Large-model-driven remote sensing image small sample target detection method and system, medium and computer equipment
By employing a large model-driven approach, combined with multi-scale feature fusion and anisotropic structure perception, the challenge of detecting small targets and slender structures in remote sensing images has been solved. This approach improves detection accuracy and adaptability to complex backgrounds, making it suitable for environmental monitoring and resource surveys.
Patent Information
- Application Number
- CN202511852781.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-10
- Publication Date
- 2026-02-10
AI Technical Summary
Existing methods for detecting small targets in remote sensing images struggle to accurately detect tiny targets and slender structures when faced with complex backgrounds and multi-scale variations. Traditional methods rely on handcrafted features and are time-consuming, while deep learning methods lack sufficient feature extraction capabilities in high-resolution remote sensing images, leading to a decline in detection performance.
A large model-driven approach is adopted, which enhances the model’s sensitivity to small targets, multi-scale targets and slender structures in remote sensing images through a scale-adaptive feature enhancement module, a visual state space multi-scale feature fusion module and an anisotropic structure perception module. The pre-trained model of the Transformer architecture and the multi-granularity self-attention mechanism are combined with large kernel convolution to construct anisotropic receptive fields and enhance the target detection capability.
It improves the accuracy and generalization ability of target detection in complex backgrounds, significantly enhances the positioning accuracy of small targets, effectively filters background noise interference, and enhances the perception ability of slender targets, making it suitable for remote sensing applications such as environmental monitoring and resource surveys.
Smart Images
Figure CN121505467A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to a large model driven remote sensing image small sample target detection method, system, medium and computer equipment, belonging to the field of image processing and computer vision. BACKGROUND
[0002] Remote sensing images usually contain multiple target objects and complex backgrounds, where the target objects may have different shapes, sizes, colors and texture characteristics, and the background may include vegetation, buildings, sandy land and other noise or interference areas. These factors make it particularly difficult to detect and identify small sample targets in high-resolution remote sensing images. Remote sensing small sample target detection method is a computer vision task aimed at complex background, scale variation and other challenges in remote sensing images, which aims to accurately locate and identify target objects in the captured image.
[0003] Although existing deep learning methods have made significant progress in handling conventional target detection tasks, in remote sensing images, as the application scenarios continue to expand, target scale variation and complex background still pose increasingly greater challenges to traditional target detection models. Currently, the commonly used target detection methods mainly include traditional method based target detection method and deep learning based target detection method. The traditional method based target detection method usually relies on manually extracted features such as color, texture, boundary and brightness. This method often shows insufficient adaptability when facing complex scenes, and the detection performance may be greatly affected. In addition, the manual feature extraction process is time-consuming and requires professional knowledge, which is a significant limitation in practical applications. On the other hand, the deep learning based target detection method automatically extracts image features through neural networks and realizes target detection through end-to-end training. Although this method greatly improves the adaptability to complex scenes and reduces human intervention, due to the weak image spectral information of high-resolution remote sensing images, the problem of multi-scale target variation and the interference problem of small changes in complex spatial structure features, the feature extraction ability and target detection performance of existing deep learning target detection models inevitably decline, limiting the application of existing technology in remote sensing images. SUMMARY
[0004] The purpose of the present application is to overcome the shortcomings in the prior art, and to provide a large model driven multi-scale feature fusion remote sensing image small sample target detection method and device. Through the scale adaptive feature enhancement module, the visual state space multi-scale feature fusion module and the anisotropic structure perception module, the sensitivity of the model to small targets, multi-scale targets and slender structures in remote sensing images is enhanced, the useful geometric structure information is highlighted, the problems of small sample condition micro target feature loss and slender target geometric perception deficiency are effectively solved, and the target detection accuracy in complex background is improved.
[0005] To achieve the above objectives, the present invention is implemented using the following technical solution:
[0006] In a first aspect, this invention proposes a large-model-driven method for detecting small-sample targets in remote sensing images, comprising:
[0007] Acquire remote sensing images containing target objects and complex backgrounds;
[0008] The remote sensing image is input into a pre-trained large model based on the Transformer architecture to obtain enhanced features. The pre-trained large model based on the Transformer architecture includes a backbone network and a feature enhancer. The backbone network is used to extract and output features, and the feature enhancer includes a scale-adaptive structure-aware feature enhancement module, which is used to perform structure-aware and scale-adaptive enhancement on the preset multi-scale image features output by the backbone network to obtain the enhanced features. The enhanced features are then input into a visual state space multi-scale feature fusion module to obtain fused features. The visual state space multi-scale feature fusion module includes a strip pooling mechanism, a lightweight bidirectional feature pyramid, and visual state space blocks, which are used to capture long-distance dependencies, achieve multi-scale information complementarity, and global context refinement.
