Satellite target detection method and system based on multi-scale feature fusion enhancement

Abstract

This invention discloses a satellite target detection method and system based on multi-scale feature fusion enhancement. The method includes: constructing a satellite target detection dataset and an initial satellite target detection model, wherein the initial satellite target detection model includes a backbone network, a neck network, and a detection head; introducing a multi-scale feature fusion module at the connection between the backbone network and the neck network, introducing an efficient upsampling module in the upsampling path of the neck network, and introducing a multi-branch fusion perception module before the detection head to obtain an improved model; training the improved model using the simulation dataset to obtain a trained model; inputting the satellite image to be detected into the trained model and outputting the detection result. This invention, through the collaborative design of three modules, systematically solves the problems of insufficient fusion of shallow and deep features, blurred upsampling edges, and a single receptive field, significantly improving the detection accuracy and robustness of weak satellite targets at multiple scales in space while maintaining a lightweight design.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a satellite target detection method and system based on multi-scale feature fusion enhancement. Background Technology

[0002] With the development of aerospace technology, accurate detection of non-cooperative targets in space (such as space debris and faulty satellites) is a prerequisite for performing tasks such as on-orbit servicing, debris removal, and space surveillance. Currently, the mainstream detection scheme is based on deep learning-based target detection models, such as the YOLO series algorithms. However, when applied to space satellite target detection, the following technical drawbacks still exist:

[0003] 1. Existing space target image datasets are scarce and feature limited scenes, mostly consisting of pure black deep space backgrounds. They lack simulations of Earth backgrounds, complex lighting, and various satellite types, resulting in insufficient model generalization ability.

[0004] 2. During multiple downsampling processes in deep networks, general detection models are prone to losing key edge and texture information of microsatellite targets. Furthermore, their traditional feature fusion methods cannot effectively bridge the semantic-detail gap between shallow geometric features and deep semantic features, resulting in weak multi-scale target detection capabilities.

[0005] 3. Conventional nearest neighbor interpolation upsampling methods in feature pyramid networks introduce blockiness and jagged edges, which destroy the contour integrity of small targets. At the same time, single receptive field convolution in front of the detector head is difficult to simultaneously take into account the overall contour of the satellite and the features of local fine components such as antennas and solar panels. Summary of the Invention

[0006] The purpose of this invention is to provide a satellite target detection method and system based on multi-scale feature fusion enhancement, which solves the problems of low accuracy, high false negative rate and incomplete feature representation caused by coarse upsampling and single receptive field in the existing technology for detecting weak satellite targets at multiple scales in space.

[0007] The first aspect of this invention provides a satellite target detection method based on multi-scale feature fusion enhancement, comprising the following steps:

[0008] A simulation dataset is constructed, which is derived from satellite targets of various types, attitudes, and complex lighting backgrounds.

[0009] An initial satellite target detection model is constructed, which includes a backbone network, a neck network, and a detection head. A multi-scale feature fusion module is introduced at the connection between the backbone network and the neck network. An efficient upsampling module is introduced in the upsampling path of the neck network. A multi-branch fusion sensing module is introduced in the feature aggregation area in front of the detection head to obtain an improved satellite target detection model.

[0010] The improved satellite target detection model is trained using the satellite target simulation dataset to obtain a trained satellite target detection model;

[0011] The satellite image to be detected is input into the trained satellite target detection model to obtain the satellite target detection result.

[0012] In this solution, the construction of the simulation dataset specifically includes:

[0013] Obtain three-dimensional models of various typical satellite structures, including at least cubic bus structures, cylindrical bus structures, and solar panel structures of different configurations.

[0014] Based on the three-dimensional model, satellite foreground material is generated by multi-angle rendering using a virtual camera along a preset observation path.

[0015] A multi-source background image library is constructed, and the satellite foreground material is fused with the background image to generate a synthetic satellite image;

[0016] The satellite targets in the synthetic satellite image are labeled to generate training label files, so as to obtain the simulation dataset of satellite targets.

[0017] In this solution, the construction of a multi-source background image library and the fusion of the satellite foreground material with the background image specifically includes:

[0018] Construct a multi-source background image library containing deep-space star images and Earth remote sensing images;

[0019] During the fusion process, the brightness histogram features of the background image are analyzed, and the pixel intensity of the foreground satellite is automatically adjusted based on the illumination consistency adaptive correction mechanism to match the illumination conditions of the foreground and the background.

[0020] Gaussian blur is applied at the boundary where the foreground and background blend to eliminate compositing artifacts.

[0021] In this scheme, the multi-scale feature fusion module is used to aggregate information from multiple receptive fields within the same feature level through a hierarchical feature extraction chain of channel decoupling and cascaded deep convolutions. Specifically, it includes the following steps:

[0022] The input feature tensor is adjusted to the hidden layer dimension through projection transformation, and channel decoupling operation is performed, dividing it into an identity mapping branch for maintaining gradient backpropagation and a processing branch for feature extraction.

[0023] A cascaded depthwise convolution sequence is constructed for the processing branch. The depthwise convolutions with different kernel sizes are sequentially connected, and dense connections are used to achieve dense reuse of features in order to capture point-like, local and global features.

[0024] The output of the identity mapping branch is concatenated with the outputs of all cascaded branches, and then aggregated through convolution to obtain multi-scale fused features.

[0025] In this scheme, the high-efficiency upsampling module is used to perform depthwise convolutional smoothing and channel shuffling operations after feature upsampling through a reconstruction-interaction mechanism, specifically including the following steps:

[0026] In the first stage, a physical upsampling operation is performed, and a depthwise convolution is immediately introduced to smooth the features of the upsampled feature map in order to eliminate the blocky effect and jagged edges introduced by interpolation.

[0027] In the second stage, a channel shuffling operation is performed to reshape and transpose the feature tensor so that the features of different groups are evenly mixed in the channel dimension, thereby breaking the channel information blockage caused by depth convolution.

[0028] In the third stage, cross-channel information is integrated through point convolution to output an enhanced feature map with clear edges.

[0029] In this scheme, the multi-branch fusion sensing module is used to integrate densely connected branches and dilated convolutional branches with different dilation rates, specifically including the following steps:

[0030] The input feature tensor is decoupled into two parts, which are then used as dense and hollow branches respectively.

[0031] In dense branches, continuous convolutional layers and dense connection structures are used to enhance the reuse of high-frequency local detail features and gradient propagation.

[0032] In the dilated branch, a set of dilated convolutions with different dilation rates are processed in parallel to exponentially expand the receptive field without reducing the resolution, so as to capture long-range contextual dependencies.

