Dynamic attention multi-scale remote sensing target detection method

Abstract

The invention discloses a dynamic attention multi-scale remote sensing target detection method. A space-channel attention enhancement mechanism, a scale dynamic fusion strategy and a lightweight efficient detection structure are fused. According to the method, a high-coupling path from feature modeling to detection output is constructed based on a multi-dimensional feature regulation and control framework, so that the target detection performance in a scene with a complex background, violent scale change and serious shielding is improved. The system has the advantages of modular structure, flexible deployment, strong compatibility and the like, is suitable for rapid and accurate detection tasks of key targets such as airplanes, vehicles, ships and the like in multi-source remote sensing images, and has good engineering suitability and industrialization potential.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision and remote sensing image processing, and specifically provides a dynamic attention multi-scale remote sensing target detection method. Background Art

[0002] With the rapid advancement of remote sensing technology's refinement and commercialization, large amounts of high-resolution image data are being widely used in key areas such as urban governance, emergency monitoring, and military perception. At the same time, targets in remote sensing images exhibit significant small size, diversity, and dense distribution. Combined with complex backgrounds and large scales, these characteristics pose significant challenges to the accuracy and robustness of traditional target detection methods.

[0003] Widely used deep learning detection frameworks, such as the YOLO series, possess end-to-end modeling capabilities, but they still face key challenges in remote sensing imagery. These include insufficient shallow features to represent fine-grained structure, a lack of task-adaptive multi-scale fusion mechanisms, and limited detection head modeling capabilities for densely populated areas. These limitations are particularly pronounced in environments with clusters of small objects or complex occlusions. Therefore, a detection architecture with dynamic perception, structural adaptability, and deployment flexibility is urgently needed to comprehensively enhance object recognition capabilities and system practicality in remote sensing imagery. Summary of the Invention

[0004] The present invention aims to propose a remote sensing target detection method with spatially adaptive attention guidance and scale dynamic aggregation capabilities, breaking through the limitations of existing structures in feature fine modeling and task adaptability, and enhancing the ability to accurately identify small, dense, and heterogeneous targets in multi-source remote sensing images while ensuring computational efficiency.

[0005] To achieve the above object, the present invention proposes the following technical solutions:

[0006] The present invention provides a dynamic attention multi-scale remote sensing target detection method, comprising the following steps:

[0007] Step 1: Input remote sensing image data, use a lightweight deep neural network as the backbone network, embed a dynamic channel-spatial soft gating mechanism, and extract multi-dimensional discriminative features;

[0008] Step 2: Construct an integrated differentiable channel-space soft gating mechanism. The feature compression module based on the information bottleneck theory dynamically suppresses background redundant responses through the minimum mutual information regularization target and strengthens the perception weight of key areas.

[0009] Step 3: Design a hierarchical programmable multi-scale feature aggregation strategy, combined with heterogeneous convolution kernel receptive field enhancement technology, to achieve adaptive balanced modeling of feature semantic depth and spatial density;

[0010] Step 4: During the detection and decoding process, the system collaboratively performs the spatial attention-guided detection task through the refined positioning path and the discrimination optimization path, significantly improving the positioning accuracy and category differentiation ability of small targets.

[0011] Step 5: Build a task training set for various remote sensing scenarios and use a joint optimization strategy to perform end-to-end training and evaluation of the detection model;

[0012] Step 6: Deploy the trained model to a heterogeneous computing platform to implement real-time reasoning and target detection tasks for resource-constrained terminals.

[0013] Furthermore, the step (1) inputs the remote sensing image and extracts basic features, specifically:

[0014] The input image is uniformly resized to 1024 × 1024 pixels. Preliminary features are extracted through convolutional layers (e.g., Conv+C3k2) and downsampling operations, and the output features are feature maps P3, P4, and P5 of different scales.