[0009] The fused features are input into the anisotropic structure perception module, and anisotropic receptive fields are constructed using directional separable large kernel convolutions to enhance the model's overall perception ability of structured targets such as roads, rivers, and bridges, and output structure-enhanced features.
[0010] The structure-enhanced features are input into the decoding structure for final detection and regression, and the target detection results are output, including bounding boxes and corresponding text phrases.
[0011] Furthermore, the pre-trained large model based on the Transformer architecture is a pre-trained GroundingDINO model, which is used as a basic detector to extract multi-scale visual features of the remote sensing image.
[0012] Further, the backbone network includes an image backbone network and a text backbone network. The image backbone network extracts multi-scale image features of the remote sensing image based on the Swin-T architecture, and the text backbone network extracts text features of the corresponding text labels of the remote sensing image based on the BERT architecture. The multi-scale image features, after being processed by the feature enhancer and the visual state space multi-scale feature fusion module, serve as the visual input to the cross-modal decoding structure and are used together with the text features for target detection. Further, the feature enhancer processes the output features of the backbone network by embedding a structure-aware feature enhancement module after each layer of image self-attention to obtain the enhanced features, including:
[0013] The output features of the backbone network are processed based on a multi-granularity self-attention mechanism to combat extreme scale changes in the image and capture the features of small objects.
[0014] Hollow local attention extracts context-enhanced local features while preserving the features of small objects, thereby further expanding the receptive field;
[0015] A spatial attention map is generated based on the local features, and attention weights are calculated.
[0016] The attention weights are applied to the output features of the backbone network to obtain the enhanced features, thus solving the scale mismatch problem of remote sensing targets under high-altitude views.
[0017] Furthermore, the step of inputting enhanced features into the visual state space multi-scale feature fusion module to obtain fused features includes:
[0018] According to the strip pooling mechanism, the enhanced features are subjected to strip-shaped average pooling along two orthogonal directions, horizontal and vertical, to aggregate the global context and establish long-distance dependencies, and to provide inductive bias for slender targets.
[0019] The feature inputs, which are subjected to elongated average pooling, are quantized into a lightweight bidirectional feature pyramid structure. A weighted fusion strategy is used to perform multi-level information complementation on feature layers of different scales from bottom to top and from top to bottom.
[0020] The weighted fusion features are input into the visual state space block, and the global features are refined using the state space model to obtain the fused features.
[0021] Furthermore, the visual state space block uses two-dimensional selective scanning as the core operator to replace part of the convolution to achieve global modeling, including:
[0022] The weighted fused feature map is expanded into a sequence set along four directions, and a selective state-space equation is applied.
[0023] Based on the sequence set as input, parameters are dynamically generated, and the forget gate and input gate are adaptively adjusted to filter out large areas of background redundancy and focus on salient targets.
[0024] By recursively updating the state and merging sequences, feature extraction of the global receptive field is achieved while maintaining linear complexity.
[0025] Furthermore, the step of inputting the fused features into the anisotropic structure sensing module includes:
[0026] Based on the large kernel decomposition theory, the fused features are processed using depthwise convolution branches, which are decomposed from a large-size convolution kernel and configured along the horizontal and vertical directions respectively.
[0027] Long-distance linear structural features along the target extension direction are captured by horizontal branching;
[0028] Capture local geometric changes in the short side of the target by using vertical branches;
[0029] The output features of the two branches are fused, and a structure-sensitive spatial attention map is generated through 1×1 convolution and activation function.
[0030] The attention map and the fused features are subjected to a Hadamard product operation to output the structure-enhanced features.
[0031] Secondly, this invention proposes a large-model-driven remote sensing image small-sample target detection system for implementing a large-model-driven remote sensing image small-sample target detection method. The system includes:
[0032] The data acquisition module is used to acquire remote sensing images and text prompts; the remote sensing images include slender structures, small target objects, and complex background scenes.
[0033] The feature extraction module is used to input the remote sensing image into a pre-trained Grounding DINO model, extract the image features of the remote sensing image using the Swin-T backbone network, and extract the text features of the text labels corresponding to the remote sensing image using the BERT backbone network.
