Remote sensing image ship target detection method based on geographic information hybrid expert model

Abstract

The invention provides a remote sensing image ship target detection method based on a geographic information hybrid expert model. The method depends on a hybrid expert model and a Mask-RCNN network in deep learning, and is a visible light remote sensing image ship target deep learning detection method suitable for a multi-spatial resolution and complex geographical environment. In a Mask-RCNN feature extraction link, two expert groups are introduced, and two feature extraction expert groups guided by resolution and latitude and longitude geographic information respectively are constructed. In this way, detection of the ship target in the multi-resolution complex background visible light remote sensing image is achieved. The method not only has an accurate multi-scale target position sensing capability, but also can effectively eliminate ground feature interference in various geographic areas and complex background environments, and makes up for the deficiency of a current deep learning ship detection method in the aspect of multi-resolution complex background image research.

Description

Technical Field

[0001] This invention is a method for ship target detection in remote sensing images based on a geographic information hybrid expert model. This method relies on a hybrid expert model in deep learning and a Mask-RCNN network, and is a deep learning-based method for ship target detection in visible light remote sensing images under multi-spatial resolution and complex geographic environments. It focuses on solving the problem of ship target detection in visible light remote sensing images under multi-spatial resolution and complex geographic environments, and technically falls under the category of target detection in multi-resolution, complex background remote sensing images. Background Technology

[0002] Among the many targets monitored by satellite remote sensing, ships possess the dual attributes of being major maritime transport vehicles and important military targets. Therefore, achieving automatic detection and identification of ships is of significant research importance. In the civilian sector, ship target detection is widely used in maritime transport monitoring, maritime rescue, and port management. In the military sector, ship target detection is widely applied in maritime surveillance and maritime combat situational awareness. However, ship detection tasks in remote sensing images have many inherent characteristics, such as the diversity of geographical areas, the inconsistency of target scale, and the complexity of the background environment. These characteristics pose significant challenges to achieving high-quality ship detection and place higher demands on ship detection methods.

[0003] However, existing methods merely employ a uniform structure to process remote sensing image information, lacking the application of crucial geographic information and struggling to adapt to the complex and diverse remote sensing scenes for ship detection. Different latitude and longitude coordinates encompass geographically diverse regions with varying environmental factors that interfere with ship target detection, such as sea ice in polar regions and clouds in tropical regions. Current research, such as IceReggionShip's efforts to reduce sea ice interference in ice-covered waters and PNBT-CR's use of Berlin noise for thin cloud removal to aid ship target detection, focuses on localized areas and processes the image's inherent characteristics without considering the geographic regional characteristics implied by latitude and longitude information. This failure to integrate latitude and longitude information into the target detection process makes the detection models ill-suited to the complexities of different geographic regions, limiting improvements in detection accuracy and reliability.

[0004] Furthermore, to address the challenge of detecting targets at varying scales, a series of image-specific methods have been applied to remote sensing vessel detection research. For example, LSKNet dynamically adjusts its spatial receptive field to better model the long-range context of various targets in the remote sensing scene, utilizing global target information in the image to assist in the detection of small targets. Fostc3net proposes a novel and effective hybrid spatial pyramid pooling method to fuse local and global information from feature maps, improving the detection performance of vessel targets at different scales. While these studies alleviate the difficulty of multi-scale target detection to some extent, they only focus on the information within the image itself, failing to recognize the importance of image spatial resolution information for multi-scale target detection.

