defense and security, environmental monitoring, and urban planning. Remote sensing image target detection technology has always been one of the research focuses in the field of remote sensing.
[0003] Current mainstream remote sensing image ship target detection methods are mainly based on backbone networks such as convolutional neural networks or Transformers to extract features. After feature pyramid processing, the detection head network is used for final category prediction and position regression.
[0004] However, most existing target detection methods focus on the feature extraction capabilities of the backbone network, while neglecting the limitations of the detection head network in processing complex features, resulting in significant bottlenecks in ship detection performance.
[0005] In deep learning models, the detection head undertakes the dual tasks of classification and localization. First, existing detection head structures generally suffer from poor decoupling capabilities. Specifically, classification tasks mainly rely on semantic features with translation invariance, while localization and angle regression tasks require geometric features with rotation sensitivity. Most methods simply share feature maps in the detection head, failing to effectively distinguish between these two fundamentally different feature requirements, leading to feature conflicts and thus reducing detection accuracy.
[0006] Secondly, existing detection heads cannot effectively identify and utilize rotation-invariant features in response to the rotational diversity of remotely sensed ship targets. Specifically, ships in remotely sensed images are usually arranged in arbitrary directions with large aspect ratios and close arrangement. Traditional convolutional detection heads are limited by fixed geometric structures and find it difficult to capture the inherent semantic relationships that change with angles. They also lack the ability to model rotation-invariant features. This makes the feature response of the model unstable when facing ships with different orientations, which easily leads to missed detections or inaccurate bounding box regression.
[0007] Therefore, how to effectively decouple task features in the detection head stage and enhance the extraction of rotation-invariant semantic features to achieve high-precision remote sensing ship detection is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0008] In view of the above problems, this invention provides a method and system for ship target detection in remote sensing images based on a graph semantic redirection module. By designing a graph semantic redirection module and optimizing the detection head structure based on it, the invention addresses the shortcomings of existing detection heads in feature processing. This module utilizes the topological modeling advantages of graph structures to capture potential semantic relationships in target features and explicitly extracts key features with rotation invariance. Simultaneously, the designed redirection mechanism can dynamically adjust the flow and distribution of semantic features according to different requirements of classification and regression tasks, achieving effective decoupling and alignment of features. The detection head network built based on the graph semantic redirection module can effectively overcome the interference caused by the large aspect ratio and arbitrary rotation of ships, significantly improving the model's feature representation ability for ship targets. This provides a more accurate and robust solution for ship target detection in remote sensing images, greatly enhancing detection performance in complex sea surface backgrounds and providing strong technical support for future marine monitoring and remote sensing applications.
[0009] To achieve the above objectives, the present invention adopts the following technical solution: A method for detecting ship targets in remote sensing images based on a graph semantic retargeting module includes the following steps: S1. Construct an improved detection head and corresponding target detection model that incorporates a graph semantic redirection module. Input the remote sensing image into the target detection model, extract features through stacked convolutional blocks, and obtain shallow feature maps, mid-level feature maps, and deep feature maps with feature information. Then, use a feature pyramid network to perform feature fusion. S2. Input the fused feature map into the improved detection head that introduces the graph semantic redirection module, introduce a multilayer perceptron for graph semantic reasoning, use pixel-level semantic similarity to construct a graph structure to extract key features with rotation invariance, and use a feature integration mechanism based on exponential moving average to perform global semantic consistency integration, thereby decoupling the feature map to obtain the category and location information of the potential target. S3. Train and test the target detection model for remote sensing image target detection tasks. Input the remote sensing image to be tested into the trained remote sensing image ship target detection model combined with the graph semantic relocation module and output the corresponding target detection results.
[0010] Preferably, the feature fusion performed by the feature pyramid network is as follows:
[0011] in, This represents the fused feature map corresponding to the i-th layer. The pyramid is a characteristic pyramid. These are the shallow feature map, middle feature map, and deep feature map obtained after feature extraction, respectively, all containing feature information. The feature map is obtained by performing feature extraction on a remote sensing image G with size H×W and 3 channels through stacked convolutional blocks.