[0034] The structure-aware optimization module is used to generate structure-sensitive attention weights by embedding the feature enhancer of the structure-aware feature enhancement module, thereby enhancing the perception of small and slender targets.
[0035] The multi-scale fusion module is used to reconstruct features through the visual state space multi-scale feature fusion module. It combines strip pooling, lightweight bidirectional feature pyramid and visual state space blocks to achieve global context refinement and multi-scale complementarity. The state space model helps to control computational overhead and focus on salient regions while maintaining global modeling capabilities.
[0036] An anisotropic perception module is used to construct an anisotropic receptive field through the anisotropic structure perception module, generate a structure-sensitive attention map to enhance the perception of slender targets, and output structure-enhanced features.
[0037] The output module is used to output target detection results based on the structure enhancement features. The target detection results include bounding box prediction and class label, supporting high-precision detection under small sample conditions.
[0038] Thirdly, the present invention proposes a computer-readable storage medium storing a computer program thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of a large model-driven remote sensing image small sample target detection method.
[0039] Fourthly, the present invention provides a computing device, comprising,
[0040] Memory, used to store computer programs;
[0041] A processor for executing the computer program to implement the steps of a large model-driven method for small sample target detection in remote sensing images.
[0042] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:
[0043] (1) This invention provides a large model-driven method for detecting small targets in remote sensing images. It designs a scale-adaptive feature enhancement module, a visual state space multi-scale feature fusion module, and an anisotropic structure perception module. It effectively solves the challenges of small targets being easily lost, slender targets being difficult to detect, and large scale differences under small sample conditions in remote sensing scenarios, and improves the generalization ability and practicality of the method in the interpretation of complex ground objects.
[0044] (2) This invention optimizes the key multi-scale features output by the backbone network by embedding a scale-adaptive structure-aware feature enhancement module in the feature enhancer. This module enhances the model’s ability to perceive small targets in remote sensing images by using multi-granularity attention and hollow local attention mechanisms, generates scale-sensitive attention weights, and significantly improves the positioning accuracy of small targets under high-altitude view.
[0045] (3) The present invention uses a visual state space multi-scale feature fusion module to strengthen the global dependency of multi-scale features. By using a strip pooling mechanism and introducing a 2D selective scanning state space model, global context refinement is achieved while maintaining near-linear computational complexity. This effectively filters out interference from redundant background information in large areas and improves the distinguishability of features in complex backgrounds.
[0046] (4) This invention introduces an anisotropic structure perception module, which uses the big kernel decomposition theory to construct an anisotropic receptive field that conforms to the geometric properties of slender targets, and generates structure-sensitive attention weights, which effectively enhances the model's perception ability of slender targets such as roads and bridges, and alleviates the problem of mismatch between the shape of the receptive field and the geometric structure of the target when the traditional isotropic convolution processes anisotropic targets.
[0047] (5) The method of the present invention provides a large model-driven remote sensing image small sample target detection method, which has important engineering promotion value in remote sensing application fields such as environmental monitoring and resource survey. Attached Figure Description
[0048] Figure 1 This is a flowchart of a large model-driven remote sensing image small sample target detection method provided by an embodiment of the present invention;
[0049] Figure 2 This is a diagram of the lightweight bidirectional feature pyramid structure in an embodiment of the present invention;
[0050] Figure 3 This is a structural diagram of the visual state space module in an embodiment of the present invention. Detailed Implementation
[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.
[0052] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.
[0053] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.
[0054] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.
[0055] It should be emphasized here that the step markers mentioned below are not a limitation on the order of the steps, but should be understood as meaning that the steps can be executed in the order mentioned in the embodiments, or in a different order than in the embodiments, or several steps can be executed simultaneously.
[0056] Example 1
[0057] One embodiment of the present invention provides a large-model-driven, multi-scale feature fusion method for small-sample target detection in remote sensing images, such as... Figure 1 As shown, the specific steps include:
[0058] The process involves acquiring remote sensing images containing target objects and complex backgrounds. Target objects are specific objects in the remote sensing scene that need to be detected and identified, such as aircraft, ships, and vehicles. These targets typically have different shapes, sizes, colors, and textures, and may exhibit low recognizability under complex backgrounds and scale variations. Background images encompass non-target areas in the remote sensing environment, including vegetation, buildings, sand dunes, water bodies, and areas affected by lighting or occlusion. These background images often contain noise and complex texture information, posing a challenge to target detection.