[0005] In general, current methods primarily rely on a single network structure to perceive the complex features of ground objects. This approach struggles to effectively identify and differentiate different types of ground objects, fails to accurately distinguish between complex background environments and vessel targets, and is prone to false detections. Hybrid expert models (MoE) effectively mitigate the interference of complex background environments on vessel target detection by assigning different features to specialized experts for differentiated processing. In conclusion, incorporating geographic information, including latitude, longitude, and resolution, into vessel target detection in remote sensing images, and utilizing hybrid expert models for adaptive processing of complex terrain environments, represents an important direction in the field of target detection technology for multi-resolution remote sensing images with complex backgrounds. Summary of the Invention

[0006] This invention aims to provide a method for ship target detection in remote sensing images based on a hybrid expert model using geographic information. This method relies on a hybrid expert model in deep learning and the Mask-RCNN network, making it a deep learning-based method for ship target detection in visible light remote sensing images with multiple spatial resolutions and complex geographic environments. In the feature extraction stage of Mask-RCNN, this method introduces two expert groups, constructing two feature extraction expert groups guided by resolution and latitude / longitude geographic information, respectively. In this way, ship targets in visible light remote sensing images with complex backgrounds at multiple resolutions can be detected. This method not only possesses accurate multi-scale target location perception capabilities but also effectively eliminates interference from ground features in diverse geographic regions and complex background environments, overcoming the shortcomings of current deep learning ship detection methods in the study of images with complex backgrounds at multiple resolutions. This invention is achieved through the following technical solutions:

[0007] This invention discloses a method for ship target detection in remote sensing images based on a geographic information hybrid expert model. The detection method is mainly implemented through a convolutional neural network and can be divided into an information encoding part, a feature extraction part, and a target detection part. The specific steps of the method are as follows:

[0008] Step 1: Computer Data Reading. First, use a computer to read the remote sensing image data and simultaneously load the corresponding latitude, longitude, and resolution information of the image; after reading the data, perform data augmentation on the image.

[0009] Step 2: Construct a deep neural network for ship target detection based on a geographic information hybrid expert model.

[0010] This invention designs a geographic information encoder to encode the loaded latitude, longitude, and resolution information separately. After geographic information encoding, this invention designs a feature extraction module (GeoMoE feature extraction module) based on a geographic information hybrid expert model to extract features from the loaded image. The GeoMoE feature extraction module can more effectively mine and extract different features of ground objects in the image. Subsequently, this invention uses a Mask-RCNN object detection module to integrate and utilize the extracted features, enabling the network to perceive ship targets in multi-resolution complex background images.

[0011] The GeoMoE (Geographic Information Hybrid Expert Model) consists of two distinct groups of experts: one guided by resolution geographic information and the other by latitude and longitude geographic information. The expert guided by resolution geographic information is responsible for perceiving ship targets at different scales; the expert guided by latitude and longitude geographic information focuses on identifying and adapting to diverse geographical features. Through the collaborative combination of these experts, complex background interference is effectively removed, thereby achieving high-quality ship target detection.

[0012] The geographic information-guided expert model is processed as follows: The loaded geographic information (latitude, longitude, and resolution) is encoded separately. For one-dimensional spatial resolution information, a multilayer perceptron and softmax function are used to encode the implicit information, transforming it into a spatial resolution expert group confidence probability distribution. For latitude and longitude information, a combination of sine and cosine coding, multilayer perceptron (MLP), and softmax is used to transform discrete latitude and longitude coordinates into a continuous encoding form that includes location features, ultimately obtaining the latitude and longitude expert group confidence probability distribution. In this way, geographic information can be incorporated into the hybrid expert model in the form of a biased expert confidence probability distribution, helping the model achieve more accurate ship target detection in complex multi-resolution backgrounds.

[0013] The final detection part uses the Mask-RCNN network framework to receive multi-scale image features containing geographic information extracted from the geographic information hybrid expert model. Then, it uses a region proposal network to generate a series of candidate regions that may contain ship targets. Finally, it uses feature maps of candidate regions of different sizes to accurately regress the location of the ship targets and obtain the ship target location.

[0014] Step 3: Train the deep neural network. After constructing the ship target detection deep neural network based on the geographic information hybrid expert model in Step 2, train the ship target detection deep neural network using labeled training data in the PyTorch deep learning framework until the performance of the target detection deep neural network reaches its optimal level, and record the parameters of the ship target detection deep neural network at this time.