[0033] By using residual connections, the input feature tensor, dense branch output, and dilated branch output are fused to obtain enhanced features that combine local details with global semantics.

[0034] A second aspect of the present invention also provides a satellite target detection system based on multi-scale feature fusion enhancement, comprising a memory and a processor. The memory includes a satellite target detection method program based on multi-scale feature fusion enhancement. When the processor executes the satellite target detection method program based on multi-scale feature fusion enhancement, it performs the following steps:

[0035] A simulation dataset is constructed, which is derived from satellite targets of various types, attitudes, and complex lighting backgrounds.

[0036] An initial satellite target detection model is constructed, which includes a backbone network, a neck network, and a detection head. A multi-scale feature fusion module is introduced at the connection between the backbone network and the neck network. An efficient upsampling module is introduced in the upsampling path of the neck network. A multi-branch fusion sensing module is introduced in the feature aggregation area in front of the detection head to obtain an improved satellite target detection model.

[0037] The improved satellite target detection model is trained using the satellite target simulation dataset to obtain a trained satellite target detection model;

[0038] The satellite image to be detected is input into the trained satellite target detection model to obtain the satellite target detection result.

[0039] A third aspect of the present invention provides a computer-readable storage medium comprising a machine program for a satellite target detection method based on multi-scale feature fusion enhancement, wherein when the program is executed by a processor, it implements the steps of the satellite target detection method based on multi-scale feature fusion enhancement as described in any of the preceding claims.

[0040] A fourth aspect of the present invention provides a computer program product comprising computer program code, wherein when the computer program code is run on a computer, the computer implements the steps of a satellite target detection method based on multi-scale feature fusion enhancement as described in any of the preceding claims.

[0041] A fifth aspect of the present invention provides an electronic device comprising: a processor and a memory; wherein the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the steps of a satellite target detection method based on multi-scale feature fusion enhancement as described in any of the preceding claims.

[0042] The satellite target detection method and system based on multi-scale feature fusion enhancement disclosed in this invention have the following beneficial effects: 1. By constructing a simulation dataset from satellite targets of various types, attitudes, and complex lighting backgrounds, the problem of scarce real on-orbit image data and single scene was solved, providing high-quality and diverse data support for model training and significantly improving the model's generalization ability.

[0043] 2. By integrating a multi-scale feature fusion module, an efficient upsampling module, and a multi-branch fusion perception module into the architecture of the initial model, structural improvements are completed directly during the model building stage. This enables the improved model to bridge the gap between shallow and deep features, restore clear edges of small targets, and capture local details and global contours simultaneously from the beginning of training.

[0044] 3. By directly applying the simulation dataset to the training of the improved model, a complete technical loop from data construction and model improvement to model training is formed, ensuring that the improved network architecture can fully learn the multi-scale and multi-pose satellite target features in the simulation data.

[0045] 4. By decoupling the channels and designing cascaded deep convolutions in the multi-scale feature fusion module, the gap between shallow geometric details and deep semantic information is effectively bridged, significantly improving the detection capability of multi-scale targets.

[0046] 5. By using the depth convolution smoothing + channel shuffling post-processing mechanism of the efficient upsampling module, the edge blurring and block effect caused by traditional upsampling are solved, the contour integrity of small targets is restored, and the positioning accuracy is improved.

[0047] 6. Through the local-global dual-path sensing mechanism of the multi-branch fusion sensing module, the model can adaptively capture the fine component features and overall contour features of the satellite simultaneously, enhancing the robustness of non-cooperative target identification in complex backgrounds. Attached Figure Description

[0048] Figure 1 The diagram illustrates the steps of a satellite target detection method based on multi-scale feature fusion enhancement according to the present invention.

[0049] Figure 2 A schematic diagram illustrating the dataset construction process of a satellite target detection method based on multi-scale feature fusion enhancement according to the present invention is shown.

[0050] Figure 3 A schematic diagram illustrating a satellite target detection method based on multi-scale feature fusion enhancement according to the present invention is shown.

[0051] Figure 4 The diagram illustrates the distribution of satellite target data generated by the satellite target detection method based on multi-scale feature fusion enhancement according to the present invention.

[0052] Figure 5 The overall network architecture diagram of ST-YOLO, a satellite target detection method based on multi-scale feature fusion enhancement according to the present invention, is shown.

[0053] Figure 6 A schematic diagram of a multi-scale feature fusion module for a satellite target detection method based on multi-scale feature fusion enhancement according to the present invention is shown.

[0054] Figure 7 A schematic diagram of an efficient upsampling reconstruction module for a satellite target detection method based on multi-scale feature fusion enhancement according to the present invention is shown.

[0055] Figure 8 A schematic diagram of a multi-branch fusion sensing module for a satellite target detection method based on multi-scale feature fusion enhancement according to the present invention is shown.

[0056] Figure 9 A block diagram of a satellite target detection system based on multi-scale feature fusion enhancement according to the present invention is shown. Detailed Implementation

[0057] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0058] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0059] Currently, the implementation schemes most similar to this invention are mainly the following: Scheme 1, Scheme 2, or Scheme 3.

[0060] Specifically, Option 1: Directly use a general target detection model (such as YOLO11) for satellite detection. Its disadvantage is that the backbone network of the model will lose the edge details of small satellite targets when downsampling multiple times, and the feature fusion method of the neck network fails to effectively combine shallow geometric information and deep semantic information, resulting in a high rate of missed detection of small targets.

[0061] Specifically, Option 2: The traditional nearest neighbor interpolation method is used for feature map upsampling. Its disadvantage is that when processing low-resolution small target feature maps, this method introduces significant blockiness and jagged edges, which destroys the integrity of the satellite target outline and affects the subsequent positioning accuracy.

[0062] Specifically, Option 3: Using a standard single-branch convolutional structure as the feature aggregation module in front of the detection head has the disadvantage that the receptive field of a single convolutional kernel is limited, making it difficult to simultaneously take into account the overall outline of the satellite body and the features of fine components such as solar panels and antennas, resulting in incomplete feature representation.

[0063] Through in-depth analysis of the above-mentioned existing technical solutions, several key drawbacks of existing technologies in solving frame-averaged images can be clearly summarized:

[0064] 1. The lack of training data and the limited scenarios in existing datasets result in limited generalization ability of the models;

[0065] 2. Downsampling leads to the loss of information about small targets, and the fusion of shallow and deep features is insufficient, resulting in weak multi-scale detection capabilities;

[0066] 3. Conventional upsampling operations cause edge blurring, and a single receptive field cannot simultaneously focus on local details and global contours, resulting in non-robust feature representation.

