Remote sensing lightweight target detection method based on local state modeling network

Abstract

The invention provides a remote sensing lightweight target detection method based on a local state modeling network, which is called as LSM Net. The method comprises the following steps: preprocessing a remote sensing image data set; then constructing an LSM Net which takes the LSM-TD as a backbone network and takes the feature pyramid as a neck network, and taking the preprocessed image as network input; in the training process, a lightweight local state modeling module (LSM Block) is combined to optimize the network, and fine feature extraction and long-range spatial dependence modeling of a tiny target are realized. And finally, inputting a to-be-detected remote sensing image into the trained LSM Net network to obtain a detection frame containing a tiny target category and the confidence coefficient. According to the method, through lightweight network design and multi-dimensional feature optimization, the complex scene adaptability and operation efficiency are remarkably improved while the detection precision is guaranteed, the method is suitable for intelligent monitoring and rapid identification of multi-field tiny targets in remote sensing images, and an efficient and feasible technical scheme is provided for resource-limited edge equipment.

Description

Lightweight Target Detection Method for Remote Sensing Based on Local State Modeling Network Technical Field

[0001] This invention belongs to the field of target detection technology, specifically relating to a lightweight remote sensing target detection method based on local state modeling networks. Background Technology

[0002] With the rapid development of UAVs and remote sensing technologies, the automatic and accurate detection of tiny targets such as vehicles, ships, and oil tanks from high-resolution aerial images has significant application value in fields such as urban planning, traffic monitoring, and environmental perception. However, targets in aerial images are usually extremely small in scale (occupying only tens of pixels in the image), have sparse texture information, and have complex backgrounds, resulting in weak features and poor discriminability.

[0003] Existing high-performance detection methods typically employ Transformer architectures or large-kernel convolutions to expand the receptive field and model long-range dependencies, but this leads to a sharp increase in the number of parameters and computational costs, making them difficult to deploy on computationally limited aerospace edge devices (such as UAV onboard platforms). On the other hand, existing lightweight network designs (such as methods based on depthwise separable convolutions or structure reparameterization) can reduce computational costs, but their feature extraction capabilities are limited, making it difficult to simultaneously capture the fine-grained structural details and long-range contextual dependencies required for small targets, resulting in a significant decrease in detection performance.

[0004] Recently, the RepMamba model has demonstrated efficient long-range modeling capabilities in image super-resolution tasks by combining a state-space model with reparameterized convolutions. However, its architecture and computational resource allocation strategy, designed for reconstruction tasks, are not directly applicable to small object detection tasks, resulting in issues such as blurred feature hierarchy and weakened spatial details. Direct transfer of these features leads to poor detection performance.

[0005] Therefore, a lightweight network architecture is needed that can effectively model local details and long-range spatial dependencies while maintaining extremely low computational overhead, thereby achieving high-precision airborne small target detection in resource-constrained environments. Summary of the Invention

[0006] To address the shortcomings of the aforementioned background technologies, this invention proposes a lightweight remote sensing target detection method based on a local state modeling network. While maintaining extremely low computational overhead and parameter quantity, it significantly improves the detection accuracy of small targets through innovative network structure design.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A lightweight remote sensing target detection method based on local state modeling networks includes the following steps:

[0009] S1: Obtain the remote sensing image dataset and perform preprocessing.

[0010] S2: Construct an LSM Net network, including a backbone network and a neck network; the backbone network is LSM-TD, and the neck network is a feature pyramid.

[0011] S3: Using the preprocessed remote sensing image as input to the LSM Net network, the LSM Net network is trained based on local state modeling.

[0012] S4: Input the remote sensing image to be detected into the trained LSM Net network. After multi-stage feature extraction, local-global joint modeling and detection head prediction, the output includes the detection results containing the small target category, location coordinates and corresponding confidence.

[0013] Furthermore, the construction of the LSM Net network includes the following steps:

[0014] S1: Constructing the LSM-TD backbone network: Design a hierarchical four-stage network architecture with the output channel ratio of the four stages being 5:5:2:1. The first and second stages are high-resolution stages used for densely stacking feature enhancement modules to preserve spatial details; the third and fourth stages are low-resolution stages used for feature extraction.

[0015] S2: Construct the feature pyramid neck network: adopt FPN as the neck network, receive the output of the LSM-TD backbone network at different stages, and generate multi-scale feature maps.