[0015] Step 4: Remote Sensing Image Target Detection. Using the deep neural network for ship target detection trained in Step 3 based on a geographic information hybrid expert model, small ship targets are detected in multi-scale complex environment remote sensing image detection data containing latitude, longitude, and resolution information, and the ship position information is output.

[0016] The advantages of this invention are: by designing a deep learning architecture based on a hybrid expert model of geographic information combined with Mask-RCNN, this invention can achieve high-precision detection of ship targets of varying scales in visible light remote sensing images with multiple spatial resolutions and complex geographic environments. The input is a remote sensing image containing latitude, longitude, and resolution information, and the output is accurate ship target location information. It has a high degree of automation, improves adaptability and accuracy in complex and diverse geographic environments, and has significant application value. Attached Figure Description

[0017] Figure 1 This is a diagram of the deep neural network structure for ship target detection based on a geographic information hybrid expert model designed in this invention.

[0018] Figure 2 This is a structural diagram of the geographic information coding module.

[0019] Figure 3 This is a structural diagram of the GeoMoE feature extraction module.

[0020] Figure 4 This is the overall detection process of the method of the present invention. Detailed Implementation

[0021] To better understand the technical solution of the present invention, the embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0022] This invention is implemented using Python within the PyTorch framework. First, the network is built according to the invention's concept. A network for object detection based on a geographic information hybrid expert model and Mask-RCNN is constructed. After the network is built, relevant parameters are configured, and then the network is trained using training data. Next, the trained network model is used to detect ships in multi-resolution, complex background visible light remote sensing images. A suitable Python version is selected and installed according to the version compatibility requirements of the PyTorch deep learning framework.

[0023] During implementation, the computer configuration used a 12th Gen Intel® Core™ i9-12900K CPU processor with a clock speed of 3.20GHz, 512GB of memory, and an NVIDIA GeForce GTX 4080 graphics card with 16GB of video memory. Figure 3 As shown, the remote sensing image target method of the present invention includes the following steps:

[0024] Step 1: Data Reading by Computer. First, a computer is used to read the remote sensing image data and its corresponding latitude, longitude, and resolution geographic information. The data used in this invention comes from Google Earth, with spatial resolutions of 1.07m, 2.15m, 4m, 8m, and 16m. This invention annotated a total of 1860 1024×1024 images and divided them into training and test sets as shown in Table 1 below. Table 1 is the parameter table of the dataset used in the experiment.

[0025]

[0026] Table 1

[0027] After the data is read in, the images need to be augmented during the training phase.

[0028] Step 2: Construct a deep neural network for ship target detection based on a geographic information hybrid expert model.

[0029] like Figure 1 The diagram shows the deep neural network for ship target detection based on a geographic information hybrid expert model proposed in this invention. The entire deep neural network for ship target detection mainly consists of three parts: a geographic information encoding module, a GeoMoE feature extraction module, and a Mask-RCNN target detection module. The geographic information encoding module is the data information encoding method designed in this invention to guide the hybrid experts. The results of the geographic information encoding module are directly introduced into the GeoMoE feature extraction module, using the geographic information of the image to adjust the confidence level of the expert scores in the extraction module, thus more accurately extracting target information in complex environments. After extracting image features, the Mask-RCNN target detection module further utilizes the features, predicting the accurate location of the target through a two-stage progressive refinement of the target location.

[0030] Processing multi-scale, complex environment remote sensing images first requires acquiring the image's geographic information and then encoding and converting this information to facilitate subsequent use by the GeoMoE feature extraction module. The specific geographic information encoding process is as follows: Figure 2 As shown, one expert group is guided by resolution geographic information, while the other is guided by latitude and longitude geographic information.