[0067] To address these shortcomings, the present invention aims to provide a satellite target detection method and system based on multi-scale feature fusion enhancement, wherein the present invention seeks to achieve the following objectives:

[0068] (1) Construct a high-fidelity, multi-condition satellite simulation dataset to provide high-quality data for model training;

[0069] (2) In the model building stage, multi-scale feature fusion, efficient upsampling and multi-branch fusion perception mechanism are introduced in an integrated manner to solve the problems of insufficient feature fusion, blurred upsampling edges and single receptive field from the architecture level;

[0070] (3) The simulation dataset is directly applied to the training of the improved model, forming a complete technology chain of data construction-model improvement-model training. While keeping the model lightweight, the detection accuracy and robustness of weak satellite targets at multiple scales in complex backgrounds are significantly improved.

[0071] Therefore, the core innovation and key protection points are:

[0072] 1. "Data-Model" Closed-Loop Solution: Construct a simulation dataset from satellite targets of multiple types, attitudes and complex lighting backgrounds, and simultaneously improve the architecture during the model building phase, so that the simulation dataset and the improved model form a direct application closed loop during the training phase.

[0073] 2. Internal structure of the multi-scale feature fusion module: Channel decoupling and cascaded deep convolutional sequences are introduced, and a method of multi-receptive field aggregation is achieved by densely connecting them at the same level.

[0074] 3. Internal structure of the high-efficiency upsampling module: It adopts a reconstruction-interaction process of upsampling + depth convolution smoothing + channel shuffling + point convolution fusion to restore the edge clarity of small targets.

[0075] 4. Internal structure of the multi-branch fusion perception module: Design of a local-global dual-path perception mechanism that integrates densely connected branches and multi-diffraction dilated convolution branches.

[0076] Specifically, Figure 1 The diagram illustrates the steps of a satellite target detection method based on multi-scale feature fusion enhancement according to the present invention.

[0077] like Figure 1 As shown, this invention discloses a satellite target detection method based on multi-scale feature fusion enhancement, comprising the following steps:

[0078] S102, Construct a simulation dataset, which is derived from satellite targets of various types, attitudes, and complex lighting backgrounds;

[0079] S104, Construct an initial satellite target detection model, which includes a backbone network, a neck network, and a detection head;

[0080] S106, The improved satellite target detection model is trained using the satellite target simulation dataset to obtain a trained satellite target detection model;

[0081] S108, input the satellite image to be detected into the trained satellite target detection model to obtain the satellite target detection result.

[0082] It should be noted that, in this embodiment, the key technical point of the present invention lies in proposing an enhanced detection network architecture comprising three modules: a multi-scale feature fusion module, an efficient upsampling module, and a multi-branch fusion perception module. This architecture is directly integrated into the model building step, forming a complete closed loop from data to training with the simulation dataset constructed by the present invention. Specifically, the present invention is applied to tasks such as on-orbit service and space debris removal in the field of space situational awareness. In these scenarios, the present invention can effectively overcome challenges such as drastic changes in space illumination, interference from the Earth's background, and large imaging distance spans, and can detect non-cooperative satellite targets in the field of view in real time and with high precision. In particular, it has a stronger perception capability for distant and weak targets. After being optimized and deployed on edge devices such as NVIDIA Jetson AGX Orin, it can achieve real-time processing, providing fast and reliable technical support for space missions.

[0083] According to an embodiment of the present invention, the construction of the simulation dataset specifically includes:

[0084] Obtain three-dimensional models of various typical satellite structures, including at least cubic bus structures, cylindrical bus structures, and solar panel structures of different configurations.

[0085] Based on the three-dimensional model, satellite foreground material is generated by multi-angle rendering using a virtual camera along a preset observation path.

[0086] A multi-source background image library is constructed, and the satellite foreground material is fused with the background image to generate a synthetic satellite image;

[0087] The satellite targets in the synthetic satellite image are labeled to generate training label files, so as to obtain the simulation dataset of satellite targets.

[0088] According to an embodiment of the present invention, the construction of a multi-source background image library and the fusion of the satellite foreground material with the background image specifically includes:

[0089] Construct a multi-source background image library containing deep-space star images and Earth remote sensing images;

[0090] During the fusion process, the brightness histogram features of the background image are analyzed, and the pixel intensity of the foreground satellite is automatically adjusted based on the illumination consistency adaptive correction mechanism to match the illumination conditions of the foreground and the background.

[0091] Gaussian blur is applied at the boundary where the foreground and background blend to eliminate compositing artifacts.

[0092] It should be noted that, in this embodiment, the training of the deep learning model requires a large amount of diverse data. Since it is difficult to capture photos of non-cooperative targets in real space, and most existing public datasets have relatively simple scenes, it is difficult to meet the requirements of high-precision detection. Therefore, this invention adopts a method based on "3D rendering and image synthesis". This method uses high-quality satellite material and then synthesizes it with the cosmic background to generate realistic training data in batches. Since non-cooperative targets in space usually involve reconnaissance or core communication assets, their real on-orbit image data is highly sensitive information and is extremely difficult to obtain through public channels. This data barrier seriously restricts the development of deep learning-based space perception technology. To this end, based on an in-depth analysis of the characteristics of existing datasets, this invention adopts a technical route based on CAD simulation and image synthesis to construct the ST-Satellite dataset.

[0093] Specifically, in this embodiment, as Figure 2As shown, this is a schematic diagram of the dataset construction process. In order to overcome the problem of the single target of existing datasets, this invention selects typical satellite 3D models as the basic data. These models cover cubic bus, cylindrical bus and solar panel structures of different configurations, which can represent the current mainstream satellite forms. In SolidWorks, a virtual camera is set up and a spherical observation path is created around the satellite model. By adjusting the camera angle, a high-resolution rendering output is performed every 30° to generate a transparent PNG format foreground material. This process ensures that the satellite edges are smooth and avoids the jagged noise that may be generated in traditional screenshot methods.

[0094] Furthermore, in this embodiment, to more realistically simulate the on-orbit imaging environment, the present invention constructs a multi-source background library containing deep-space star images and real Earth remote sensing images, and designs an illumination consistency adaptive correction mechanism. This mechanism automatically adjusts the pixel intensity of the foreground satellite by analyzing the brightness histogram features of the background image. For example, when the background is a shadowed area of ​​the Earth, the satellite brightness is automatically reduced and the contrast is compressed; while when the background is a brightly lit area, the highlight reflection features are enhanced. In addition, at the boundary where the foreground and background merge, the present invention applies Gaussian blur to eliminate hard artifacts that may be introduced by artificially synthesized images, further enhancing the geometric realism and scale diversity of the image.