[0016] S3: Insert a Local State Modeling Module (LSMBlock) into the backbone of the network to enhance the features.

[0017] S4: The input end of the target detection head is connected to the output end of the feature pyramid neck network, and is used to predict the target category and bounding box based on the enhanced multi-scale features.

[0018] Furthermore, the Local State Modeling Module (LSMBlock) as a whole comprises the following modules:

[0019] The local feature extraction unit is used to extract local neighborhood information of the input features through depthwise separable convolution.

[0020] A two-dimensional sequence modeling unit is used to perform long-range dependency modeling on the local neighborhood information along the spatial height and width directions, respectively.

[0021] The channel mixing unit is used to perform inter-channel interaction and integration of features that incorporate long-range dependency information.

[0022] The feature fusion unit is used to fuse the output of the channel mixing unit with the original input features through residual connections to obtain enhanced output features.

[0023] Furthermore, the overall implementation method of the Local State Modeling Module (LSMBlock) includes the following steps:

[0024] Input feature map ,in For batch size, For the number of channels, and These refer to the space's height and width, respectively.

[0025] S1: Local feature extraction:

[0026] For input features implement Depthwise separable convolution operations are used to extract local neighborhood information and obtain preliminary local features. Its operation can be expressed as:

[0027] ;

[0028] To stabilize training and enhance nonlinear representation capabilities, batch normalization (BN) and an activation function (such as GELU or SiLU) are typically added after convolution. Therefore, complete local feature extraction can be represented as:

[0029] ;

[0030] in, This represents the activation function.

[0031] S2: Lightweight 2D State-Space Modeling (L-2DSSM):

[0032] To capture long-range spatial dependencies, a lightweight two-dimensional state-space modeling method is introduced. This method avoids unfolding feature maps into one-dimensional sequences and instead directly simulates state propagation in two-dimensional space.

[0033] The local features obtained in the previous step Input a lightweight two-dimensional state-space model (L-2DSSM) and perform sequence modeling along the height and width dimensions to capture long-range dependency features. The mathematical definition of this process is:

[0034]

[0035] This formula clarifies that the function of the L-2DSSM operation is to transform local features. Mapped to features implying long-range spatial dependence .

[0036] L-2DSSM is implemented through row- and column-separated depthwise convolutions:

[0037] use The depthwise convolutional kernel performs convolutions along the height direction, capturing long-range inter-line dependencies:

[0038]

[0039] use The depthwise convolutional kernel performs convolution in the width direction, capturing long-range inter-column dependencies:

[0040]

[0041] Will The receptive fields are decomposed into sequence. and Convolution can expand the effective receptive field at minimal cost, while maintaining channel independence and fine-grained spatial structure that is crucial for the localization of small targets.

[0042] Merge information from two directions:

[0043]

[0044] Finally, use groups of 8. Grouped convolutions are used to perform channel fusion, resulting in a state-space modeled output:

[0045]

[0046] S3: Channel Mixing (MLP):

[0047] After capturing spatial dependencies, a lightweight multilayer perceptron (referred to as RepVitMLP in the paper) is used for channel mixing to enhance feature representation. Its operation is as follows:

[0048]

[0049] The first pointwise convolutional layer reduces the channel dimension from... Down to The second layer of grouped convolutions reduces the channel dimension from Restore to .

[0050] S4: Feature Fusion and Output

[0051] By using residual connections, the original input of the module is... Features after the above processing The features are added together. Finally, through batch normalization (BN) integration, the final output features of LSMBlock are obtained. :

[0052]

[0053] Furthermore, the method for training the LSM Net network based on local state modeling includes the following steps:

[0054] S1: Network Initialization: Initialize the weight parameters of the LSM Net network.

[0055] S2: Forward propagation: A batch of preprocessed training images are input into the network and processed sequentially through the LSM-TD backbone network, the local state modeling module, the feature pyramid neck network, and the target detection head to obtain the prediction results.

[0056] S3: Loss Calculation: Calculate the classification loss and localization loss based on the prediction results and the true labels of the image annotations.

[0057] S4: Backpropagation and parameter update: Using an optimization algorithm, calculate the gradient based on the loss and update all trainable parameters of the network.

[0058] S5: Iterative optimization: Repeat steps S2 to S4 until the preset training rounds or performance indicators are reached.