[0031] Because geographic information contains relative spatial location information, which differs from the simple quantitative relationship of spatial resolution, the encoding methods for latitude and longitude information differ from those for spatial resolution. The processing flow for both is shown in the following formula:

[0032] Conf Loc =Φ softmax (Φ MLPl (Φ PE (I Loc )))

[0033] Conf Res =Φ softmax (Φ MLPr (I Res ))

[0034] Among them, I Loc ∈R 2 It is the center latitude and longitude location information of the image, I Res ∈R 1 This refers to the spatial latitude and longitude information of the image. Φ MLPl Φ MLPr These are a latitude and longitude multilayer perceptron and a spatial resolution multilayer perceptron, respectively, which convert latitude and longitude location information and spatial resolution information into scores for each expert in their respective expert groups. To ensure that these scores possess probabilistic characteristics as the basis for the probability distribution of expert assignments, this invention uses Φ... softmax The function transforms these scores into a probability distribution such that each expert's score is between 0 and 1, and the sum of all experts' scores in the expert group is 1. Φ PE It is a function that performs sine and cosine encoding on latitude and longitude. Φ softmax Φ PE The specific formula is as follows:

[0035]

[0036] Where N is the total number of experts in the expert group, and in this invention, the number of experts in both expert groups is set to 4. i It is the probability distribution for each expert. For each expert, the probability distribution value z i Perform exponential operations with the natural constant e as the base. Φ PE The latitude and longitude of the image are encoded and converted, where lat and lon are the latitude and longitude values, respectively, converting the range from [-90, 90] and (-180, 180] to (-1, 1] and (-1, 1]. model It is the encoded latitude and longitude feature latitude, which needs to be an integer multiple of 4. In this invention, it is set to 256.

[0037] The GeoMoE feature extraction module used in this invention is specifically as follows: Figure 1 As shown, the overall framework is an improvement on Swin-Transformer. Swin-Transformer stacks a series of structurally similar TransformerBlocks. The differences between different blocks are concentrated in the window partitioning and self-attention calculation range of the multi-head attention module. The improvement in this invention does not involve this part. This invention mainly uses Geo-MOE to replace the FFN of some blocks. The process in each block is described as follows:

[0038] T i =T i-1 +Φ Norm (Φ Attn (T i-1 ))+Φ Norm (Φ GMF (T i-1 +Φ Norm (Φ Attn (T i-1 )),I Geo ))

[0039] Where, Φ Attn The multi-head self-attention layer is performed in a local window, Φ Norm The normalization layer normalizes the input features, Φ GMF yes Figure 1 The GeoMoE & FFN modules shown contain FFN modules from the original Swin-Transformer in some layers, and Geo-MOE modules in others. i-1 T i T represents the tokens for the input and output of the i-th block, respectively. i-1 T i ∈R L×d L is the length of the input tokens, and d is the dimension of the feature. Geo ∈R L×d I Geo The geographic information corresponding to the image is specifically the image spatial resolution, the longitude of the image center, and the latitude of the image center.

[0040] And Φ GMF The specific processing flow of the GeoMoE & FFN modules represented is as follows:

[0041] T out =Φ Loc_FFN (T in )·P Loc +Φ Res_FFN (T in )·PRes

[0042] Among them, T in T out These are the characteristics of a single token at the input and output of the Geo-MOE module, respectively, T in T out ∈R d Φ Loc_FFN Φ Res_FFN These are the FFN expert networks with the highest confidence levels from the location expert group and resolution expert group, respectively, after being filtered by the router. And P... Loc P Res This means that the router is targeting T. in The confidence level of the expert with the highest confidence among the two expert groups selected. The selection process for experts from different geographic information-guided expert groups by the router is shown in the following formula:

[0043] P Loc =max(Φ g (T in )·Φ Loc (I Loc ))

[0044] P Res =max(Φ g (T in )·Φ Res (I Res ))

[0045] Where, Φ g (T in )∈R e Φ represents the number of experts in the expert panel. Loc Φ Res These represent the location-information-guided weight bias function and the resolution-information-guided weight bias function, respectively. The specific calculation process of the weight bias function is shown in the introduction of the geographic information coding module. Finally, the value with the highest confidence level among each expert group is selected as the confidence level, and the corresponding expert is selected for feature processing.