[0015] Furthermore, the step (2) integrates the microchannel-spatial soft gating mechanism to enhance feature expression, specifically:

[0016] This part consists of two core sub-modules:

[0017] a. DynamicTanh activation function replaces the normalization layer:

[0018] As shown in the following formula (1), the traditional LayerNorm can be expressed as:

[0019] As shown in the following formula (2), DyT is replaced by the following dynamic nonlinear formula: DyT(x)=γ·tanh(αx)+β (2)

[0020] The dynamic hyperbolic tangent activation function (DyT) replaces the traditional normalization layer to eliminate the dependence on mean / variance calculation, reducing computational overhead while retaining the ability to align feature distributions.

[0021] b. TSSA self-attention mechanism:

[0022] As shown in the following formula (3), its goal is to minimize the following information bottleneck objective function ΔR(Z, Π) and improve attention to key spatial locations:

[0023] This mechanism is based on minimum mutual information regularization (MCR 2) theoretical framework, which realizes the intrinsic dimension compression of the feature manifold by constructing the variational upper bound of the second-order statistics, and drives the attention module to extract the key area response with strong semantic representation from the high-redundancy feature space, which is suitable for heterogeneous target detection in remote sensing images.

[0024] Furthermore, the step (3) is specifically as follows:

[0025] In order to fully integrate semantic and spatial information at different levels, this paper proposes a multi-scale adaptive fusion network. This module introduces three types of substructures in feature fusion:

[0026] a. SAF: Scale Alignment Unit

[0027] As shown in the following formula (4), the deep, medium and shallow features are scaled and concatenated: P n '=δ(C(Down(P n-1 )),P n ,U(P n+1 )) (4)

[0028] By downsampling, upsampling and channel compression of features at different scales, cross-level feature space-channel alignment is achieved, thereby realizing effective cross-layer information fusion.

[0029] b. AAF: History-aware aggregation unit

[0030] As shown in the following formula (5), the output of the previous layer and the surrounding layer features are added to each layer for deep fusion: P n ”=δ(Concat(C(Down(P' n-1 )),P' n ,C(U(P' n+1 )),P n-1 ”)) (5)

[0031] The historical layer output and adjacent scale features are further introduced to achieve multi-directional information coupling, thereby enhancing the semantic consistency across frames.

[0032] c. RepHELAN: Heterogeneous Convolution Receptive Field Enhancement Module

[0033] This architecture adopts a parallel processing mechanism of heterogeneous convolution kernel groups. Through the coordinated expansion of multi-scale receptive fields and resolution preservation technology, it achieves high-fidelity retention of shallow details and robust fusion of deep semantics in a unified feature space.

[0034] Furthermore, the step (4) is specifically as follows:

[0035] The present invention upgrades the structure of the traditional detection head and embeds spatial enhanced attention mechanisms in the regression and classification sub-branches to improve detection accuracy.

[0036] a. Regression branch

[0037] As shown in the following formula (6), the output is a four-directional distribution reg max ×4, decoded into coordinates by DFL (distributed focus loss):

[0038] We guide boundary distribution prediction via spatially enhanced attention and combine it with distributed focal loss (DFL) to decode continuous coordinates.

[0039] b. Classification branches

[0040] As shown in the following formula (7), using depth-wise separable convolution DWConv and SEAM:

[0041] A cascade structure of depthwise separable convolution and SEAM is used to improve the inter-class discrimination of small targets.

[0042] Furthermore, the step (6) is specifically as follows:

[0043] Construct a remote sensing image dataset containing multiple types of targets (such as aircraft, vehicles, ships, etc.) and use a multi-target joint loss function in the training phase, as shown in the following formula (8):

[0044] in: is the category cross entropy loss, Distributed focal loss for regression branch, is the IoU object box alignment loss.