[0095] Furthermore, in this embodiment, the dataset constructed by the present invention takes into account the diversity of targets during its design, ensuring that the model can handle satellite targets of different sizes, locations and background conditions. However, the complexity of the dataset is still limited. For example, the performance of satellites under extreme lighting conditions and the performance of highly reflective materials in specific environments may not be fully covered. In the future, the dataset can be further expanded to consider more environmental changes and satellite types.

[0096] Furthermore, in this embodiment, as Figure 3 As shown, the diagram is labeled, in which, Figure 3 In this context, "satellite" refers to a satellite. During the annotation process, the LabelImg tool was used for manual annotation to ensure accuracy and efficiency. LabelImg is a widely used image annotation tool that allows users to select targets in an image and generate labels. Using this tool, researchers can quickly and accurately label satellite targets in images and generate label files conforming to the YOLO format. Furthermore, the automatically generated labels were manually verified during the annotation process to ensure high accuracy. This process provides precise data support for subsequent model training and ensures the consistency of annotations across the dataset.

[0097] Furthermore, in this embodiment, as Figure 4As shown, the diagram illustrates the distribution of generated satellite target data. Analysis reveals that the targets are positioned relatively evenly across the image, almost covering the entire image area. Most targets are distributed in the lower x and y coordinate regions, indicating that these targets may be concentrated in the image center or other common areas. Furthermore, there is a correlation between the width and height of the targets; as the width increases, the height also shows a certain upward trend. Most targets are relatively small, with only a small number being larger. This data distribution reflects the rationality of the dataset used in this invention. The uniform distribution of targets ensures the model's target detection capability at different locations, while the relationship between target sizes provides valuable statistical information for training. In this way, the model can learn the features of targets at different scales, ensuring accurate target detection under diverse location and scale conditions. In particular, the diversity of the dataset helps improve the model's generalization ability, enabling it to handle different backgrounds, lighting conditions, and target sizes.

[0098] Finally, in this embodiment, the constructed ST-Satellite dataset contains 2876 high-resolution samples. To ensure the rigor and sufficiency of the experimental evaluation, the dataset is randomly divided into training set, validation set and test set in a 6:2:2 ratio. This division strategy not only ensures that the model can be fully trained, but also reserves enough test data to evaluate the model's generalization ability in unknown and complex scenarios.

[0099] It should be noted that, in this embodiment, in response to the challenges of large scale span, weak texture features and strong background interference in the space non-cooperative target detection task, the present invention improves on the basic architecture of YOLO11 and proposes an enhanced satellite target detection network - ST-YOLO. Although YOLO11 performs well in general target detection, it is prone to losing edge detail information for small targets during deep network downsampling.

[0100] To address this, ST-YOLO has optimized its structure in three key dimensions: feature fusion, upsampling reconstruction, and receptive field expansion. Specifically, this includes: introducing a multi-scale feature fusion module (MSF) at the connection between the main body and the neck to solve the problem of insufficient fusion of shallow geometric information and deep semantic information; replacing traditional interpolation upsampling with an efficient upsample block (EUB) in the downlink path of the feature pyramid to solve the blurring problem of small target edges during feature map magnification; and designing a multi-branch fusion perception module (MBF) in the feature aggregation area in front of the detection head to address the issue that a single receptive field cannot simultaneously capture the features of both the satellite body and fine components. Accordingly, the overall network architecture of ST-YOLO is as follows: Figure 5 As shown.

[0101] According to an embodiment of the present invention, the multi-scale feature fusion module is used to aggregate multiple receptive field information within the same feature level through a hierarchical feature extraction chain of channel decoupling and cascaded deep convolution, specifically including the following steps:

[0102] The input feature tensor is adjusted to the hidden layer dimension through projection transformation, and channel decoupling operation is performed, dividing it into an identity mapping branch for maintaining gradient backpropagation and a processing branch for feature extraction.

[0103] A cascaded depthwise convolution sequence is constructed for the processing branch. The depthwise convolutions with different kernel sizes are sequentially connected, and dense connections are used to achieve dense reuse of features in order to capture point-like, local and global features.

[0104] The output of the identity mapping branch is concatenated with the outputs of all cascaded branches, and then aggregated through convolution to obtain multi-scale fused features.

[0105] It should be noted that, in this embodiment, in the deep convolutional neural network, shallow features are rich in geometric details, while deep features possess strong semantic information. To address the semantic-detail gap problem in satellite target detection, this invention proposes a multi-scale feature fusion module (MSF) based on cascaded deep convolutions, specifically as follows: Figure 6As shown, this module aims to aggregate information from multiple receptive fields within the same feature level by constructing a hierarchical feature extraction chain. Furthermore, the multi-scale feature fusion module abandons the traditional single-path convolution extraction method and instead adopts a strategy that combines channel decoupling and cascaded flow. Its core idea is to divide the feature channels into a "reservation group" and a "processing group": the reservation group directly transmits the original features to maintain the effective backpropagation of gradients; the processing group performs serial feature extraction through a set of incremental convolution kernels. This design not only captures point-like, local, and global features using different kernel sizes, but also implicitly expands the equivalent receptive field through the cascaded structure.

[0106] Specifically, in this embodiment, the input feature tensor is defined as... First, the channel is adjusted to the hidden layer dimension through projection transformation. And perform channel decoupling operation:

[0107] ;

[0108] ;

[0109] in, It is a 3×3 convolution. , These represent the identity mapping branch and the branch to be processed, respectively. For the processing branch... Construct a cascaded depthwise convolution sequence, and set... For the first The output of the level feature extraction is recursively defined as follows:

[0110] ;

[0111] Here, Indicates the kernel size as The depthwise separable convolutions and concatenated inputs employ a concatenation operation to achieve dense feature reuse. Finally, the features of the identity branch and all concatenated branches are aggregated and mapped. .

[0112] According to an embodiment of the present invention, the efficient upsampling module is used to perform depthwise convolutional smoothing and channel shuffling operations after feature upsampling through a reconstruction-interaction mechanism, specifically including the following steps:

[0113] In the first stage, a physical upsampling operation is performed, and a depthwise convolution is immediately introduced to smooth the features of the upsampled feature map in order to eliminate the blocky effect and jagged edges introduced by interpolation.

[0114] In the second stage, a channel shuffling operation is performed to reshape and transpose the feature tensor so that the features of different groups are evenly mixed in the channel dimension, thereby breaking the channel information blockage caused by depth convolution.

[0115] In the third stage, cross-channel information is integrated through point convolution to output an enhanced feature map with clear edges.