[0046] This invention utilizes Mask-RCNN as a detection module after acquiring the ground feature information of the image to achieve detection. The detection process is as follows: Figure 1As shown in the network structure above, after acquiring image features, multi-scale feature fusion is first performed through the Feature Pyramid Network (FPN) to facilitate the exchange of high-level and low-level features. Then, the Region Proposal Network (RPN) generates candidate regions that may contain the target. For the target candidate regions obtained by the RPN, Region of Interest Align (RoIAlign) uses bilinear interpolation to more accurately align candidate regions of different sizes onto a fixed-size feature map. Finally, the feature maps extracted by RoIAlign are fed into a classification and regression head to predict the category of each candidate region and further refine the precise target bounding box.

[0047] The loss functions used during training mainly consist of ship target classification loss, bounding box regression loss, and candidate target classification loss and bounding box regression loss. These losses measure the difference between the predicted candidate target and final target categories and the true categories, as well as the difference between the predicted bounding boxes and the true bounding boxes. The specific loss calculation is shown in the following formula:

[0048] L=α·L rpn-cls +β·L rpn-bbox +γ·L cls +δ·L bbox

[0049]

[0050] In this invention, the four values ​​α, β, γ, and δ represent the candidate target classification loss, localization loss, and the final detected target classification loss and localization loss, respectively, and are all set to 1. ij y ij Let x represent the predicted probability of the i-th target being in the j-th class, and y represent the true probability (0 or 1) of the i-th target being in the j-th class. This invention predicts targets for a single class, ships; therefore, there are only two possibilities: background (class 0) and ship (class 1). M and N represent the number of predicted candidate targets and the number of ultimately predicted targets, respectively. i y i w i h i These are the x and y coordinates of the candidate target's center, as well as the candidate target's width and height dimensions. And... These are the center coordinates of the paired ground truth target and the width and height dimensions of the ground truth target, respectively.

[0051] Step 3: After constructing the deep neural network for ship target detection based on the geographic information hybrid expert model in Step 2, train it using the PyTorch deep learning framework. First, in the PyTorch environment, a truncated normal distribution initialization method is used to assign values ​​to the weights and bias parameters of each layer of the network, setting the standard deviation to 0.02. This method ensures that the initial parameters are randomly distributed within a specific range, effectively avoiding the model getting trapped in local optima due to excessive concentration of initial parameter values. Then, image data and its corresponding image center latitude and longitude, and image spatial resolution information are input into the network to obtain the predicted detection results. This output provides information about the ship target, specifically covering the classification confidence and bounding box (Bbox) information for both candidate and final ship targets. Following the loss function calculation method described in Step 3, the loss function value between the predicted result and the real data label is calculated. The Adam optimizer is then used to iteratively optimize the weights and bias parameters of each layer of the network, continuously pushing the network towards reducing the loss function value. During the training process, the convergence of the network output results is closely recorded and observed. Convergence is determined by periodically evaluating model performance on the test set, with specific metrics including mAP, mAP50, mAP75, mAPs, mAPm, and mAPl. mAP refers to the average precision, specifically the average AP value when the Intersection over Union (IoU) ranges from 0.5 to 0.95 and the step size is 0.05. mAP50 and mAP75 correspond to the AP values ​​at IoU thresholds of 0.5 and 0.75, respectively. When these metrics stabilize over several consecutive training epochs, the weights and bias parameters of each layer of the network at this point are saved; these parameters constitute the trained network parameters.