[0059] A pre-trained Grounding DINO model is used as the base detector to extract multi-scale visual features from remote sensing images. The Grounding DINO model includes a backbone network and a feature enhancer. The backbone network extracts and outputs features, comprising an image backbone network and a text backbone network. The image backbone network extracts multi-scale image features from remote sensing images based on the Swin-T architecture, while the text backbone network extracts text features corresponding to text labels in remote sensing images based on the BERT architecture. The multi-scale image features and text features are used as the output features of the backbone network and input into the feature enhancer for preliminary view-text feature fusion. The input multi-scale feature set is represented as follows:
[0060] (1)
[0061] in This represents features from different downsampling levels; This is a newly added high-resolution feature stream used to preserve rich spatial details of small targets.
[0062] The output features of the backbone network are fed into the feature enhancer, and a scale-adaptive feature enhancement module is embedded after the self-attention of each layer of the image to alleviate the feature weakening problem of small targets in high-altitude overhead scenes.
[0063] Employing a dilated local attention mechanism: for any feature layer Extracting context-enhanced local features using dilated convolution. ,in H and W represent the number of channels, the height of the feature map, and the width of the feature map, respectively. Indicates the expansion rate Depth-separable convolutions. By adjusting... It can expand the receptive field without reducing the resolution;
[0064] Attention weight generation and recalibration: based on local features Generate spatial attention map And perform element-wise reweighting on the input. Obtain attention weights ,in It is the product of Hadamard.
[0065] Perform residual output: Apply attention weights to the output features of the backbone network to obtain enhanced features; the enhanced output is... ,in The response to small and medium-sized targets is significantly enhanced.
[0066] The features output from the structure-aware feature enhancement module are fed into the multi-scale state space feature fusion module, and then sequentially pass through strip pooling, a lightweight bidirectional feature pyramid, and a visual state space block to achieve a refined representation with global context. The specific steps are as follows:
[0067] Based on the strip pooling mechanism, the features output by the structure-aware feature enhancement module are... Perform strip-shaped average pooling along two orthogonal directions (horizontal and vertical) to obtain horizontal strip pooling output. and vertical strip pooling output By aggregating global context and establishing long-range dependencies, and providing inductive biases for elongated targets, the fused features are obtained. ,in, The output is a horizontal strip pooling output with dimension . , The output is a vertical strip pooling output with dimension . ; These represent the channel, row index, and column index, respectively. Expanding the expression, Stretch to Same space dimensions. This represents the fused features after strip pooling is integrated into the global context.
[0068] The feature input, after undergoing elongated average pooling, is quantized into a lightweight bidirectional feature pyramid structure. A weighted fusion strategy is then used to perform multi-level information complementation on feature layers of different scales, both bottom-up and top-down. Figure 2 As shown:
[0069] Adopting a three-layer structure Multi-scale compensation is achieved through weighted fusion; for example, the third layer of fusion is...
[0070] (2)
[0071] in These represent feature pyramid layers at different scales; This is the input feature for the 3rd level; This refers to the intermediate fusion features in the bottom-up path of level 3; This represents the final output feature in the top-down path of level 3; This indicates a downsampling operation used to align feature resolution; This indicates weighted feature fusion.
[0072] The features to be weighted and fused are input into the visual state space block, such as Figure 3 As shown, preprocessing is first performed through linear layers, activation functions, and layer normalization. Then, a 2D selective scanning module executes the state space equation across the entire domain to capture long-range dependencies and refine features. Subsequently, local details are further fused through linear layers, activation functions, depthwise separable convolutions, linear layers, and layer normalization, outputting the final fused features. The visual state space block utilizes 2D selective scanning instead of traditional convolution, including:
[0073] The feature maps that undergo weighted fusion are expanded into a sequence set along four directions, and a selective state-space equation is applied.
[0074] Based on the serialized input features at time step vector The input parameters are dynamically generated, and the forget gate and input gate are adaptively adjusted. The dynamic parameters are generated by the following formula:
[0075] (3)
[0076] in, The step size parameter is dynamically generated from the input and is used to adjust the scale of memory and forgetting. For input Dynamically generated control parameters enable adaptive content modeling. For activation functions; It is a linear projection layer.
