A lightweight target detection model adaptable to various remote sensing scenarios

Abstract

This invention discloses a lightweight target detection model adaptable to various remote sensing scenarios. An improved reparameterizable convolutional block is introduced into the backbone network to achieve multi-scale feature enhancement and detail capture with low computational cost. A simplified bottom-up feature pyramid structure is adopted in the neck network, combined with low-frequency feature extraction and cross-layer residual concatenation, significantly reducing computational cost and parameter count while effectively fusing low-level details with high-level semantics. A lightweight task alignment structure is set in the detection head, with an attention mechanism embedded in the classification branch and deformable convolution in the regression branch, thereby improving the accuracy of small target detection. Normalized Gaussian Wasserstein distance is introduced in the bounding box regression to enhance the stability and robustness of localization, balancing model lightweighting and detection accuracy, enabling efficient and stable small target detection in complex remote sensing scenarios.

Description

Technical Field

[0001] This invention relates to the fields of image processing and computer vision technology, specifically to a lightweight target detection model that can adapt to remote sensing in multiple scenarios. Background Technology

[0002] Target detection in remote sensing images is one of the core tasks in remote sensing image analysis and processing, and is widely used in fields such as environmental monitoring, urban planning, disaster assessment, and military reconnaissance. With the continuous development of remote sensing technology, image resolution is gradually improving, and the amount of data acquired is constantly increasing, which places higher demands on the accuracy and efficiency of target detection.

[0003] Existing methods for target detection in remote sensing images mainly include traditional image processing methods and deep learning-based methods. Traditional methods often rely on manually designed features, such as edges, textures, and shapes, combined with classifiers to achieve target recognition. These methods can achieve certain results under specific conditions, but their detection performance significantly degrades when faced with complex scenes, varying backgrounds, and scale differences in remote sensing images. In recent years, deep learning methods, especially convolutional neural networks (CNNs) and their derived detection models (such as the YOLO series), have made significant progress in target detection tasks for both natural and remote sensing images. By automatically learning image features, deep learning methods have improved the accuracy and robustness of detection to some extent. Nevertheless, existing technologies still have the following shortcomings:

[0004] 1. Large scale variation: The target size in remote sensing images varies significantly, and existing methods are prone to detection failure when dealing with extreme scale variations.

[0005] 2. Complex background: Remote sensing images usually contain complex and varied ground features and noise information, which can easily lead to false detections and missed detections.

[0006] 3. Difficulty in detecting small targets: Many remote sensing targets are small in size and are often ignored or confused with the background, resulting in insufficient detection accuracy.

[0007] 4. Difficulty in balancing accuracy and efficiency: To improve the performance of small target detection, some methods add multi-scale feature fusion or high-resolution input, but this significantly increases the computational load and storage requirements, making it difficult to meet real-time requirements.

[0008] 5. Insufficient generalization ability: Due to the significant differences in different satellite platforms, sensors and imaging conditions, the performance of existing detection models degrades significantly when applied across domains, lacking good generalization ability.

[0009] In view of this, the present invention proposes a lightweight target detection model that can adapt to remote sensing in multiple scenarios. Summary of the Invention

[0010] The purpose of this invention is to provide a lightweight target detection model that can adapt to multiple remote sensing scenarios, aiming to solve the problems of difficulty in identifying small targets in remote sensing image target detection, easy false detection and missed detection in complex backgrounds, difficulty in balancing accuracy and efficiency, and insufficient cross-domain generalization ability.

[0011] In a first aspect, the present invention provides a lightweight target detection model that can adapt to remote sensing in multiple scenarios, comprising a backbone network, a neck network, and a detection head, wherein:

[0012] The backbone network is used to extract multi-scale features through improved reparameterizable convolutional blocks and generate single-branch convolutional kernels by parameter fusion of multi-scale features, thereby enhancing the ability to express multi-scale features and capture local details under low computational conditions.

[0013] The neck network module is used to upsample and stitch together the branch feature layers output by the backbone network through a simplified bottom-up feature pyramid network. After stitching, a low-frequency feature extraction module and a cross-layer residual stitching structure are introduced to achieve multi-level feature fusion.