[0012] Preferably, the improved detection head that incorporates the graph semantic redirection module contains the following: Feature graph decoupling is achieved through a graph semantic redirection module:
[0013] in, To optimize the feature map output by the graph semantic redirection module, For graph semantic redirection module; Predict the category and location information of potential targets using classification and regression functions:
[0014]
[0015] in, This represents the classification information obtained from the prediction. This indicates the predicted location information. Represents the classification function, This represents the location regression function.
[0016] Preferably, step S2 includes the following: S21. Perform channel-level feature splitting on the input fused feature map to obtain the split sub-maps; S22. The refined feature maps of each subgraph are obtained by refining the subgraph information through the first multilayer perceptron; S23. Calculate the initial graph structure information of each sub-feature map based on the semantic similarity at the pixel level, and then filter the graph structure information through a second multilayer perceptron. S24. Introduce a feature integration mechanism based on exponential moving average to integrate the graph structure information of each sub-feature map in a consistent manner, and dynamically weight and average the features of different branches to construct a semantic prototype with global consistency. S25. The shape and number of channels of the feature map are adjusted by feature fusion and convolution to obtain the output feature map after calculation by the graph semantic redirection module.
[0017] Preferably, the specific content of pixel-level semantic similarity calculation is as follows: For each input sub-feature map, assign two-dimensional spatial coordinates to each spatial location. The pixel at position i is mapped to a one-dimensional node with index i, and each node... The initial feature vector is the C-dimensional feature vector of the corresponding pixel on the corresponding sub-feature map. All pixels in the feature map constitute the node set of the graph. Edges are determined based on pixel-level semantic similarity between nodes and after normalization. and weight .
[0018] Preferably, in step S3, the network is trained for remote sensing image ship target detection using focus loss and rotation box intersection-union loss.
[0019] A remote sensing image ship target detection system based on a graph semantic redirection module, based on the aforementioned remote sensing image ship target detection method based on a graph semantic redirection module, includes: a model training and testing module and a trained remote sensing image ship target detection model combined with the graph semantic redirection module; The model training and testing module is used to train and test the target detection model for remote sensing image target detection tasks. The trained remote sensing image ship target detection model combined with the graph semantic retargeting module is used to input the remote sensing image to be tested and output the corresponding target detection results. The remote sensing image ship target detection model combined with the graph semantic redirection module includes a feature extraction module, a feature fusion module, and an improved detection head that incorporates the graph semantic redirection module; The feature extraction module is used to extract features from the input remote sensing image through stacked convolutional blocks to obtain shallow feature maps, mid-level feature maps, and deep feature maps with feature information. The feature fusion module is used to perform feature fusion using a feature pyramid network. An improved detection head is introduced through a graph semantic redirection module to perform graph semantic reasoning on the fused feature map. A multilayer perceptron is introduced to construct a graph structure using pixel-level semantic similarity to extract key features with rotation invariance. A feature integration mechanism based on exponential moving average is used to perform global semantic consistency integration, thereby decoupling the feature map to obtain the category and location information of potential targets.
[0020] Preferably, the improved detection head that incorporates a graph semantic redirection module includes the graph semantic redirection module as well as classification and regression functions; The graph semantic redirection module is used to decouple feature maps; Classification and regression functions are used to obtain information about the category and location of potential targets; The graph semantic redirection module includes a channel splitting module, a multilayer perceptron module, a graph building mechanism based on semantic similarity, an exponential average moving average mechanism, a feature fusion module, and a basic convolution module. The channel splitting module is used to perform channel-level feature splitting on the input fused feature map to obtain the split sub-maps; The multilayer perceptron module is used to obtain the refined feature maps of each subgraph by refining the subgraph information through the first multilayer perceptron, and is also used to filter the graph structure information through the second multilayer perceptron. A semantic similarity-based graph construction mechanism is used to calculate the initial graph structure information of each sub-feature map based on pixel-level semantic similarity from multiple obtained sub-feature maps. The exponential moving average mechanism is used to consistently integrate the graph structure information of each sub-feature map by grouping them, and dynamically weighting and averaging the features of different branches to construct a semantic prototype with global consistency. The feature fusion module and the basic convolution module are used to adjust the shape and number of channels of the feature map through feature fusion and convolution to obtain the output feature map after being calculated by the graph semantic relocation module.
[0021] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for detecting ship targets in remote sensing images based on a graph semantic retargeting module.