[0077] The discretized state is updated as follows
[0078] (4)
[0079] in, For time step The hidden state is used to store historical context; These are the state-space parameters after zero-order preserved discretization; This refers to the output features after two-dimensional selective scanning.
[0080] By recursively updating the state and merging sequences, feature extraction of the global receptive field is achieved while maintaining linear complexity.
[0081] The features output from the multi-scale state space are input into the anisotropic structure perception (ASAM) module to perform geometric enhancement on slender structures such as bridges and roads, including:
[0082] Based on the theory of large kernel decomposition, large kernel convolution is... The algorithm is decomposed into horizontal and vertical depthwise convolution sequences. The horizontal branch captures long-range linear structural features along the target's extension direction, while the vertical branch captures local geometric changes along the target's short side, resulting in the output features. Output features The calculation formula is:
[0083] (5)
[0084] in, For input, it refers to the features of the multi-scale state space output. A horizontal depthwise convolution sequence, containing and convolution kernel, It is a vertical depthwise convolution sequence, whose convolution kernel is symmetrical to the horizontal convolution kernel.
[0085] The output features of the two branches are fused to generate a structure-sensitive spatial attention map. The formula for calculating the structure-sensitive attention map is as follows:
[0086] (6)
[0087] in, The Sigmoid function is used to... The output response is compressed to between 0 and 1, forming a gating weight.
[0088] The attention map and the fused features are then subjected to a Hadamard product operation to output the structure-enhanced features. The calculation formula is as follows:
[0089] (7)
[0090] in, This represents structural enhancement features, the enhanced version. It can form a continuous high response band along the direction of slender targets and suppress background noise.
[0091] Enhance the final structural features The input is fed into the Cross-Modality Decoder, which generates the final detection results, including predicted bounding boxes and class labels, through multi-layer attention interactions.
[0092] Example 2
[0093] Based on Embodiment 1, this embodiment proposes a large-model-driven remote sensing image small-sample target detection system, including:
[0094] The data acquisition module is used to acquire remote sensing images;
[0095] The feature extraction module is used to extract image features of remote sensing images using the Swin-T backbone network and to extract text features of text labels corresponding to the remote sensing images using the BERT backbone network.
[0096] The structure-aware optimization module is used to generate structure-sensitive attention weights by embedding the feature enhancer of the structure-aware feature enhancement module, thereby enhancing the perception of small and slender targets.
[0097] The multi-scale fusion module is used to reconstruct features through the visual state space multi-scale feature fusion module, and combines strip pooling, bidirectional feature pyramid and visual state space blocks to achieve global context refinement and multi-scale complementarity.
[0098] An anisotropic perception module is used to construct an anisotropic receptive field through the anisotropic structure perception module, generate a structure-sensitive attention map to enhance the perception of slender targets, and output structure-enhanced features.
[0099] The output module is used to output target detection results based on structural enhancement features.
[0100] Example 3
[0101] Based on Embodiment 1, this embodiment proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements a large-model-driven method for detecting small-sample targets in remote sensing images.
[0102] Example 4
[0103] This embodiment, based on Embodiment 1, proposes a computing device, including,
[0104] Memory, used to store computer programs;
[0105] A processor for executing the computer program to implement the steps of a large model-driven method for small sample target detection in remote sensing images.
[0106] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0107] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0108] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0109] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A large-model-driven method for detecting small-sample targets in remote sensing images, characterized in that, include: Acquire remote sensing images containing target objects and complex backgrounds; The remote sensing image is input into a pre-trained large model based on the Transformer architecture to obtain enhanced features. The pre-trained large model based on the Transformer architecture includes a backbone network and a feature enhancer. The backbone network is used to extract and output features. The feature enhancer includes a scale-adaptive structure-aware feature enhancement module, which is used to perform structure-aware and scale-adaptive enhancement on the preset multi-scale image features output by the backbone network to obtain the enhanced features. The enhanced features are input into the visual state space multi-scale feature fusion module to obtain the fused features. The visual state space multi-scale feature fusion module includes a strip pooling mechanism, a lightweight bidirectional feature pyramid, and a visual state space block. The fused features are input into the anisotropic structure perception module, and anisotropic receptive fields are constructed using orientation-separable large kernel convolutions to output structure-enhanced features. The structure-enhanced features are input into the decoding structure for final detection and regression, and the target detection results are output, including bounding boxes and corresponding text phrases.