[0014] The detection head is used to perform classification and regression based on a lightweight task alignment structure. It extracts the target region based on the classification branch embedding auxiliary attention mechanism, and the regression branch uses deformable convolution to improve the bounding box localization accuracy of the target region, thereby achieving high-precision detection of small targets.

[0015] As a preferred technical solution of the present invention, the improved reparameterizable convolutional block includes the organic integration of multi-branch convolutional structure, pointwise convolution and attention mechanism;

[0016] Multi-scale features are obtained by mapping branches to multi-branch convolutional structures, thereby enhancing edge texture features and mid-scale semantic features.

[0017] Pointwise convolutional layers and spatial attention units are placed after the multi-branch convolutional structure to achieve linear combination and feature recombination between channels while maintaining low computational cost.

[0018] During the inference phase, multi-branch convolutional kernels and bias terms are equivalently fused to generate single-branch convolutional kernels, thereby reducing computational complexity. These kernels are then connected in series with pointwise convolutional layers to adaptively adjust the weights of features in each channel, highlighting key target regions and suppressing background interference.

[0019] As a preferred technical solution of the present invention, the multi-branch convolutional structure extracts features in parallel through 3×3 convolution, 1×1 convolution and identity mapping branches during the training phase, and the parameter fusion module performs equivalent fusion of the above-mentioned branch convolutions during the inference phase.

[0020] As a preferred embodiment of the present invention, the application logic of the simplified bottom-up feature pyramid network is as follows:

[0021] From the multi-layer feature map output by the backbone network, only three layers, P3 to P5, are selected as input; the P5 feature map is upsampled and concatenated with the P4 feature map; then the fused feature is upsampled and concatenated with the P3 feature map, thereby realizing the layer-by-layer transmission of semantic information from bottom to top.

[0022] A low-frequency feature extraction module is set up after the above splicing to highlight the edge and texture details of small targets and avoid the loss of details during the multi-layer feature fusion process;

[0023] The residual information of the downsampling path is further introduced into the upsampling path to maintain the balance between high-level semantics and low-level details and improve the robustness of feature fusion.

[0024] As a preferred embodiment of the present invention, the low-frequency feature extraction module includes:

[0025] The convolutional filtering unit is used to perform low-pass convolution on the fused feature map to extract low-frequency components;

[0026] Feature normalization units are used to suppress background noise;

[0027] Nonlinear mapping units are used to enhance the saliency of small targets in complex backgrounds; and

[0028] The feature output unit is used to concatenate the enhanced low-frequency features with the cross-layer residual information and transmit them to the detection head.

[0029] As a preferred embodiment of the present invention, the lightweight task alignment structure includes:

[0030] Shared convolutional stem: used to perform unified feature extraction on the input fusion features to avoid redundant calculations between classification and regression branches, thereby reducing the number of parameters and computational overhead;

[0031] Classification branch: It is connected to the shared convolutional stem and embeds an auxiliary attention mechanism to enhance the weights of salient target regions and suppress background interference, so as to achieve accurate classification of small targets;

[0032] Regression branch: Connected to the shared convolutional stem, it uses deformable convolutional units to adaptively adjust the convolutional sampling position to improve the flexibility and accuracy of bounding box regression;

[0033] Task alignment unit: Used to establish consistency constraints on feature distribution between classification and regression branches, so that classification results and localization results are coordinated, thereby improving overall detection accuracy.

[0034] As a preferred embodiment of the present invention, the classification branch of the detection head adopts a channel attention mechanism, and the regression branch adopts a second type of deformable convolution.

[0035] As a preferred technical solution of the present invention, a normalized Gauss-Wasestein distance is introduced into the regression branch for bounding box localization accuracy regression. The Gauss-Wasestein distance is modeled by the difference between the mean vector and covariance matrix of the predicted box and the true box, thereby minimizing the bounding box localization error.