[0022] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned method for detecting ship targets in remote sensing images based on a graph semantic retargeting module.
[0023] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for remote sensing image ship target detection based on a graph semantic retargeting module, which has the following beneficial effects: By constructing a graph structure using pixel-level semantic similarity through the graph semantic retargeting module, the inherent semantic relationships of ship targets can be explicitly captured across spatial distances. This effectively extracts key features with rotation invariance, enabling the model to maintain a stable feature response when facing ships with large aspect ratios and arbitrary angles in remote sensing images. This effectively solves the problems of feature misalignment and missed detection caused by target rotation. The detection head structure based on graph semantic redirection optimization solves the feature conflict problem between classification tasks (relying on translation-invariant features) and localization tasks (relying on rotation-sensitive features) in traditional detection heads by introducing a multilayer perceptron (MLP) for graph semantic reasoning and utilizing the exponential moving average (EMA) mechanism for global semantic consistency integration. This structure not only achieves efficient decoupling and redirection of features, but also ensures accurate alignment between high-level semantic information and low-level geometric details by eliminating semantic ambiguity in the fusion process, thereby achieving accurate classification and high-precision bounding box regression of remotely sensed ship targets. Attached Figure Description
[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0025] Figure 1 This is a schematic diagram of a remote sensing image ship target detection model combining a graph semantic redirection module provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the graph semantic redirection module provided in an embodiment of the present invention. Detailed Implementation
[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0027] This invention discloses a method for detecting ship targets in remote sensing images based on a graph semantic redirection module, such as... Figure 1 This includes the following steps: S1. Construct an improved detection head and corresponding target detection model that incorporates a graph semantic redirection module. Input the remote sensing image into the target detection model, extract features through stacked convolutional blocks, and obtain shallow feature maps, mid-level feature maps, and deep feature maps with feature information. Then, use a feature pyramid network to perform feature fusion. S2. Input the fused feature map into the improved detection head that introduces the graph semantic redirection module, introduce a multilayer perceptron for graph semantic reasoning, use pixel-level semantic similarity to construct a graph structure to extract key features with rotation invariance, and use a feature integration mechanism based on exponential moving average to perform global semantic consistency integration, thereby decoupling the feature map to obtain the category and location information of the potential target. S3. Train and test the target detection model for remote sensing image target detection tasks. Input the remote sensing image to be tested into the trained remote sensing image ship target detection model combined with the graph semantic relocation module and output the corresponding target detection results.
[0028] To further implement the above technical solution, the feature pyramid network performs feature fusion as follows:
[0029] in, This represents the fused feature map corresponding to the i-th layer. The pyramid is a characteristic pyramid. These are the shallow feature map, middle feature map, and deep feature map obtained after feature extraction, respectively, all containing feature information. The feature map is obtained by performing feature extraction on a remote sensing image G with size H×W and 3 channels through stacked convolutional blocks.
[0030] To further implement the above technical solution, the improved detection head of the graph semantic redirection module is introduced as follows: Feature graph decoupling is achieved through a graph semantic redirection module:
[0031] in, To optimize the feature map output by the graph semantic redirection module, For graph semantic redirection module; Predict the category and location information of potential targets using classification and regression functions:
[0032]
[0033] in, This represents the classification information obtained from the prediction. This indicates the predicted location information. Represents the classification function, This represents the location regression function.
[0034] To further implement the above technical solutions, such as Figure 2 The specific content of step S2 includes: S21. Perform channel-level feature splitting on the input fused feature map to obtain the split sub-maps; S22. The refined feature maps of each subgraph are obtained by refining the subgraph information through the first multilayer perceptron; S23. Calculate the initial graph structure information of each sub-feature map based on the pixel-level semantic similarity, i.e. the corresponding point-edge relationship, and filter the graph structure information through a second multilayer perceptron. S24. In order to alleviate the problems of information misalignment and semantic ambiguity in the structure of each subgraph and further improve the decoupling ability of the detection head for classification and regression tasks, a feature integration mechanism based on exponential moving average is introduced. By utilizing the smoothing property of EMA, the graph structure information of each sub-feature graph is grouped and integrated in a consistent manner. The features of different branches are dynamically weighted and averaged to construct a semantic prototype with global consistency. This not only eliminates random noise in single-layer features, but also promotes the effective complementarity between deep abstract semantics and shallow detailed information. S25. The shape and number of channels of the feature map are adjusted by feature fusion and convolution to obtain the output feature map after calculation by the graph semantic redirection module.