[0116] It should be noted that, in this embodiment, feature upsampling is a crucial step for the decoder to restore spatial resolution. The native YOLO11 architecture uses nearest-neighbor interpolation. While this method is computationally efficient, it introduces significant blockiness and jagged edges when processing low-resolution feature maps of small targets, disrupting the contour integrity of the satellite target. Therefore, this invention designs an Efficient Upsample Block (EUB), specifically as follows... Figure 7 As shown.

[0117] Specifically, in this embodiment, the design of the efficient upsampling reconstruction module is based on the reconstruction-interaction concept. First, depthwise convolution is introduced immediately after the upsampling operation for feature smoothing, using the learnable parameters of the convolution kernel to correct noise caused by interpolation. Second, given that depthwise convolution can cause information blockage between channels, a channel shuffling mechanism is introduced to promote information flow between groups by rearranging the channel order, thereby enhancing the robustness of the features. Here, let the input features be... The computation process of the efficient upsampling reconstruction module consists of three stages:

[0118] The first stage involves physical upsampling and feature correction, defining the upsampling operator. With the correction operator :

[0119] ;

[0120] in, It is a 3×3 depthwise convolution. As the activation function, this step improves resolution while performing preliminary filtering of interpolation noise.

[0121] The second stage is channel interaction, which involves channel shuffling to overcome the channel independence limitation imposed by depthwise convolution. :

[0122] ;

[0123] in, For the number of groups, this operation reshapes the feature tensor into... Then, the features of different groups are transposed to ensure uniform mixing across the channel dimension.

[0124] The third stage is linear fusion, achieved through point convolution. Integrate information across channels:

[0125] ;

[0126] Among them, the high-efficiency upsampling reconstruction module effectively improves the feature distortion problem in the upsampling process while maintaining lightweight computing costs, which is especially important for preserving the edge information of small satellite targets.

[0127] According to an embodiment of the present invention, the multi-branch fusion sensing module is used to integrate densely connected branches and dilated convolutional branches with different dilation rates, specifically including the following steps:

[0128] The input feature tensor is decoupled into two parts, which are then used as dense and hollow branches respectively.

[0129] In dense branches, continuous convolutional layers and dense connection structures are used to enhance the reuse of high-frequency local detail features and gradient propagation.

[0130] In the dilated branch, a set of dilated convolutions with different dilation rates are processed in parallel to exponentially expand the receptive field without reducing the resolution, so as to capture long-range contextual dependencies.

[0131] By using residual connections, the input feature tensor, dense branch output, and dilated branch output are fused to obtain enhanced features that combine local details with global semantics.

[0132] It should be noted that, in this embodiment, the visual morphology of spatial targets has highly unstructured features, and background interference is random. Traditional single-receptive-field convolution is difficult to simultaneously focus on local details and suppress global background. Therefore, this invention proposes a multi-branch fusion perception module (MBF), specifically as follows: Figure 8 As shown, a local-global dual-path perception mechanism is constructed by integrating dense connections and dilated convolution.

[0133] Specifically, in this embodiment, the multi-branch fusion perception module includes dense branches and dilated branches. The dense branches utilize consecutive small convolutional kernels and dense connection structures to enhance feature reuse and gradient propagation, focusing on extracting high-frequency local details. The dilated branches utilize dilated convolutions with different dilation rates to exponentially expand the receptive field without reducing resolution, focusing on capturing long-range contextual dependencies. The input feature tensor is also decoupled into... and In this part, dense path A utilizes continuous convolutions and dense connections to enhance high-frequency detail features. Reuse:

[0134] ;

[0135] Void path B utilizes a set of different expansion rates Dilated convolutions capture long-range context. :

[0136] ;

[0137] Finally, the original features and the two-stream features are fused through residual connections:

[0138] .

[0139] Furthermore, in this embodiment, the improved satellite target detection model is then trained using the satellite target simulation dataset to obtain a trained satellite target detection model. Specifically, in practical applications, the satellite image to be detected is input into the trained satellite target detection model to obtain the satellite target detection result.

[0140] Specifically, in this embodiment, the feasibility of the present invention is verified by experimental results and analysis. In order to comprehensively evaluate the performance of the ST-YOLO algorithm in the task of non-cooperative spatial target detection, the present invention conducts a systematic experimental study on the constructed ST dataset. The experimental content mainly includes three parts: First, the specific environment configuration, hyperparameter settings and evaluation index system of the experiment are described; second, ST-YOLO is compared with the current mainstream target detection algorithms to verify its comprehensive performance advantages; finally, ST-YOLO is deployed in embedded devices to evaluate its application effect in resource-constrained environments.

[0141] Furthermore, in this embodiment, the platform is equipped with an Intel(R) Core(TM) i7-9700K CPU @3.60GHz processor and an NVIDIA GeForce RTX 3090 (24GB VRAM) graphics card, providing sufficient computing power for model training and inference. The software environment is based on the Windows 10 operating system, the deep learning framework uses PyTorch 2.5.0, and CUDA is used for GPU acceleration. The model training uses the stochastic gradient descent (SGD) optimizer, with an initial learning rate set to 0.001. To ensure that the model converges sufficiently with limited data and does not fall into overfitting, the number of training epochs is set to 300, and the batch size is set to 8, as shown in Table 1.

[0142] Table 1. Detailed Configuration of Experimental Platform .

[0143] Furthermore, in this embodiment, in order to comprehensively evaluate the performance and applicability of the proposed ST-YOLO model, the present invention conducted systematic comparative experiments on three datasets. The composition of these datasets and the training / validation / testing partitions are summarized in Table 2, providing a clear basis for subsequent evaluation. Several classic object detection models and recent advanced deep learning algorithms were selected as benchmark methods for comparison. Through comprehensive evaluation on different datasets and models, the advantages and potential limitations of ST-YOLO in satellite target detection were analyzed from multiple perspectives. This helps to more accurately evaluate the performance and generalization ability of the model in real-world scenarios.

[0144] Table 2. Experimental Dataset .

[0145] Furthermore, in this embodiment, in order to comprehensively evaluate the performance of the proposed model, this study used four key indicators for quantitative analysis: precision, recall, mAP50, and mAP50-95. These indicators provide a multi-dimensional and objective evaluation of the model's detection capability. Table 3 shows the ablation experiment results for each module. As can be seen from the data in Table 3, each individual module contributes to the model's performance, with the EUB module showing the most significant improvement.

[0146] Table 3. Ablation experiments for each module .