[0059] Furthermore, the optimization algorithm used for backpropagation and parameter update is AdamW, and a learning rate preheating and step decay strategy is adopted.

[0060] The beneficial effects of this invention are:

[0061] This invention, through core modules and appropriate resource allocation, significantly reduces model complexity while effectively enhancing feature representation capabilities, achieving the best balance between computational efficiency and detection performance, and providing an efficient and lightweight detection solution for edge computing scenarios such as spaceborne and airborne applications.

[0062] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 is a flowchart illustrating a lightweight remote sensing target detection method based on a local state modeling network according to an embodiment of the present invention.

[0065] Figure 2 is a schematic diagram of the results of a lightweight remote sensing target detection method based on a local state modeling network according to an embodiment of the present invention. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] As shown in Figure 1, this embodiment of the invention provides a lightweight remote sensing target detection method based on a local state modeling network. The specific steps are as follows:

[0068] S1: Obtain the remote sensing image dataset and perform preprocessing; specifically, the dataset used in this embodiment is the AI-TODv2 dataset. AI-TODv2 is a benchmark dataset specifically for the detection of small targets in aerial images. It focuses on targets with extremely small pixel sizes in aerial photography scenes and is used to train and evaluate the detection capabilities of algorithms in such small targets and complex background scenes, supporting the technological development in fields such as UAV inspection, satellite image analysis, and industrial micro-object recognition.

[0069] The preprocessing workflow for this dataset is as follows: First, the images and annotations are loaded. Then, the images are uniformly scaled to 800x800 pixels while maintaining the original aspect ratio to prevent distortion. Next, data augmentation is performed by random horizontal flipping. Then, the images are normalized using the mean and standard deviation of ImageNet. Finally, the images are padded to integer multiples of 32 and converted to the tensor format required for model training. This series of operations aims to improve the robustness of the model in detecting small targets in remote sensing images.

[0070] S2: Construct an LSMNet network with LSM-TD as the backbone network and a feature pyramid as the neck network, and pre-train the LSMNet network.

[0071] The lightweight detection task of small targets in remote sensing images adopts a detection framework based on the FCOS anchorless framework, with LSM-TD as the dedicated backbone network. Compared with general state space models (such as RepMamba), LSM Net enhances the feature capture and localization accuracy of small targets while reducing the number of parameters and computational cost, making it suitable for resource-constrained scenarios such as real-time inspection by UAVs and rapid parsing by satellite terminals.

[0072] As a lightweight backbone designed specifically for remote sensing of small targets, LSM-TD uses LSM blocks and downsampling blocks as basic units. It achieves accurate feature extraction by reconstructing the allocation of computational resources in each stage: the number of LSM blocks in the first two high-resolution feature map stages is set to 5, and the number of LSM blocks in the last two stages is set to 2 and 1, which prioritizes the preservation of fine-grained spatial information at the bottom layer and meets the detection needs of small targets; the channel dimension adopts a lightweight design with 96 channels, and the number of channels in the four stages are 64, 128, 256 and 512 respectively, which maintains feature expression ability while controlling complexity.

[0073] The neck section adopts the FPN structure of FCOS, which fuses and connects the multi-scale feature maps output by LSM-TD from top to bottom and laterally to generate feature layers that are adapted to targets of different scales. Combined with optimization strategies, the number of FPN output channels is set to 96, which enhances the ability to aggregate cross-scale information of small targets while balancing computational cost and feature capacity.

[0074] The detection head is based on an improved FCOS architecture and adopts a lightweight combination of "Conv2d + deformable convolution DCN". It retains two layers of convolution stacking and replaces traditional convolution with DCN to dynamically adjust the receptive field and accurately capture the irregular shape of small targets. At the same time, it optimizes the channel dimension and parameter configuration, and divides them into three branches: classification, regression and centrality. They output the class probability, target box coordinates and centrality score respectively. With the help of Focal Loss (classification loss), DIoU Loss (regression loss) and cross-entropy loss (centrality loss), it effectively filters low-quality detection boxes and improves the localization accuracy.

[0075] During the model training phase, the LSM-TD backbone was fine-tuned on the AI-TODv2 dataset.

[0076] S3: Using the preprocessed remote sensing image as input to the LSM Net network, the remote sensing lightweight target detection method based on local state modeling network is used to train the LSM Net network.