[0035] In this embodiment, step S21, the channel-level feature splitting operation, is as follows:
[0036] in, This indicates a channel-level feature splitting operation. The resulting sub-image has 1 / 4 the number of channels of the original image. Input feature map; The first method for refining subgraph information using a multilayer perceptron is as follows:
[0037]
[0038]
[0039]
[0040] in, This indicates the first multilayer perceptron module. These are the feature maps obtained after each subgraph has been refined using MLP. The method for obtaining initial graph structure information based on pixel-level semantic similarity calculation is as follows:
[0041]
[0042]
[0043]
[0044] in, This indicates a pixel-level semantic similarity calculation module. The calculated graph structure representation information; The content of the second multilayer perceptron filter graph structure information is as follows:
[0045]
[0046]
[0047]
[0048] in, This indicates the second multilayer perceptron module. These are the feature maps obtained after each subgraph has been refined using MLP. Considering the stronger complementarity of features at adjacent scales, the graph structure information of each sub-feature map is divided into two groups for consistent integration. The specific content is as follows:
[0049]
[0050] in, and These represent the features of the integrated subgraph. Indicates an exponential moving average; The output feature map after calculation by the graph semantic redirection module is:
[0051] in, This is the output feature map calculated by the graph semantic redirection module. This is a feature fusion module, which typically uses feature concatenation and feature addition, and can be adjusted according to the task. This is a convolution used to adjust the shape and channels of the feature map.
[0052] To further implement the above technical solution, the specific details of pixel-level semantic similarity calculation are as follows: For each input sub-feature map, assign two-dimensional spatial coordinates to each spatial location. The pixel at position i is mapped to a one-dimensional node with index i, and each node... The initial feature vector is the C-dimensional feature vector of the corresponding pixel on the corresponding sub-feature map. All pixels in the feature map constitute the node set of the graph. Edges are determined based on pixel-level semantic similarity between nodes and after normalization. and weight .
[0053] In this embodiment, for any two nodes and Semantic similarity is obtained by calculating the similarity measure between feature vectors. The calculation process usually uses a metric function such as dot product or cosine similarity.
[0054] in, and For feature transformation functions, such as 1×1 convolution, Represents a node and The original similarity score between them; The similarity scores are normalized using the Softmax function to obtain the connection nodes. and Edge weights between:
[0055] N represents the total number of nodes in the graph, and n represents the weight of the loop variable in the summation process. The semantic association strength between two pixels was quantified. Some target pixels may be spatially separated, but highly related in semantic features, such as the bow and stern of a ship, thus establishing a close connection in the graph.
[0056] To further implement the above technical solution, in step S3, the network is trained for remote sensing image target detection task using focus loss and rotation box intersection-union loss.
[0057] In this embodiment, the loss function for training the remote sensing image ship target detection task is:
[0058] in, This represents the focus loss, used in the classification and prediction branch for target detection tasks in remote sensing images; This represents the IoU loss function for the rotated bounding box, used to calculate the distance between the predicted bounding box and the ground truth label of the target in the input remote sensing image; This represents the total loss used in network training; Once the network training reaches a point where the loss no longer decreases, it indicates that the network training has reached a stable state, and the training process ends. The trained object detection model, combined with the graph semantic redirection module, is used to test the remote sensing image to be tested, as shown below:
[0059] in, This represents the trained object detection model incorporating the graph semantic redirection module. This represents the remote sensing image to be tested. This indicates the corresponding target detection result.