[0147] Specifically, in this embodiment, the accuracy improved from 92.7% to 95.6%, an increase of 3.1%, thus significantly reducing false positives. However, the improvement of the MBF module alone was relatively limited, and some metrics even decreased slightly. This indicates that when used alone, feature redundancy or noise may have been designed. When these modules are used in combination, the performance is significantly improved, especially the MSF+EUB combination, which increased the recall rate to 83.4%, an improvement of 3.6% over the baseline, while mAP50 and mAP50-95 increased to 89.4% and 51.0%, respectively, improvements of 3.2% and 1.1%. This highlights the synergistic effect of multi-scale features and enhanced feature extraction, which significantly improves the model's target detection and localization accuracy. Finally, when all three modules are integrated into the ST-YOLO model, all performance metrics reach peak levels, with recall increasing to 84.2%, mAP50 to 90.1%, and mAP50-95 reaching 52.9%, representing improvements of 4.4%, 3.9%, and 3.0% respectively compared to baseline data.

[0148] In summary, the experimental results of this invention fully demonstrate the effectiveness of each module. The combination of MSF and EUB significantly improves the accuracy of recall and localization, while the ST-YOLO model that fully integrates these three modules performs best on all metrics. This indicates that the synergistic optimization of multiple modules can greatly improve the model's target detection capability, making it more adaptable and robust in complex environments.

[0149] Furthermore, this invention also conducted comparative experiments with mainstream detection models. Specifically, a systematic performance comparison of mainstream object detection models was performed on the ST-Dataset. The experimental results are shown in Table 4. ST-YOLO demonstrated significant advantages in key evaluation metrics, including precision, recall, mAP50, and mAP50-95, clearly demonstrating its powerful object detection capabilities and multi-scale adaptability. Specifically, ST-YOLO achieved a precision of 93.4%, 0.7% higher than YOLO11, indicating a significant improvement in suppressing false detections and higher reliability in classification. Its recall reached 84.2%, 4.4% higher than YOLO11, effectively reducing false negatives. Especially when detecting challenging targets, such as small components, weak texture structures, and complex backgrounds, it exhibited higher detection sensitivity. Moreover, ST-YOLO performed better in mAP50 and mAP50-95, further demonstrating that its feature fusion strategy and multi-branch architecture more effectively capture discriminative information from multi-scale satellite targets, reflecting the effectiveness and robustness of the model design.

[0150] Table 4. Performance comparison with other models on the ST dataset .

[0151] Furthermore, in this embodiment, the proposed satellite target detection model was systematically and comprehensively evaluated on the SciCap dataset. This dataset, constructed using 3D models and rendering techniques, contains high-quality synthetic spacecraft images generated from multiple angles, under different lighting conditions, and in various background environments. These images effectively simulate real on-orbit observation scenarios. The dataset is large in scale, containing 3117 images of satellites and space stations, each with a resolution of 1280×720. It includes 3667 spacecraft instances and 10350 component-level segmentation masks, covering three key parts: the main body, solar panels, and antennas. This dataset has rich structural details and significant diversity, making it an important benchmark for evaluating the generalization ability of spacecraft detection algorithms.

[0152] Furthermore, in this embodiment, the results are shown in Table 5. In terms of accuracy, ST-YOLO achieves 0.783, which is the highest among all models. This indicates that it has stronger positive detection capability in spacecraft detection tasks and can effectively reduce false detections. In contrast, the accuracy of YOLOv9 and YOLO11 remains around 0.75. Based on these results, although ST-YOLO has slightly decreased in other indicators, it still maintains a significant advantage in accuracy. This indicates that it has stronger robustness and resistance to false detections when distinguishing spacecraft targets. At the same time, its overall mAP performance is comparable to the comparative models, further confirming that ST-YOLO can still maintain reliable detection performance in complex multi-scale synthetic spacecraft scenarios.

[0153] Table 5. Performance comparison of different models on the SciCap dataset .

[0154] Finally, in this embodiment, when applied, the present invention obtained a publicly available spacecraft detection dataset from the Roboflow platform to further evaluate the model's generalization ability on synthetic data from different sources. This dataset contains 9448 satellite images generated through synthetic rendering, covering various typical satellite shapes, attitude variations, imaging angles, and background conditions, effectively simulating the diversity of satellite targets in on-orbit observation scenarios. The experimental results are shown in Table 6, where ST-YOLO... It performs best on key metrics, including precision, mAP50, and mAP50-95, demonstrating strong generalization ability on large-scale synthetic satellite imagery. Specifically, ST-YOLO's precision is 0.751, ranking first among all models, indicating its consistent effectiveness in suppressing false detections. Regarding mAP scores, ST-YOLO shows a significant advantage: its mAP50 reaches 0.785, surpassing all comparative models, and its mAP50-95 reaches 0.651, also the highest. These results demonstrate that ST-YOLO maintains stable and accurate recognition capabilities in satellite target identification at different scales and levels of detail complexity, especially in detection tasks involving cross-viewpoints, multiple components, and complex backgrounds, where its advantages are most significant.

[0155] Table 6. Performance comparison of different models on the spacecraft inspection dataset .

[0156] In summary, based on the experimental results of these three datasets, ST-YOLO maintained relatively stable detection performance on satellite target data from different sources, scales, and imaging conditions. On the self-built real dataset, ST-YOLO significantly outperformed YOLO11 in both accuracy and recall. On the SciCap simulation dataset, the model achieved the highest accuracy and maintained strong performance in multi-scale scenarios. On the large-scale simulated Roboflow dataset, ST-YOLO performed comparable to or even better than mainstream YOLO series models in terms of overall mAP. Overall, ST-YOLO demonstrated excellent robustness and generalization ability in both real and simulated scenarios, confirming the effectiveness of its structural design in satellite target detection tasks.

[0157] Furthermore, in this embodiment, as shown in Table 7, the present invention compares and analyzes ST-YOLO with RT-MDET, YOLOX, YOLOv9, YOLOv10, and YOLO11 in terms of inference time, detection frame rate, and number of parameters. In terms of inference time, ST-YOLO's single-image inference time is 9.6 ms, slightly higher than YOLO11's 8.5 ms, but significantly better than RT-MDET (12.6 ms), YOLOX (12.0 ms), and YOLOv9 (14.7 ms), and only slightly lower than YOLOv10 (9.8 ms). This indicates that although the introduction of the improved structure brings some additional computational overhead, ST-YOLO still maintains a relatively fast inference speed overall.

[0158] In terms of detection speed, ST-YOLO achieved a frame rate of 90.09 FPS. Although this is lower than YOLO11's 100 FPS and YOLOv10's 97.09 FPS, it is still significantly higher than RT-MDET's 79.5 FPS and YOLOv9's 62.11 FPS. Typically, real-time object detection tasks require a model speed of 30 FPS or higher, and ST-YOLO far exceeds this standard, indicating that it can meet the application requirements of real-time detection scenarios. At the same time, the detection speed of over 90 FPS also shows that the model still has good operating efficiency in video stream processing, edge deployment, and continuous object detection tasks.