[0077] To further accurately capture minute targets, this invention designs a lightweight remote sensing target detection method based on a local state modeling network. LSMNet uses an improved lightweight backbone network, LSM-TD, as its feature extraction backbone, and introduces a local state modeling module, LSMBlock, at a key stage of the backbone network. This enables collaborative modeling of the local structural information and long-range spatial dependencies of minute targets, thereby obtaining feature representations with both fine spatial perception and global context modeling capabilities without relying on self-attention mechanisms or complex sequence scanning. Structurally, LSMNet mainly includes an input processing module, a lightweight LSM-TD backbone network, the LSMBlock local state modeling module, and a detection head module. The LSM-TD backbone network adopts a staged hierarchical structure design. By reallocating network depth and channel resources, it significantly enhances the modeling capability of high-resolution features in the shallow stages, allowing the network to fully preserve the edge, shape, and local texture information upon which minute targets depend from the early stages. Simultaneously, downsampling and channel expansion are gradually introduced in subsequent stages to ensure necessary semantic extraction capabilities. This structure avoids the problem of small target information loss caused by premature downsampling in traditional lightweight networks.

[0078] In the key feature extraction stage of the backbone network, LSMNet introduces the local state modeling module LSMBlock. Based on convolution operations, this module combines local convolutional feature extraction with two-dimensional state space modeling. Through separable depthwise separable convolutions in the row and column directions, it simulates the state propagation process of features in the spatial dimension, thereby effectively expanding the receptive field and establishing long-distance spatial dependencies while maintaining the two-dimensional feature structure without unfolding. Compared to modeling methods based on sequence scanning or self-attention, this design significantly reduces tensor rearrangement, memory access, and computational overhead, making it more suitable for deployment on resource-constrained remote sensing edge devices.

[0079] The overall process is as follows: First, the remote sensing image is input into the LSM-TD backbone network, and then feature representations at different scales are obtained through multi-stage convolutional feature extraction. Among them, shallow features focus on maintaining spatial resolution and fine-grained structural information, while deep features gradually enhance semantic expressive power. Finally, in the feature extraction process at each stage, LSMBlock first performs local convolutional enhancement on the input features, and then realizes long-range information interaction across rows and columns through lightweight two-dimensional state modeling, ultimately fusing local detailed information with global contextual information for output.

[0080] Through the above structural design, LSM Net achieves a unified approach to remotely sensed small targets: clear visibility, long-range connectivity, and low computational cost. On one hand, it effectively improves the detection stability for densely packed small targets and complex background scenes; on the other hand, it significantly outperforms existing detection networks based on Transformers or large convolutional kernels in terms of parameter scale and computational cost, demonstrating good engineering practicality and deployment value. This network structure can be flexibly adapted to various mainstream detection frameworks, exhibiting strong versatility and scalability.

[0081] Through the above process, this invention improves the detection accuracy of small targets on public remote sensing datasets while reducing the number of parameters and computational complexity. It successfully solves the problems of extremely small target scale, weak texture information, complex background and high sensitivity to computing resources in traditional methods, and provides a feasible lightweight detection solution for edge computing scenarios.

[0082] Specifically, in the embodiments of this application, the LSM Net network is trained based on the local state modeling network as follows:

[0083] S3.1: Network Initialization: Initialize the weight parameters of the LSM Net network.

[0084] S3.2: Forward Propagation: A batch of preprocessed training images are input into the network and processed sequentially through the LSM-TD backbone network, the local state modeling module, the feature pyramid neck network, and the target detection head to obtain the prediction results.

[0085] S3.3: Loss Calculation: Based on the prediction results and the true labels of the image annotations, calculate the classification loss and localization loss.

[0086] S3.4: Backpropagation and parameter update: Using an optimization algorithm, calculate the gradient based on the loss and update all trainable parameters of the network.

[0087] S3.4: Iterative optimization: Repeat steps (2) to (4) until the preset training rounds or performance indicators are reached.

[0088] S4: Input the remote sensing image to be detected into the trained LSM Net network. After multi-stage feature extraction, local-global joint modeling and detection head prediction, the output includes the detection results containing the small target category, location coordinates and corresponding confidence.