[0036] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0037] This invention achieves multi-scale feature enhancement and detail capture under low computational requirements by introducing improved reparameterizable convolutional blocks into the backbone network, ensuring both lightweight design and high efficiency. A simplified bottom-up feature pyramid structure is employed in the neck network, combined with low-frequency feature extraction and cross-layer residual concatenation, enabling efficient fusion of low-level detail information and high-level semantic information, preventing the loss of small target features during transmission. In the detection head, a lightweight task alignment structure is used to reasonably decouple classification and regression tasks, and optimization is achieved through attention mechanisms and deformable convolutions, significantly improving the detection accuracy of small targets in complex backgrounds. Normalized Gaussian Wasserstein distance is introduced in the bounding box regression stage to improve the stability and robustness of the fitting between predicted and ground truth boxes through distribution modeling. Overall, this invention balances lightweight detection models, cross-scale expressive capabilities, and small target detection accuracy, enabling efficient, stable, and well-generalized target detection in various remote sensing scenarios. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0039] Figure 1 This is a structural diagram of the lightweight target detection model of the present invention, which can be adapted to remote sensing in multiple scenarios.

[0040] Figure 2 This is a flowchart of the lightweight and multi-scenario adaptable target detection method of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings.

[0042] Throughout the accompanying drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. The described embodiments are only a part of the embodiments of this application, not all of them. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0043] To address the challenge of balancing accuracy and efficiency in existing small target detection technologies, this paper presents a collaborative optimization design for a backbone network, feature fusion network, detection head, and bounding box regression module. This design, based on in-depth analysis of feature extraction, semantic transfer, and target localization mechanisms in small target detection tasks, significantly improves detection performance while effectively reducing computational resource consumption. The design approach for each module combination has been systematically analyzed and experimentally verified, demonstrating its irreplaceability and non-obviousness.

[0044] Example 1

[0045] Please see Figure 1-2 As shown, this embodiment provides a lightweight target detection model that can adapt to multiple remote sensing scenarios, including three parts: a backbone network, a neck network, and a detection head;

[0046] The backbone network includes an improved reparameterizable convolutional block; the improved reparameterizable convolutional block organically integrates multi-branch convolutional structures, pointwise convolutions and attention mechanisms, enhancing multi-scale feature representation capabilities and local detail capture capabilities while maintaining lightweight design.

[0047] It should be noted that the input features are first processed through residual depthwise convolution to extract fine-grained information, then pointwise convolution and Gaussian error linear unit activation functions are used to enhance nonlinear expressive power, and global response normalization is combined to enhance spatial information interaction and channel correlation. Compared with existing convolutional backbones, the improved reparameterizable convolutional blocks significantly improve the accuracy of small target detection while reducing computational cost. Its design logic stems from a systematic analysis of the limitations of traditional convolution and multi-scale feature extraction, which is obvious to those not skilled in the art. In addition, the backbone network structure is a combination of conventional convolution and improved reparameterizable convolution.

[0048] Specifically, the improved reparameterizable convolutional block includes:

[0049] The multi-branch convolutional structure consists of parallel 3×3 convolutions, 1×1 convolutions, and identity mapping branches, and is used to improve the diversity of feature representations during the training phase.

[0050] Pointwise convolutional layers, placed after multi-branch convolutional structures, are used to achieve linear combination and feature recombination between channels while maintaining low computational cost.

[0051] The parameter fusion module is used to equivalently fuse multi-branch convolution kernels and bias terms during the inference phase to generate single-branch convolution kernels, thereby reducing computational complexity.

[0052] The attention mechanism unit, set in series with the pointwise convolutional layer, is used to adaptively adjust the weights of features in each channel, highlighting key target regions and suppressing background interference.

[0053] In other words, the improved reparameterizable convolutional block includes a multi-branch convolutional structure, a pointwise convolutional layer, a parameter fusion module, and an attention mechanism unit. The multi-branch convolutional structure extracts features in parallel during the training phase through 3×3 convolution, 1×1 convolution, and identity mapping branches. The parameter fusion module performs equivalent fusion of the above-mentioned branch convolutions during the inference phase to reduce computational overhead.