[0159] In terms of model size, ST-YOLO has 3.08 M parameters, which is higher than YOLO11 (2.58 M), YOLOv10 (2.69 M), and YOLOv9 (1.97 M), but lower than YOLOX (5.03 M) and RT-MDET (4.87 M). This indicates that ST-YOLO did not pursue extreme lightweighting in its model design, but rather enhanced the network's feature extraction and expression capabilities within a relatively small parameter increment range, enabling the model to achieve more stable detection results in complex scenarios. From an overall scale perspective, the 3.08 M parameter count still falls within the lightweight model category, indicating that the model has the potential to further improve detection performance while maintaining low storage and deployment costs.

[0160] Comprehensive analysis shows that ST-YOLO achieves a good balance between inference time, detection speed, and parameter count. Although its inference speed is slightly lower than YOLO11 and its parameter count is increased, this cost is relatively limited, and the model as a whole remains within an acceptable range for real-time detection and lightweight deployment. Therefore, ST-YOLO demonstrates good overall performance while balancing computational efficiency and model complexity, providing effective support for improving subsequent detection accuracy and possessing certain practical application value.

[0161] Table 7. Performance Comparison of Each Model .

[0162] Furthermore, in this embodiment, to verify the feasibility of ST-YOLO in practical applications, the optimized model was deployed on the NVIDIA Jetson AGX Orin development kit for testing. Edge devices typically face challenges of limited computing power and high inference speed requirements; therefore, optimization and deployment become key issues in practical applications. Based on this, this invention uses TensorRT to optimize ST-YOLO, improving inference speed and real-time performance while ensuring high accuracy on resource-constrained devices. During the optimization process, TensorRT significantly improves the inference efficiency of ST-YOLO through graph optimization, operation fusion, FP16 precision conversion, and automatic kernel tuning. Specifically, as shown in Table 8, the inference time of the optimized ST-YOLO model decreased from 54.4 ms to 22.8 ms, an improvement of 58.1%; the end-to-end processing time decreased from 58.8 ms to 28.6 ms, an improvement of 51.4%. The corresponding frame rate increased from 17 FPS to 35 FPS, an improvement of 106%, enabling the model to achieve real-time target detection on edge devices.

[0163] Table 8. Performance Comparison on Embedded Devices .

[0164] Despite the optimization, TensorRT had a very limited impact on model accuracy, with the model's mAP50 decreasing by only 0.3 percentage points and mAP50-95 decreasing by 0.4 percentage points. These minor changes in accuracy demonstrate that TensorRT optimization improves performance with almost no loss in the model's detection capabilities. Through this optimization, ST-YOLO can achieve real-time processing on resource-constrained embedded devices such as NVIDIA Jetson AGX Orin, fully meeting the high requirements for real-time performance and accuracy in space missions. The optimized ST-YOLO significantly improves inference speed while maintaining detection accuracy, providing strong support for rapid response in space missions.

[0165] Figure 9 A block diagram of a satellite target detection system based on multi-scale feature fusion enhancement according to the present invention is shown.

[0166] like Figure 9 As shown, this invention discloses a satellite target detection system based on multi-scale feature fusion enhancement, including a memory and a processor. The memory includes a satellite target detection method program based on multi-scale feature fusion enhancement. When the processor executes the satellite target detection method program based on multi-scale feature fusion enhancement, it performs the following steps:

[0167] A simulation dataset is constructed, which is derived from satellite targets of various types, attitudes, and complex lighting backgrounds.

[0168] An initial satellite target detection model is constructed, which includes a backbone network, a neck network, and a detection head. A multi-scale feature fusion module is introduced at the connection between the backbone network and the neck network. An efficient upsampling module is introduced in the upsampling path of the neck network. A multi-branch fusion sensing module is introduced in the feature aggregation area in front of the detection head to obtain an improved satellite target detection model.

[0169] The improved satellite target detection model is trained using the satellite target simulation dataset to obtain a trained satellite target detection model;

[0170] The satellite image to be detected is input into the trained satellite target detection model to obtain the satellite target detection result.

[0171] It should be noted that when the satellite target detection system based on multi-scale feature fusion enhancement disclosed in this invention is applied, the specific process corresponds to the satellite target detection method based on multi-scale feature fusion enhancement described in the above embodiments. Since the specific implementation details of the system application are consistent with the content of the satellite target detection method based on multi-scale feature fusion enhancement described above, no further details will be provided in this embodiment.

[0172] A third aspect of the present invention provides a computer-readable storage medium comprising a satellite target detection method program based on multi-scale feature fusion enhancement, wherein when the satellite target detection method program based on multi-scale feature fusion enhancement is executed by a processor, the program implements the steps of the satellite target detection method based on multi-scale feature fusion enhancement as described in any of the preceding claims.

[0173] The fourth aspect of the present invention provides a computer program product comprising: computer program code, which, when run on a computer, causes the computer to execute any of the methods described in the embodiments of the satellite target detection method based on multi-scale feature fusion enhancement.

[0174] A fifth aspect of the present invention provides an electronic device comprising: a processor and a memory; wherein the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the steps of a satellite target detection method based on multi-scale feature fusion enhancement as described in any of the preceding claims.

[0175] The terms “component,” “module,” “system,” etc., used in this specification are used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution. For example, a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program, and / or a computer. As illustrated, applications running on computing devices and computing devices can both be components. One or more components may reside in a process and / or an execution thread, and components may be located on a single computer and / or distributed among two or more computers. Furthermore, these components can be executed from various computer-readable media on which various data structures are stored. Components can communicate, for example, via local and / or remote processes based on signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system, and / or a network, such as the Internet interacting with other systems via signals).

[0176] Those skilled in the art will recognize that the various illustrative logical blocks and steps described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0177] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0178] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0179] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0180] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0181] In the above embodiments, the functions of each functional unit can be implemented entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. A computer program product includes one or more computer instructions (programs). When the computer program instructions (programs) are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs (DVDs), or semiconductor media (e.g., solid-state disks (SSDs)).

[0182] This invention discloses a satellite target detection method and system based on multi-scale feature fusion enhancement. It overcomes the problem of scarce space target data through an innovative high-fidelity data simulation process and directly applies the constructed simulation dataset to the training process of the improved model. By integrating a multi-scale feature fusion module, an efficient upsampling module, and a multi-branch fusion perception module in the model construction stage, it systematically solves the problems of existing models losing details of small targets, blurring upsampling edges, and having a single receptive field from the architectural level, forming a complete technical closed loop of "data construction - model improvement - model training".