[0052] Step Four: Remote Sensing Image Target Detection. In the actual detection process, the remote sensing image to be detected is input into the network trained in Step Three. This image carries corresponding latitude and longitude geographic information, which is used to accurately locate the image's position in geographic space; it also contains spatial resolution geographic information, which clarifies the actual ground distance represented by a single pixel in the image. The network uses a geographic information encoding module to encode the resolution and latitude / longitude information, and then sends the encoding results to the GeoMoE feature extraction module. This module uses this encoded information to assist in extracting ground feature information from the remote sensing image. Based on the extracted features, the network sequentially performs candidate target detection and fine-grained detection of candidate targets, filtering out targets with a confidence level higher than a threshold of 0.05. Finally, the network outputs the ship target location information in the form of bounding boxes (Bboxes), which is the network's detection result. This invention provides an efficient and convenient method for automated data analysis and information extraction of massive amounts of remote sensing data.

[0053] The following is a complete embodiment to demonstrate the practicality of the present invention:

[0054] The dataset used in this invention originates from visible light remote sensing image data from Google Earth, covering five levels of spatial resolution: 1.07m, 2.15m, 4m, 8m, and 16m. In the experiments, to ensure consistency of training parameters across different models, the initial learning rate was set to 0.0001, momentum to 0.9, and weight decay coefficient to 0.05. The model was trained for 100 epochs, with the learning rate reduced to one-tenth of its original value during the 30th and 60th training epochs.

[0055] The evaluation method uses COCO's evaluation metrics, specifically including mAP, mAP50, mAP75, mAPs, mAPm, and mAPl. Here, mAP represents the average accuracy, which is the average AP value when the Intersection over Union (IoU) is between 0.5 and 0.95, with a step size of 0.05; mAP50 and mAP75 correspond to the AP values ​​when the IoU thresholds are 0.5 and 0.75, respectively. mAPs, mAPm, and mAPl are used to evaluate the network's ability to detect ships at multiple scales, with mAPs representing small-scale ships (pixel area < 32 pixels). 2 The mAP value of a medium-sized ship (32) is given by mAPm. 2 <Pixel area<96 2 The mAP value, where mAPl represents a large-scale ship (pixel area > 96). 2 The mAP values ​​of the samples are shown in Table 2, which compares them with other detection methods.

[0056]

[0057]

[0058] Table 2

[0059] Experimental results show that the target detection network designed in this invention achieves a significant improvement in detection accuracy compared to the Mask-RCNN prototype, effectively validating the effectiveness of the various structures involved in this invention. A comprehensive analysis of the various metrics of this invention compared to other detection methods clearly demonstrates its advantages. Taking the mAP50 metric as an example, this invention, with an excellent score of 0.618, surpasses all comparable methods. This means that this invention exhibits superior detection performance even with relatively relaxed requirements for target overlap. In practical applications, this invention can stably and reliably achieve target detection for small- and medium-scale ship targets, as well as tasks with more flexible requirements for detection accuracy, highlighting its high practical value.

Claims

1. A method for detecting ship targets in remote sensing images based on a geographic information hybrid expert model, characterized in that: The steps are as follows: Step 1: Computer reads data; First, use a computer to read the remote sensing image data and simultaneously load the latitude, longitude, and resolution information corresponding to the image; After reading the data, perform data augmentation on the image; Step 2: Construct a deep neural network for ship target detection based on a geographic information hybrid expert model; A geographic information encoder is designed to encode the loaded latitude, longitude, and resolution information. After geographic information encoding, a feature extraction module based on a geographic information hybrid expert model, namely the GeoMoE feature extraction module, is designed to extract features from the loaded image. Then, a Mask-RCNN object detection module is used to integrate the extracted features, enabling the network to perceive ship targets in multi-resolution complex background images. Finally, the detection part uses the Mask-RCNN network framework to receive multi-scale image features containing geographic information extracted from the geographic information hybrid expert model. Then, a region proposal network is used to generate a series of candidate regions that may contain ship targets. The position of the ship target is accurately regressed using the feature maps of candidate regions of different sizes to obtain the ship target position. Step 3: Train the deep neural network; In the PyTorch deep learning framework, the ship target detection deep neural network is trained using labeled training data until the performance of the target detection deep neural network reaches its optimal value, and the parameters of the ship target detection deep neural network at this time are recorded. Step 4: Remote sensing image target detection; using a trained deep neural network for ship target detection based on a geographic information hybrid expert model, small ship targets are detected in multi-scale complex environment remote sensing image detection data containing latitude, longitude and resolution information, and the ship position information is output.