[0054] To further explain, the application logic of the improved reparameterizable convolutional block is as follows:

[0055] The call is performed in both the shallow and middle layers of the backbone network. In the shallow layer, the branch features of local edges and texture details are extracted through a multi-branch convolutional structure. In the middle layer, the above branch features are fused to form a single-branch convolutional kernel through parameter fusion.

[0056] The single-branch convolutional kernel is combined with a residual structure and a spatial attention mechanism to highlight salient target regions and suppress background interference, and the output enhanced feature map is used to pass to the neck network.

[0057] In other words, the improved reparameterizable convolutional block enhances feature diversity through multi-branch convolution during the training phase, reduces computational overhead through parameter fusion during the inference phase, and works in conjunction with cross-stage partial residual structures and spatial attention mechanisms to highlight target regions in complex contexts, thereby ensuring that features still have high discriminative power when passed to the neck network.

[0058] The neck network comprises a simplified bottom-up feature pyramid network to replace the traditional path aggregation structure. Through a systematic analysis of the roles of low-level edge information and high-level semantic features in small target detection, this invention proposes feature fusion only at layers P3 to P5: P3 is responsible for capturing low-level details, while P4-P5 enhance semantic expression through bottom-up connections. Experiments have verified that adding layer P2 increases redundant computation, while using only layers P4-P5 cannot construct a complete bottom-up pyramid network, resulting in limited low-level information transmission. This design ensures both semantic integrity and fine-grained information transmission, while reducing computation by 19.3% and parameter count by 38.2% simply by replacing the original pyramid structure, demonstrating the irreplaceable nature of the module combination. After each concatenation, the neck network includes a low-frequency feature enhancement module, maintaining the integrity of global semantics and low-frequency information during cross-scale feature fusion, thereby enhancing the robustness and accuracy of remote sensing target detection / classification.

[0059] The detection head employs a lightweight task-aligned structure, including a shared convolutional stem, a classification branch, and a regression branch. By using a task decoupling strategy through shared convolution, the fused features are decomposed into classification and regression tasks, and global contextual information is incorporated to enhance expressive power. Specifically, the regression branch uses deformable convolutions to achieve high-precision target localization, while the classification branch introduces an auxiliary attention mechanism to highlight key regions. Experiments show that traditional detection head structures cannot simultaneously meet the requirements of lightweight design and multi-scale small target detection. The combined design of this invention significantly improves detection performance through collaborative optimization.

[0060] An auxiliary attention mechanism is embedded in the classification branch to highlight salient regions; deformable convolution is used in the regression branch to improve the flexibility and accuracy of bounding box localization. In the bounding box regression process, the predicted and ground truth bounding boxes are modeled as two-dimensional Gaussian distributions. This embodiment further introduces normalized Gaussian Wasserstein distance as a loss function, which is used to fit the relationship between the predicted and ground truth bounding boxes through the differences in the distributions of the mean and covariance, thereby improving the stability and robustness of small target localization.

[0061] In summary, this invention constructs a complete small target detection system for remote sensing images through the collaborative design of an improved backbone network, a simplified bottom-up feature pyramid network, a lightweight task-aligned detection head, and a normalized Gaussian Wasserstein regression method. This system effectively addresses the technical problems of difficulty in identifying small targets, susceptibility to false positives and false negatives in complex backgrounds, and insufficient cross-domain generalization ability in remote sensing image target detection. By improving convolutional blocks to enhance multi-scale feature representation capabilities, utilizing a simplified feature pyramid network for effective fusion, employing a lightweight task-aligned detection head to improve classification and localization capabilities, and achieving stable and accurate small target regression based on normalized Gaussian Wasserstein distance, this invention solves the technical problems of difficulty in detecting small targets in remote sensing images, susceptibility to false positives in complex backgrounds, difficulty in balancing accuracy and efficiency, and insufficient cross-domain generalization ability.