[0183] Extensive experiments on self-built datasets and multiple public datasets demonstrate that this invention significantly outperforms existing mainstream models in core metrics such as precision, recall, and mAP, exhibiting particularly excellent robustness in complex backgrounds and low-light scenarios. Furthermore, while maintaining a lightweight design (only 3.08M parameters), this invention, accelerated and optimized with TensorRT, achieves real-time processing performance of 35 FPS on edge devices such as Jetson AGX Orin, successfully meeting the stringent requirements of high precision, high real-time performance, and low power consumption for space-based on-orbit missions.

[0184] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A satellite target detection method based on multi-scale feature fusion enhancement, characterized in that, Includes the following steps: A simulation dataset is constructed, which is derived from satellite targets of various types, attitudes, and complex lighting backgrounds. An initial satellite target detection model is constructed, which includes a backbone network, a neck network, and a detection head. A multi-scale feature fusion module is introduced at the connection between the backbone network and the neck network. An efficient upsampling module is introduced in the upsampling path of the neck network. A multi-branch fusion sensing module is introduced in the feature aggregation area in front of the detection head to obtain an improved satellite target detection model. The improved satellite target detection model is trained using the satellite target simulation dataset to obtain a trained satellite target detection model; The satellite image to be detected is input into the trained satellite target detection model to obtain the satellite target detection result.

2. The satellite target detection method based on multi-scale feature fusion enhancement according to claim 1, characterized in that, The construction of the simulation dataset specifically includes: Obtain three-dimensional models of various typical satellite structures, including at least cubic bus structures, cylindrical bus structures, and solar panel structures of different configurations. Based on the three-dimensional model, satellite foreground material is generated by multi-angle rendering using a virtual camera along a preset observation path. A multi-source background image library is constructed, and the satellite foreground material is fused with the background image to generate a synthetic satellite image; The satellite targets in the synthetic satellite image are labeled to generate training label files, so as to obtain the simulation dataset of satellite targets.

3. The satellite target detection method based on multi-scale feature fusion enhancement according to claim 2, characterized in that, The construction of a multi-source background image library and the fusion of the satellite foreground material with the background image specifically includes: Construct a multi-source background image library containing deep-space star images and Earth remote sensing images; During the fusion process, the brightness histogram features of the background image are analyzed, and the pixel intensity of the foreground satellite is automatically adjusted based on the illumination consistency adaptive correction mechanism to match the illumination conditions of the foreground and the background. Gaussian blur is applied at the boundary where the foreground and background blend to eliminate compositing artifacts.

4. The satellite target detection method based on multi-scale feature fusion enhancement according to claim 3, characterized in that, The multi-scale feature fusion module is used to aggregate information from multiple receptive fields within the same feature level through a hierarchical feature extraction chain of channel decoupling and cascaded deep convolutions. Specifically, it includes the following steps: The input feature tensor is adjusted to the hidden layer dimension through projection transformation, and channel decoupling operation is performed, dividing it into an identity mapping branch for maintaining gradient backpropagation and a processing branch for feature extraction. A cascaded depthwise convolution sequence is constructed for the processing branch. The depthwise convolutions with different kernel sizes are sequentially connected, and dense connections are used to achieve dense reuse of features in order to capture point-like, local and global features. The output of the identity mapping branch is concatenated with the outputs of all cascaded branches, and then aggregated through convolution to obtain multi-scale fused features.

5. The satellite target detection method based on multi-scale feature fusion enhancement according to claim 4, characterized in that, The high-efficiency upsampling module is used to perform depthwise convolutional smoothing and channel shuffling operations after feature upsampling through a reconstruction-interaction mechanism, specifically including the following steps: In the first stage, a physical upsampling operation is performed, and a depthwise convolution is immediately introduced to smooth the features of the upsampled feature map in order to eliminate the blocky effect and jagged edges introduced by interpolation. In the second stage, a channel shuffling operation is performed to reshape and transpose the feature tensor so that the features of different groups are evenly mixed in the channel dimension, thereby breaking the channel information blockage caused by depth convolution. In the third stage, cross-channel information is integrated through point convolution to output an enhanced feature map with clear edges.

6. The satellite target detection method based on multi-scale feature fusion enhancement according to claim 5, characterized in that, The multi-branch fusion sensing module is used to integrate densely connected branches and dilated convolutional branches with different dilation rates, specifically including the following steps: The input feature tensor is decoupled into two parts, which are then used as dense and hollow branches respectively. In dense branches, continuous convolutional layers and dense connection structures are used to enhance the reuse of high-frequency local detail features and gradient propagation. In the dilated branch, a set of dilated convolutions with different dilation rates are processed in parallel to exponentially expand the receptive field without reducing the resolution, so as to capture long-range contextual dependencies. By using residual connections, the input feature tensor, dense branch output, and dilated branch output are fused to obtain enhanced features that combine local details with global semantics.

7. A satellite target detection system based on multi-scale feature fusion enhancement, characterized in that, The system includes a memory and a processor. The memory contains a program for a satellite target detection method based on multi-scale feature fusion enhancement. When the processor executes the program, the satellite target detection method based on multi-scale feature fusion enhancement performs the following steps: A simulation dataset is constructed, which is derived from satellite targets of various types, attitudes, and complex lighting backgrounds. An initial satellite target detection model is constructed, which includes a backbone network, a neck network, and a detection head. A multi-scale feature fusion module is introduced at the connection between the backbone network and the neck network. An efficient upsampling module is introduced in the upsampling path of the neck network. A multi-branch fusion sensing module is introduced in the feature aggregation area in front of the detection head to obtain an improved satellite target detection model. The improved satellite target detection model is trained using the satellite target simulation dataset to obtain a trained satellite target detection model; The satellite image to be detected is input into the trained satellite target detection model to obtain the satellite target detection result.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a satellite target detection method program based on multi-scale feature fusion enhancement. When the satellite target detection method program based on multi-scale feature fusion enhancement is executed by a processor, it implements the steps of the satellite target detection method based on multi-scale feature fusion enhancement as described in any one of claims 1 to 6.

9. A computer program product, characterized in that, The computer program product includes computer program code, which, when run on a computer, causes the computer to implement the steps of a satellite target detection method based on multi-scale feature fusion enhancement as described in any one of claims 1 to 6.

10. An electronic device, characterized in that, The electronic device includes a processor and a memory; wherein the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the steps of a satellite target detection method based on multi-scale feature fusion enhancement as described in any one of claims 1 to 6.

Images