2. The method for detecting ship targets in remote sensing images based on a geographic information hybrid expert model according to claim 1, characterized in that: In step two, the GeoMoE hybrid expert model based on geographic information consists of two different groups of experts: one guided by resolution geographic information and the other guided by latitude and longitude geographic information. The experts guided by resolution geographic information are responsible for perceiving ship targets at different scales, while the experts guided by latitude and longitude geographic information focus on identifying and adapting to the diverse characteristics of geographic regions.

3. The method for detecting ship targets in remote sensing images based on a geographic information hybrid expert model according to claim 1, characterized in that: In step two, the processing method for geographic information-guided experts is as follows: the loaded geographic information is encoded separately; for one-dimensional spatial resolution information, the implicit information is encoded using a multilayer perceptron and softmax function, and transformed into a spatial resolution expert group confidence probability distribution; for latitude and longitude information, sine and cosine coding combined with multilayer perceptron (MLP) and softmax is used to transform discrete latitude and longitude coordinates into a continuous encoding form that includes location features, and finally obtain the latitude and longitude expert group confidence probability distribution.

4. A method for detecting ship targets in remote sensing images based on a geographic information hybrid expert model according to claim 1, 2, or 3, characterized in that: Because geographic information contains relative spatial location information, which differs from the simple quantitative relationship of spatial resolution, the encoding methods for latitude and longitude information differ from those for spatial resolution. The processing flow for both is shown in the following formula: Conference Loc =Φ softmax (F MLPl (F PE (I Loc ))) Conference Res =Φ softmax (F MLPr (I Res )) Among them, I Loc ∈R 2 It is the center latitude and longitude location information of the image, I Res ∈R 1 It is the spatial latitude and longitude information of the image; Φ MLPl Φ MLPr These are latitude and longitude multilayer perceptrons and spatial resolution multilayer perceptrons, which respectively convert latitude and longitude location information and spatial resolution information into scores for each expert in their respective expert groups.

5. The method for detecting ship targets in remote sensing images based on a geographic information hybrid expert model according to claim 4, characterized in that: Use Φ softmax The function transforms the scores into a probability distribution such that each expert's score is between 0 and 1, and the sum of all experts' scores in the expert group is 1; Φ PE It is a function that performs sine and cosine encoding on latitude and longitude; Φ softmax Φ PE The specific formula is as follows: Where N is the total number of experts in the expert group, and the number of experts in both expert groups is set to 4. i It is the probability distribution for each expert. For each expert, the probability distribution value z i Perform exponential operations with the natural constant e as the base; Φ PE The latitude and longitude of the image are encoded and converted, where lat and lon are the latitude and longitude values, respectively, transforming the range from [-90, 90] and (-180, 180] to (-1, 1] and (-1, 1]; d model It is the encoded latitude and longitude feature latitude, which must be an integer multiple of 4.

6. A method for detecting ship targets in remote sensing images based on a geographic information hybrid expert model according to claim 1 or 2, characterized in that: The GeoMoE feature extraction module replaced the FFN of a portion of the blocks; the process in each block is described in the following equation: T i =T i-1 +F Norm (F Attn (T i-1 ))+Φ Norm (F GMF (T i-1 +F Norm (F Attn (T i-1 )),I Geo )) Where, Φ Attn The multi-head self-attention layer is performed in a local window, Φ Norm The normalization layer normalizes the input features; some layers use the original FFN from the Swin-Transformer, while others use the GeoMOE module; T i-1 T i T represents the tokens for the input and output of the i-th block, respectively. i-1 T i ∈R L×d L is the length of the input tokens, and d is the dimension of the feature; I Geo ∈R L×d I Geo The geographic information corresponding to the image is specifically the image spatial resolution, the longitude of the image center, and the latitude of the image center.