[0089] The method for object detection is as follows:

[0090] S4.1: Multi-stage feature extraction and high-resolution information preservation: The input remote sensing image is fed into the lightweight backbone network LSM-TD of LSMNet, and image features are extracted layer by layer through multi-stage convolution operations. During the feature extraction process, more computing resources are allocated in the shallow stage and the downsampling operation is delayed to enhance the modeling ability of high-resolution spatial features, so as to preserve the edge, shape and local texture information contained in the small target and avoid information loss of the small target in the early stage of feature extraction.

[0091] S4.2: The intermediate features output from the backbone network are input into the local state modeling module LSMBlock for further enhancement. LSMBlock first performs local convolution operations on the input features to extract local structural information in the neighborhood of the small target. Then, through separate two-dimensional state modeling operations along the row and column directions of the feature map, long-range information propagation of the features in the spatial dimension is achieved. This establishes the correlation between different spatial locations without unfolding the two-dimensional feature structure, thereby enhancing the ability to distinguish between the small target and its context.

[0092] The overall implementation method of LSMBlock is as follows:

[0093] Input feature map ,in For batch size, For the number of channels, and These refer to the space's height and width, respectively.

[0094] Local feature extraction:

[0095] For input features implement Depthwise separable convolution operations are used to extract local neighborhood information and obtain preliminary local features. Its operation can be expressed as:

[0096] ;

[0097] To stabilize training and enhance nonlinear representation capabilities, batch normalization (BN) and an activation function (such as GELU or SiLU) are typically added after convolution. Therefore, complete local feature extraction can be represented as:

[0098] ;

[0099] in, Represents the activation function;

[0100] Lightweight 2D State-Space Modeling (L-2DSSM):

[0101] To capture long-range spatial dependencies, a lightweight two-dimensional state-space modeling method is introduced. This method avoids unfolding feature maps into one-dimensional sequences and instead directly simulates state propagation in two-dimensional space.

[0102] The local features obtained in the previous step Input a lightweight two-dimensional state-space model (L-2DSSM) and perform sequence modeling along the height and width dimensions to capture long-range dependency features. The mathematical definition of this process is:

[0103]

[0104] This formula clarifies that the function of the L-2DSSM operation is to transform local features. Mapped to features implying long-range spatial dependence .

[0105] L-2DSSM is implemented through row- and column-separated depthwise convolutions:

[0106] use The depthwise convolutional kernel performs convolutions along the height direction, capturing long-range inter-line dependencies:

[0107]

[0108] use The depthwise convolutional kernel performs convolution in the width direction, capturing long-range inter-column dependencies:

[0109]

[0110] Will The receptive fields are decomposed into sequence. and Convolution can expand the effective receptive field at minimal cost, while maintaining channel independence and fine-grained spatial structure that is crucial for the localization of small targets.

[0111] Merge information from two directions:

[0112]

[0113] Finally, use groups of 8. Grouped convolutions are used to perform channel fusion, resulting in a state-space modeled output:

[0114]

[0115] Channel Mixing (MLP):

[0116] After capturing spatial dependencies, a lightweight multilayer perceptron (referred to as RepVitMLP in the paper) is used for channel mixing to enhance feature representation. Its operation is as follows:

[0117]

[0118] The first pointwise convolutional layer reduces the channel dimension from... Down to The second layer of grouped convolutions reduces the channel dimension from Restore to .

[0119] Feature fusion and output:

[0120] By using residual connections, the original input of the module is... Features after the above processing The features are added together. Finally, through batch normalization (BN) integration, the final output features of LSMBlock are obtained. :

[0121]

[0122] S4.3: The LSMBlock-enhanced features are fed into the detection head network to jointly predict the target category, location parameters, and confidence. The detection head regresses the corresponding target bounding box position based on the response value of each spatial position in the feature map, and outputs the target category and confidence score corresponding to each bounding box. Finally, the detection result containing information on the small target category, spatial position, and detection confidence is obtained.

[0123] To verify the effectiveness of the improved backbone, further ablation experiments were conducted, as shown in Table 1. These experiments demonstrate the performance comparison of different detectors combined with different backbone networks on the AI-TODv2 dataset. The proposed LSM-TD backbone network achieves a good balance between detection accuracy and computational efficiency across various detection frameworks. Particularly in the ATSS detector, LSM-TD achieves the highest detection accuracy among all contrasted backbone networks with an AP of 16.0, while its parameter count is only 10.81M, maintaining the lowest scale among all evaluation models, approximately one-third that of ResNet50.