[0045] The beneficial effects of the present invention are as follows: (1) By introducing a dynamic spatial-channel attention module, a multi-scale adaptive feature fusion network, and a spatial enhancement detection module, the present invention achieves full-process optimization from feature extraction to detection output, significantly enhancing the model's ability to perceive, fuse, and discriminate small-scale targets, occluded targets, and complex backgrounds in remote sensing images, effectively improving the integrity and accuracy of target recognition. (2) The overall method of the present invention fully considers the structural scalability and computational controllability of the model in its design, is compatible with the existing mainstream detection framework structure (YOLO / Transformer), is based on a modular quantization strategy and a tensor reparameterization mechanism, supports real-time remote sensing target detection under low power consumption conditions, and supports flexible deployment on multiple types of edge computing platforms. (3) The present invention is applicable to the detection of various remote sensing targets such as aircraft, vehicles, ships, and buildings, and is suitable for various high-resolution remote sensing image analysis scenarios such as urban monitoring, traffic management, military reconnaissance, and disaster assessment. It has good versatility and industrial application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A schematic diagram of the overall process provided by an embodiment of the present invention;

[0047] Figure 2 A schematic diagram of the redesigned YOLOv11 structure provided in an embodiment of the present invention;

[0048] Figure 3 A schematic diagram of the improved attention module structure provided by an embodiment of the present invention;

[0049] Figure 4 A schematic diagram of a multi-scale feature fusion structure constructed according to an embodiment of the present invention;

[0050] Figure 5 This is the detection result diagram of the basic model on the remote sensing dataset;

[0051] Figure 6 This is a graph of detection results on a remote sensing dataset provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0053] In this example, the above method was applied to the DOTA-v2.0 dataset, which contains 15 target categories: aircraft, ships, storage tanks, baseball fields, tennis courts, basketball courts, athletic fields, ports, bridges, large vehicles, small vehicles, helicopters, roundabouts, soccer fields, and swimming pools. Objects in this remote sensing image dataset exhibit a wide range of scales, diverse orientations, and significant dense occlusion. Using this dataset enables specific commercial scenarios such as urban regulation, port awareness, disaster assessment, and low-altitude emergency response, effectively validating the improved detection accuracy and reduced computational complexity achieved by our proposed method.

[0054] Figure 1 The flow chart of a dynamic attention multi-scale remote sensing target detection method of the present invention is shown. As shown in the figure, the present invention is implemented as follows, including the following steps:

[0055] S1. Cut the original remote sensing image into 1024×1024 pixels and perform image preprocessing;

[0056] S2. Integrating a direction-sensitive dynamic activation function, a second-order sparse regularized attention mechanism, and a channel-enhanced feedback structure into the backbone network to improve the ability to capture key areas and detailed information;

[0057] S3. Integrate a multi-scale adaptive feature pyramid network into the neck network and enhance the multi-scale feature representation through adaptive fusion units and heterogeneous convolution;

[0058] S4. Integrate the spatial enhanced attention mechanism into the detection head to improve the localization accuracy of small targets in complex backgrounds;

[0059] S5. Use the target detection method and YOLOv11 to build a new target detection model;

[0060] S6. Deploy this method in an edge-side heterogeneous computing architecture and observe the remote sensing image detection results.

[0061] Furthermore, the image preprocessing in step S1 is as follows: the input image is scaled to a uniform resolution of 1024×1024 by overlap compensation, RGB three-channel input is used, and multi-scale features are extracted through the C3k2 residual convolution group to completely avoid the risk of scale attenuation of small objects.

[0062] Furthermore, the direction-sensitive dynamic activation function, the second-order sparse regularized attention mechanism and the channel enhancement feedback structure in step S2 are specifically:

[0063] S21: The Dynamic Tanh activation layer uses the learnable parameters γ and β and the dynamic scaling factor α in formula (2) and replaces the traditional normalization operation with the nonlinear compression function γ·tanh(αx)+β. While retaining the feature distribution alignment function, it completely eliminates the mean / variance calculation overhead and significantly reduces the computational complexity.

[0064] S22: The AttentionTSSA mechanism adopts formula (3) in the above technical solution to construct a feature optimization target based on second-order statistics. It quantifies the global semantic association strength through the logarithm operation of the determinant of the feature matrix to achieve redundant direction compression and key feature retention.