7. The method for detecting ship targets in remote sensing images based on a geographic information hybrid expert model according to claim 6, characterized in that: Φ GMF The specific processing flow of the GeoMoE & FFN modules represented is as follows: T out =Φ Loc_FFN (T in )·P Loc +Φ Res_FFN (T in )·P Rws Among them, T in T out These are the characteristics of a single token at the input and output of the GeoMOE module, respectively, T in T out ∈R d ;Φ Loc_FFN Φ Res_FFN These are the FFN expert networks with the highest confidence levels from the location expert group and resolution expert group, respectively, after being filtered by the router; while P Loc P Res This means that the router is targeting T. in The confidence level of the expert with the highest confidence level among the two expert groups selected; the selection process of experts from different geographic information-guided expert groups by the router is shown in the following formula: P Loc =max(Φ g (T in )·F Loc (I Loc )) P Res =max(Φ g (T in )·F Res (I Res )) Where, Φ g (T in )∈R e , e represents the number of experts in the expert group; Φ Loc Φ Res These represent the location-information-oriented weight bias function and the resolution-information-oriented weight bias function, respectively.

8. A method for detecting ship targets in remote sensing images based on a geographic information hybrid expert model according to claim 1, 2, or 3, characterized in that: The loss function used during training consists of ship target classification loss, bounding box regression loss, and candidate target classification loss and bounding box regression loss. These losses measure the difference between the predicted candidate target and final target categories and the true categories, as well as the difference between the predicted bounding boxes and the true bounding boxes. The specific loss calculation is shown in the following formula: L=α·L rpn-cls +β·L rpn-bbox +γ·L cls +δ·L bbox Where α, β, γ, and δ are the weights of the candidate target classification loss, localization loss, and final target detection classification loss, respectively, and are set to 1; p ij y ij Let x and y represent the predicted probability of the i-th target being of class j, and the true probability of the i-th target being of class j, respectively. Since the prediction is performed for a single class of ships, there are only two possibilities: background class 0 and ship class 1. M and N represent the number of predicted candidate targets and the number of ultimately predicted targets, respectively. i y i w i h i These are the center coordinates of the candidate target and its width and height dimensions; while These are the center coordinates of the paired ground truth target and the width and height dimensions of the ground truth target, respectively.

9. The method for detecting ship targets in remote sensing images based on a geographic information hybrid expert model according to claim 1, characterized in that: In step three, specifically: First, in the PyTorch environment, the weights and bias parameters of each layer of the network are assigned values ​​using the truncated normal distribution initialization method, with a standard deviation of 0.02; the initial parameters are randomly distributed within a specific range. Then, image data and its corresponding image center latitude and longitude, and image spatial resolution information are input into the network to obtain the predicted detection results; the output of the results is related to the ship target, specifically covering the classification confidence and bounding box information of both the candidate ship target and the final target; the loss function value between the predicted results and the real data labels is calculated according to the loss function calculation method; and the weights and bias parameters of each layer of the network are iteratively optimized using the Adam optimizer. During the training process, the convergence of the network output results is closely recorded and observed; the convergence status is determined by periodically evaluating the model performance on the test set.

10. The method for detecting ship targets in remote sensing images based on a geographic information hybrid expert model according to claim 1, characterized in that: In step four, the remote sensing image to be detected is input into the network trained in step three. The image carries the corresponding latitude and longitude geographic information, which is used to accurately locate the image's position in geographic space. It also contains spatial resolution geographic information, which clarifies the actual ground distance represented by a single pixel in the image. The geographic information encoding module encodes the resolution information and latitude and longitude information, and then sends the encoding result to the GeoMoE feature extraction module. This module assists in extracting ground feature information from remote sensing images; Then, candidate target detection and fine detection of candidate targets are carried out in sequence, and targets with a confidence level higher than the 0.05 threshold are selected. Finally, the network outputs the ship target location information in the form of bounding boxes (Bboxes), which is the network's detection result.

Images