[0062] Example 2

[0063] The parts not detailed in this embodiment are shown in Embodiment 1. To verify the effectiveness of the lightweight and multi-scenario adaptable remote sensing image target detection method described in this invention, a comparative experiment was conducted with existing mainstream detection models YOLOv13s, YOLOv11s, YOLOv8s, and YOLOv5s. The experimental datasets were selected from publicly available remote sensing image datasets AI-TODv1, AI-TODv2, VisDrone2019, and CODrone, covering various scenarios such as satellite remote sensing imagery, UAV low-altitude imagery, and dynamic aerial imagery. This invention contains only 6.5M parameters and 20.0 GFLOPs, making it the lightest among all the compared models. Compared with YOLOv5s (9.1M) and YOLOv8s (11.1M), the number of parameters is reduced by 38%–71%, demonstrating efficient model compression and computational capabilities.

[0064] Therefore, the comparison metrics include model parameter count, computational cost (GFLOPs), accuracy, recall, and mean AP. The experimental results are shown in Tables 1–4, where:

[0065] Table 1: Comparative test data on the AI-TODv1 public dataset;

[0066]

[0067] As shown in Table 1, this invention achieves the highest average precision (44.7%) and recall (44.6%) while having fewer parameters and lower computational cost compared to the comparative model. This demonstrates that this invention can significantly improve the detection accuracy and coverage of dense small targets while maintaining lightweight design, thus solving the problem of existing methods struggling to balance accuracy and efficiency.

[0068] Table 2: Comparative test data on the AI-TODv2 public dataset

[0069]

[0070]

[0071] As shown in Table 2, this invention maintains the highest average precision (54.4%) and recall (51.9%) in complex backgrounds and larger-scale remote sensing images, outperforming YOLOv8s (52.6%) and YOLOv5s (51.3%). This indicates that this invention has stronger robustness and generalization ability in complex cross-domain scenarios, overcoming the performance degradation shortcomings of existing technologies under different imaging conditions.

[0072] Table 3: Comparative test data on the VisDrone 2019 public dataset;

[0073]

[0074] As shown in Table 3, the present invention achieved the best results in terms of average precision (41.4%) and recall (40.2%) in low-altitude UAV imagery, which is higher than YOLOv5s (39.4%) and YOLOv8s (39.9%). This indicates that the present invention has higher stability and reliability when dealing with complex backgrounds and small-sized targets, and can effectively solve the problem of unstable detection of small targets under the perspective of low-altitude UAVs.

[0075] Table 4: Comparative test data on the CODrone public dataset;

[0076]

[0077]

[0078] As shown in Table 4, in dynamic aerial photography scenarios, the present invention outperforms the comparative model in both average precision (28.8%) and recall (31.3%). Especially in the presence of rapid scale changes and motion blur, the present invention maintains high detection robustness and target localization accuracy, thus overcoming the shortcomings of existing methods in performance degradation under dynamic and variable scenarios.

[0079] Experimental results in complex remote sensing scenarios show that this invention achieves the highest average precision (44.7%) and recall (44.6%), surpassing all baseline models. Compared to models such as YOLOv13s and YOLOv11s, it provides higher confidence, more stable, and better aligned bounding boxes. We also annotated several instances demonstrating that this invention successfully detected targets missed by other models, further highlighting its reliability and robustness under dynamic UAV perspectives.

[0080] In summary, the comparative experiment results on four publicly available remote sensing datasets show that:

[0081] While maintaining lightweight design, it achieves higher detection accuracy and recall on various public remote sensing datasets. In particular, it outperforms existing technologies in small target detection, complex background suppression, and cross-domain generalization capabilities, thus effectively solving the technical problems of difficult small target detection, numerous false positives and false negatives in complex backgrounds, difficulty in balancing accuracy and efficiency, and insufficient robustness across different scenarios in remote sensing image target detection.