2. The method for small-sample target detection in remote sensing images driven by a large model according to claim 1, characterized in that, The pre-trained large model based on the Transformer architecture is a pre-trained Grounding DINO model. The pre-trained Grounding DINO model is used as a basic detector to extract multi-scale visual features of the remote sensing image.
3. The method for small-sample target detection in remote sensing images driven by a large model according to claim 1, characterized in that, The backbone network includes an image backbone network and a text backbone network. The image backbone network extracts multi-scale image features of the remote sensing image based on the Swin-T architecture, and the text backbone network extracts text features of the corresponding text labels of the remote sensing image based on the BERT architecture. The multi-scale image features are processed by the feature enhancer and the visual state space multi-scale feature fusion module, and then used as the visual input of the cross-modal decoding structure, together with the text features, for target detection.
4. The method for small-sample target detection in remote sensing images driven by a large model according to claim 1, characterized in that, The feature enhancer, based on a scale-adaptive structure-aware feature enhancement module, processes the output features of the backbone network to obtain the enhanced features, including: The output features of the backbone network are processed based on a multi-granularity self-attention mechanism to combat extreme scale changes in the image and capture the features of small objects. Hollow local attention extracts context-enhanced local features while preserving the features of small objects, thus further expanding the receptive field; A spatial attention map is generated based on the local features, and attention weights are calculated. The attention weights are applied to the output features of the backbone network to obtain the enhanced features.
5. The large-model-driven remote sensing image small-sample target detection method according to claim 1, characterized in that, The step of inputting enhanced features into the visual state space multi-scale feature fusion module to obtain fused features includes: According to the strip pooling mechanism, the enhanced features are subjected to strip-shaped average pooling along two orthogonal directions, horizontal and vertical, to aggregate the global context and establish long-distance dependencies, and to provide inductive bias for slender targets. The feature inputs, which undergo elongated average pooling, are quantized into a lightweight bidirectional feature pyramid structure. A weighted fusion strategy is used to perform multi-level information complementation on feature layers of different scales, both bottom-up and top-down. The weighted fusion features are input into the visual state space block, and the global features are refined using the state space model to obtain the fused features.
6. The large-model-driven remote sensing image small-sample target detection method according to claim 5, characterized in that, The visual state space block uses two-dimensional selective scanning as its core operator to replace part of the convolution to achieve global modeling, including: The weighted fused feature map is expanded into a sequence set along four directions, and a selective state-space equation is applied. Based on the sequence set as input, parameters are dynamically generated to adaptively adjust the forget gate and the input gate; By recursively updating the state and merging sequences, feature extraction of the global receptive field is achieved while maintaining linear complexity.
7. The method for small-sample target detection in remote sensing images driven by a large model according to claim 1, characterized in that, The step of inputting the fused features into the anisotropic structure sensing module includes: Based on the large kernel decomposition theory, the fused features are processed using depthwise convolution branches, which are decomposed from a large-size convolution kernel and configured along the horizontal and vertical directions respectively. Long-distance linear structural features along the target extension direction are captured by horizontal branching; Capture local geometric changes in the short side of the target by using vertical branches; The output features of the two branches are fused to generate a structure-sensitive spatial attention map; The attention map and the fused features are subjected to a Hadamard product operation to output the structure-enhanced features.
8. A large-model-driven remote sensing image small-sample target detection system, characterized in that, The system is used to implement the large model-driven remote sensing image small sample target detection method according to any one of claims 1 to 7, the system comprising: The data acquisition module is used to acquire remote sensing images; The feature extraction module is used to extract image features of remote sensing images using the Swin-T backbone network and to extract text features of text labels corresponding to the remote sensing images using the BERT backbone network. The structure-aware optimization module is used to generate structure-sensitive attention weights through the scale-adaptive structure-aware feature enhancement module, thereby enhancing the perception of small and slender targets. The multi-scale fusion module is used to reconstruct features through the visual state space multi-scale feature fusion module, and combines strip pooling, bidirectional feature pyramid and visual state space blocks to achieve global context refinement and multi-scale complementarity. An anisotropic perception module is used to construct an anisotropic receptive field through the anisotropic structure perception module, generate a structure-sensitive attention map to enhance the perception of slender targets, and output structure-enhanced features. The output module is used to output the target detection result based on the structural enhancement features.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the large model-driven remote sensing image small sample target detection method as described in claims 1 to 7.
10. A computing device, characterized in that, include, Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the large model-driven remote sensing image small sample target detection method according to any one of claims 1 to 7.