[0124] Compared to the ResNet50 backbone, LSM-TD achieves a 23.1% relative improvement in overall detection performance, and demonstrates a more significant advantage in AP75 metrics and sensitivity to small target detection. Similar performance trends were observed under the FCOS and RetinaNet frameworks, where LSM-TD consistently achieves superior detection accuracy compared to other lightweight backbone networks while maintaining its lightweight characteristics.

[0125] In summary, the experimental results fully demonstrate the effectiveness of LSM-TD as an efficient and lightweight backbone network: it can significantly reduce model complexity while steadily improving the detection performance of small targets.

[0126]

[0127] Compared with other methods, this study systematically improves upon the lightweight baseline model RepMamba. RepMamba achieves only 6.0% AP with a computational cost of 51.22 GFLOPs and a parameter scale of 6.61M. As shown in Table 2, the proposed LSM Net achieves continuous and significant improvements in both computational efficiency and detection accuracy.

[0128] Specifically, LSM Net reduces computational overhead from 51.22 GFLOPs to 26.89 GFLOPs, a reduction of approximately 47%; simultaneously, the number of parameters is compressed from 6.61M to 4.14M, demonstrating outstanding lightweight characteristics at the network architecture level. Meanwhile, detection performance is significantly enhanced, with mAP increasing from 6.0% to 17.7%, and AP50 and AP75 increasing to 42.5% and 11.4%, respectively. These results demonstrate that LSM Net effectively enhances target localization accuracy and classification confidence while maintaining high computational efficiency, thus significantly alleviating the shortcomings of the original baseline model in terms of detection capabilities.

[0129] Further comparison with the high-precision, lightweight detector LTDNet reveals that LTDNet achieves an AP of 23.0% with a computational cost of 38.19 GFLOPs, while LSMNet achieves approximately a 30% reduction in computational cost, at the expense of only a moderate difference in accuracy. This result indicates that LSMNet's accuracy-efficiency trade-off curve leans more towards a design approach that prioritizes extreme efficiency and practical deployment friendliness.

[0130]

[0131] The target detection technology of this invention can accurately identify and locate targets (such as vehicles, buildings, ships, and various facilities) in different scenarios, and the labeled boxes clearly and intuitively present the detection results. Its visualization results reveal the model's current advantages and challenges. The model exhibits good robustness in scenarios with dense vehicle distribution, oil tank scenarios, complex backgrounds, and ship scenarios, verifying its reliable basic detection capability under common small targets and complex background conditions. However, the failed sample (f) and the representative high-difficulty sample (e, airport scenario) together expose two core problems that still need to be solved: insufficient ability to model rotating targets and limited adaptability to extreme multi-scale changes.

[0132] Specifically, for targets with clear orientation, such as airport aircraft, traditional horizontal bounding box regression methods struggle to accurately fit their orientation, leading to decreased positioning accuracy. Furthermore, extreme scale variations in the scene pose a challenge to existing feature pyramid networks for effectively representing targets of different scales at the same level, resulting in missed or false detections within certain scale ranges.

[0133] Future improvements should focus on introducing lightweight rotating sensing heads and dynamic multi-scale feature fusion mechanisms to enhance the localization accuracy of targets with arbitrary orientations and improve the robustness of the model in the face of extreme scale changes.

[0134] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A lightweight remote sensing target detection method based on local state modeling networks, characterized in that, Includes the following steps: S1: Obtain the remote sensing image dataset and perform preprocessing; S2: Construct an LSM Net network, including a backbone network and a neck network; wherein the backbone network is LSM-TD and the neck network is a feature pyramid; S3: Use the preprocessed remote sensing image as input to the LSM Net network and train the LSM Net network based on local state modeling; S4: Input the remote sensing image to be detected into the trained LSM Net network. After multi-stage feature extraction, local-global joint modeling and detection head prediction, the output includes the detection results containing the small target category, location coordinates and corresponding confidence.