[0060] A remote sensing image ship target detection system based on a graph semantic redirection module, which is based on a remote sensing image ship target detection method based on a graph semantic redirection module, includes: a model training and testing module and a trained remote sensing image ship target detection model combined with the graph semantic redirection module; The model training and testing module is used to train and test the target detection model for remote sensing image target detection tasks. The trained remote sensing image ship target detection model combined with the graph semantic retargeting module is used to input the remote sensing image to be tested and output the corresponding target detection results. The remote sensing image ship target detection model combined with the graph semantic redirection module includes a feature extraction module, a feature fusion module, and an improved detection head that incorporates the graph semantic redirection module; The feature extraction module is used to extract features from the input remote sensing image through stacked convolutional blocks to obtain shallow feature maps, mid-level feature maps, and deep feature maps with feature information. The feature fusion module is used to perform feature fusion using a feature pyramid network. An improved detection head is introduced through a graph semantic redirection module to perform graph semantic reasoning on the fused feature map. A multilayer perceptron is introduced to construct a graph structure using pixel-level semantic similarity to extract key features with rotation invariance. A feature integration mechanism based on exponential moving average is used to perform global semantic consistency integration, thereby decoupling the feature map to obtain the category and location information of potential targets.
[0061] To further implement the above technical solution, an improved detection head is introduced, which includes a graph semantic redirection module as well as classification and regression functions; The graph semantic redirection module is used to decouple feature maps; Classification and regression functions are used to obtain information about the category and location of potential targets; The graph semantic redirection module includes a channel splitting module, a multilayer perceptron module, a graph building mechanism based on semantic similarity, an exponential average moving average mechanism, a feature fusion module, and a basic convolution module. The channel splitting module is used to perform channel-level feature splitting on the input fused feature map to obtain the split sub-maps; The multilayer perceptron module is used to obtain the refined feature maps of each subgraph by refining the subgraph information through the first multilayer perceptron, and is also used to filter the graph structure information through the second multilayer perceptron. A semantic similarity-based graph construction mechanism is used to calculate the initial graph structure information of each sub-feature map based on pixel-level semantic similarity from multiple obtained sub-feature maps. The exponential moving average mechanism is used to consistently integrate the graph structure information of each sub-feature map by grouping them, and dynamically weighting and averaging the features of different branches to construct a semantic prototype with global consistency. The feature fusion module and the basic convolution module are used to adjust the shape and number of channels of the feature map through feature fusion and convolution to obtain the output feature map after being calculated by the graph semantic relocation module.
[0062] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a remote sensing image ship target detection method based on a graph semantic retargeting module.
[0063] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements a remote sensing image ship target detection method based on a graph semantic retargeting module.
[0064] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0065] 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 ship targets in remote sensing images based on a graph semantic retargeting module, characterized in that, Includes the following steps: S1. Construct an improved detection head and a corresponding remote sensing image ship target detection model by introducing a graph semantic redirection module. Input the remote sensing image into the target detection model, extract features through stacked convolutional blocks, and obtain shallow feature maps, middle feature maps and deep feature maps with feature information. Then, use a feature pyramid network to perform feature fusion. S2. Input the fused feature map into the improved detection head that introduces the graph semantic redirection module, introduce a multilayer perceptron for graph semantic reasoning, use pixel-level semantic similarity to construct a graph structure to extract key features with rotation invariance, and use a feature integration mechanism based on exponential moving average to perform global semantic consistency integration, thereby decoupling the feature map to obtain the category and location information of the potential target. S3. Train and test the target detection model for remote sensing image target detection tasks. Input the remote sensing image to be tested into the trained remote sensing image ship target detection model combined with the graph semantic relocation module and output the corresponding target detection results.
2. The remote sensing image ship target detection method based on a graph semantic redirection module as described in claim 1, characterized in that, The feature pyramid network performs feature fusion as follows: in, This represents the fused feature map corresponding to the i-th layer. The pyramid is a characteristic pyramid. These are the shallow feature map, middle feature map, and deep feature map obtained after feature extraction, respectively, all containing feature information. The feature map is obtained by performing feature extraction on a remote sensing image G with size H×W and 3 channels through stacked convolutional blocks.
3. The remote sensing image ship target detection method based on a graph semantic redirection module as described in claim 2, characterized in that, The specific content of the improved detection head introduced by the graph semantic redirection module is as follows: Feature graph decoupling is achieved through a graph semantic redirection module: in, To optimize the feature map output by the graph semantic redirection module, For graph semantic redirection module; Predict the category and location information of potential targets using classification and regression functions: in, This represents the classification information obtained from the prediction. This indicates the predicted location information. Represents the classification function, This represents the location regression function.