[0082] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A lightweight and multi-scene adaptable remote sensing target detection model, characterized in that, The backbone network, the neck network and the detection head are included, wherein: The backbone network is used for extracting multi-scale features through an improved re-parameterizable convolution block, generating a single-branch convolution kernel through a parameter fusion module, and enhancing the multi-scale feature expression ability and local detail capture ability under low computational load conditions; the improved re-parameterizable convolution block includes organic integration of a multi-branch convolution structure, a point-wise convolution and an attention mechanism; The multi-branch convolution structure is mapped to obtain multi-scale features, and the edge texture features and the mesoscale semantic features are enhanced; The point-wise convolution layer and the spatial attention unit are arranged after the multi-branch convolution structure, and are used for realizing linear combination and feature reorganization between channels under the premise of maintaining low computational load; In the inference stage, the multi-branch convolution kernel and the bias term are equivalently fused to generate a single-branch convolution kernel, thereby reducing the computational complexity; The neck network module is used for performing step-by-step upsampling and splicing on the branch feature layers output by the backbone network through a simplified bottom-up feature pyramid network, introducing a low-frequency feature extraction module and a cross-layer residual splicing structure after splicing to realize multi-level feature fusion; The detection head is used for performing classification and regression based on a lightweight task alignment structure, extracting a target region based on a classification branch embedded auxiliary attention mechanism, and improving the positioning accuracy of the bounding box of the target region by using a deformable convolution in a regression branch, so as to realize high-precision detection of small targets.

2. The target detection model for lightweight and multi-scene adaptable remote sensing according to claim 1, wherein, The multi-branch convolution structure extracts features in parallel through 3*3 convolution, 1*1 convolution and identity mapping branch in the training stage, and the parameter fusion module equivalently fuses the above branch convolutions in the inference stage.

3. The target detection model for lightweight and multi-scene adaptable remote sensing according to claim 2, characterized in that, The application logic of the neck network is as follows: From the multi-layer feature maps output by the backbone network, only three levels P3 to P5 are selected as inputs; the P5 feature map is upsampled, spliced with the P4 feature map, and then the fused feature is upsampled and spliced with the P3 feature map, so as to realize the layer-by-layer transmission of semantic information from bottom to top; The low-frequency feature extraction module is arranged after the above splicing, which is used to highlight the edge and texture detail information of small targets and avoid detail loss in the multi-level feature fusion process; The residual information of the down-sampling path is further introduced in the up-sampling path to maintain the balance between high-level semantics and low-level details and improve the robustness of feature fusion.

4. The target detection model for lightweight and multi-scene adaptable remote sensing according to claim 3, characterized in that, The low-frequency feature extraction module includes: A convolution filter unit is used to perform low-pass convolution on the fused feature map to extract low-frequency components; A feature normalization unit is used to suppress background noise; A nonlinear mapping unit is used to enhance the saliency of small targets in complex backgrounds; and A feature output unit is used to splice and transmit the enhanced low-frequency features and cross-layer residual information to the detection head.

5. The target detection model for lightweight and multi-scene adaptable remote sensing according to claim 4, characterized in that, The lightweight task alignment structure includes: A shared convolution stem is used to perform unified feature extraction on the input fused features to avoid repeated calculation of the classification branch and the regression branch, thereby reducing the parameter quantity and computational overhead; A classification branch is connected with the shared convolution stem, and an auxiliary attention mechanism is embedded to enhance the weight of the salient target region and suppress background interference, so as to realize accurate classification of small targets; and A regression branch adopts a deformable convolution to improve the positioning accuracy of the bounding box of the target region, thereby realizing high-precision detection of small targets. The regression branch is connected with a shared convolutional long connection, adopts a deformable convolution unit, and is used for adaptively adjusting a convolution sampling position, so as to improve flexibility and precision of bounding box regression; The task alignment unit is used for establishing a consistency constraint of feature distribution between the classification branch and the regression branch, so that the classification result and the positioning result are kept in cooperation, and thus overall detection precision is improved.

6. The target detection model for lightweight and multi-scene adaptable remote sensing according to claim 5, characterized in that, The classification branch of the detection head adopts a channel attention mechanism, and the regression branch adopts a deformable convolution.

7. The target detection model for lightweight and multi-scene adaptable remote sensing according to claim 6, characterized in that, In the regression branch, a normalized Gaussian Wasserstein distance is introduced in bounding box positioning precision regression, the Gaussian Wasserstein distance is modeled by a difference between a mean vector and a covariance matrix of a predicted frame and a real frame, and a bounding box positioning error is minimized.

Images