2. The lightweight remote sensing target detection method based on local state modeling networks according to claim 1, characterized in that, The construction of the LSM Net network in step S2 includes the following steps: S1: Constructing the LSM-TD backbone network: Design a hierarchical four-stage network architecture with an output channel ratio of 5:5:2:1 for the four stages. The first and second stages are high-resolution stages used for densely stacking feature enhancement modules to preserve spatial details; the third and fourth stages are low-resolution stages used for feature extraction. S2: Constructing the feature pyramid neck network: Use FPN as the neck network to receive the outputs from different stages of the LSM-TD backbone network and generate multi-scale feature maps. S3: Insert a local state modeling module into the backbone of the network to enhance the features. S4: The input end of the target detection head is connected to the output end of the feature pyramid neck network, and is used to predict the target category and bounding box based on the enhanced multi-scale features.

3. The lightweight remote sensing target detection method based on local state modeling networks according to claim 2, characterized in that, The local state modeling module comprises the following modules: a local feature extraction unit, used to extract local neighborhood information of input features through depthwise separable convolution; a two-dimensional sequence modeling unit, used to perform long-range dependency modeling on the local neighborhood information along the spatial height and width directions respectively; and a channel mixing unit, used to perform inter-channel interaction and integration on features that have fused long-range dependency information. The feature fusion unit is used to fuse the output of the channel mixing unit with the original input features through residual connections to obtain enhanced output features.

4. The lightweight remote sensing target detection method based on local state modeling networks according to claim 2, characterized in that, The overall implementation method of the local state modeling module includes the following steps: inputting feature maps. ,in For batch size, For the number of channels, and These represent spatial height and width, respectively; S1: Local feature extraction: for input features implement Depthwise separable convolution operations are used to extract local neighborhood information and obtain preliminary local features. Its operation can be expressed as: To stabilize training and enhance nonlinear expressive power, batch normalization and activation functions are typically added after convolution; therefore, complete local feature extraction can be represented as: ;in, S2: Lightweight 2D State Space Modeling: To capture long-range spatial dependencies, lightweight 2D state space modeling is introduced; this method avoids unfolding the feature map into a one-dimensional sequence, but instead directly simulates state propagation in a two-dimensional space; the local features obtained in the previous step are... Input a lightweight two-dimensional state-space model and perform sequence modeling along the height and width dimensions to capture long-range dependency features; the mathematical definition of this process is: This formula clarifies that the function of the L-2DSSM operation is to transform local features... Mapped to features implying long-range spatial dependence L-2DSSM is implemented through row- and column-separated depthwise convolutions: using The depthwise convolutional kernel performs convolutions along the height direction, capturing long-range inter-line dependencies: ;use The depthwise convolutional kernel performs convolution in the width direction, capturing long-range inter-column dependencies: ;Will The receptive fields are decomposed into sequence. and Convolution can expand the effective receptive field with minimal cost, while maintaining channel independence and fine-grained spatial structure crucial for the localization of small targets; it combines information from two directions. Finally, use groups of 8. Grouped convolutions are used to perform channel fusion, resulting in a state-space modeled output: S3: Channel Blending: After capturing spatial dependencies, a lightweight multilayer perceptron is used for channel blending to enhance feature representations; its operation is as follows: The first layer of pointwise convolution reduces the channel dimension from... Down to The second layer of grouped convolutions reduces the channel dimension from Restore to S4: Feature Fusion and Output: Through residual connections, the original input of the module is fused and output. Features after the above processing Add them together; finally, through batch normalization integration, the final output features of the local state modeling module are obtained. ： 5. The lightweight remote sensing target detection method based on local state modeling networks as described in claim 1, characterized in that, The method for training the LSM Net network in step S3 Includes the following steps: S1: Network initialization: Initialize the weight parameters of the LSM Net network; S2: Forward Propagation: Input a batch of preprocessed training images into the network, which are then processed sequentially through the LSM-TD backbone network, the local state modeling module, the feature pyramid neck network, and the target detection head to obtain the prediction results; S3: Loss Calculation: Calculate the classification loss and localization loss based on the prediction results and the true labels of the images; S4: Backpropagation and Parameter Update: Calculate the gradient based on the loss using an optimization algorithm and update all trainable parameters of the network. S5: Iterative optimization: Repeat steps S2 to S4 until the preset training rounds or performance indicators are reached.

6. The lightweight remote sensing target detection method based on local state modeling networks according to claim 5, characterized in that, The optimization algorithm used in step S4 is AdamW, and a learning rate warm-up and step decay strategy is adopted.

Images