4. The remote sensing image ship target detection method based on a graph semantic redirection module as described in claim 1, characterized in that, The specific content of step S2 includes: S21. Perform channel-level feature splitting on the input fused feature map to obtain the split sub-maps; S22. The refined feature maps of each subgraph are obtained by refining the subgraph information through the first multilayer perceptron; S23. Calculate the initial graph structure information of each sub-feature map based on the semantic similarity at the pixel level, and then filter the graph structure information through a second multilayer perceptron. S24. Introduce a feature integration mechanism based on exponential moving average to integrate the graph structure information of each sub-feature map in a consistent manner, and dynamically weight and average the features of different branches to construct a semantic prototype with global consistency. S25. The shape and number of channels of the feature map are adjusted by feature fusion and convolution to obtain the output feature map after calculation by the graph semantic redirection module.
5. The remote sensing image ship target detection method based on a graph semantic redirection module as described in claim 4, characterized in that, The specific details of pixel-level semantic similarity calculation are as follows: For each input sub-feature map, assign two-dimensional spatial coordinates to each spatial location. The pixel at position i is mapped to a one-dimensional node with index i, and each node... The initial feature vector is the C-dimensional feature vector of the corresponding pixel on the corresponding sub-feature map. All pixels in the feature map constitute the node set of the graph. Edges are determined based on pixel-level semantic similarity between nodes and after normalization. and weight .
6. The remote sensing image ship target detection method based on a graph semantic redirection module as described in claim 1, characterized in that, In step S3, the network is trained for remote sensing image ship target detection using focus loss and rotation box intersection-union loss.
7. A remote sensing image ship target detection system based on a graph semantic redirection module, characterized in that, A remote sensing image ship target detection method based on a graph semantic redirection module according to any one of claims 1-6 includes: a model training and testing module and a trained remote sensing image ship target detection model combined with the graph semantic redirection module. The model training and testing module is used to train and test the target detection model for remote sensing image target detection tasks. The trained remote sensing image ship target detection model combined with the graph semantic retargeting module is used to input the remote sensing image to be tested and output the corresponding target detection results. The remote sensing image ship target detection model combined with the graph semantic redirection module includes a feature extraction module, a feature fusion module, and an improved detection head that incorporates the graph semantic redirection module; The feature extraction module is used to extract features from the input remote sensing image through stacked convolutional blocks to obtain shallow feature maps, mid-level feature maps, and deep feature maps with feature information. The feature fusion module is used to perform feature fusion using a feature pyramid network. An improved detection head is introduced through a graph semantic redirection module to perform graph semantic reasoning on the fused feature map. A multilayer perceptron is introduced to construct a graph structure using pixel-level semantic similarity to extract key features with rotation invariance. A feature integration mechanism based on exponential moving average is used to perform global semantic consistency integration, thereby decoupling the feature map to obtain the category and location information of potential targets.
8. A remote sensing image ship target detection system based on a graph semantic redirection module as described in claim 7, characterized in that, The improved detection head, which introduces a graph semantic redirection module, includes the graph semantic redirection module as well as classification and regression functions; The graph semantic redirection module is used to decouple feature maps; Classification and regression functions are used to obtain information about the category and location of potential targets; The graph semantic redirection module includes a channel splitting module, a multilayer perceptron module, a graph building mechanism based on semantic similarity, an exponential average moving average mechanism, a feature fusion module, and a basic convolution module. The channel splitting module is used to perform channel-level feature splitting on the input fused feature map to obtain the split sub-maps; The multilayer perceptron module is used to obtain the refined feature maps of each subgraph by refining the subgraph information through the first multilayer perceptron, and is also used to filter the graph structure information through the second multilayer perceptron. A semantic similarity-based graph construction mechanism is used to calculate the initial graph structure information of each sub-feature map based on pixel-level semantic similarity from multiple obtained sub-feature maps. The exponential moving average mechanism is used to consistently integrate the graph structure information of each sub-feature map by grouping them, and dynamically weighting and averaging the features of different branches to construct a semantic prototype with global consistency. The feature fusion module and the basic convolution module are used to adjust the shape and number of channels of the feature map through feature fusion and convolution to obtain the output feature map after being calculated by the graph semantic relocation module.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements a remote sensing image ship target detection method based on a graph semantic relocation module as described in any one of claims 1 to 6.
10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements a remote sensing image ship target detection method based on a graph semantic retargeting module as described in any one of claims 1 to 6.