[0065] S23: The feedforward network constructs a feature enhancement closed loop through the residual gating mechanism: the input features are manifold-adaptively reshaped by the Dynamic Tanh layer, significantly improving the inter-class separability of small targets and complex backgrounds; the activated two-dimensional tensor is input to the AttentionTSSA module, which implements global spatial dependency modeling with O(n) linear complexity based on the variational upper bound derivation and orthogonal projection optimization; finally, through dimensionality restoration and residual fusion, the edge gradient information is fully preserved while suppressing 90%+ of the background entropy source, forming a complete optimization system for channel-space joint enhancement.

[0066] Furthermore, the multi-scale feature fusion method in step S3 is specifically as follows:

[0067] S31: The mathematical formula of the SAF module is the formula (4) in the above technical solution, where the P2 layer performs channel compression through 3×3 convolution downsampling and 1×1 convolution, the P3 layer directly inputs, the P4 layer aligns the resolution through upsampling, and the three-way features are spliced ​​along the channel dimension to form an enhanced feature map;

[0068] S32: Implementation of the History-Aware Aggregation Unit (AAF) (Formula (5)). The AAF module fuses four-dimensional input features (shallow high / low resolution, same-layer features, and historical output) to achieve cross-frame semantic consistency improvement.

[0069] S33: RepHELAN multi-scale receptive field enhancement. This module uses convolution kernels of different scales (3×3 to 9×9) to enhance local receptive field capabilities and extract multi-scale features, enabling the model to maintain high-resolution features while also having strong semantic expression capabilities.

[0070] Furthermore, the spatially enhanced target detection method of step S4 adopts a dual-branch processing flow, using multi-level feature map input to perform bounding box regression and target classification respectively. Specifically:

[0071] S41: Implementation of the regression branch. In the regression branch, basic features are first extracted through a set of convolution operations, and then enter the spatial enhancement attention module to enhance the key area response. The enhanced features are passed through the Conv2d layer to predict the discrete boundary distribution in four directions, and are decoded into continuous bounding box positions through the DFL (distributed focal loss) module;

[0072] S42: Construction of the classification branch. The classification branch uses depthwise separable convolution and convolutional fusion to enhance the discrimination ability of each target category, and finally generates confidence maps for each target category.

[0073] S43: Processing in the inference phase. In the inference phase, the regression output is concatenated with the classification result, and the detection result is restored to the original image scale through anchor box and step size decoding.

[0074] Furthermore, the step S5-6 constructs a remote sensing image dataset, and uses the target detection model to train and test the remote sensing image dataset, and observes the target detection results, specifically:

[0075] S61: The embodiment here uses the dataset DOTA-v2.0 (2806 images / 188,282 instances, training: validation: test = 1:0.33:0.67);

[0076] S62: Use the remote sensing image target detection method and YOLOv11 to jointly build a new target detection model, set corresponding hyperparameters, and train and test the remote sensing dataset;

[0077] S63. After the training is completed, the performance evaluation and result analysis are carried out. The model performance is evaluated using the precision, recall rate and average precision mean, while the model parameter number and floating point operation number are used to evaluate the model complexity. Figure 5 and Figure 6 As shown in the figure, a comparison diagram of the detection of the model after using the S1-5 method and the basic model on the remote sensing image can be seen.

[0078] Traditional remote sensing target detection methods have inherent limitations when dealing with large-scale spans, dense occlusions, and complex backgrounds: degradation of feature representation capabilities leads to reduced detection accuracy and increased computational redundancy. In contrast, the dynamic attention multi-scale remote sensing target detection method designed in this paper utilizes a topology-aware feature compression mechanism and a cross-level adaptive fusion strategy to achieve breakthrough high-fidelity extraction of key semantic information while maintaining differential conservation of spatial details, significantly improving remote sensing image detection performance.

[0079] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without the need for creative work are still within the scope of protection of the present invention.

Claims

1. A dynamic attention multi-scale remote sensing target detection method, characterized in that: The following steps are involved: (1) Perform scale normalization and structural preprocessing on remote sensing images to generate standardized input data; (2) Extracting multi-scale semantic features through a backbone network with spatially adjustable mechanisms and channel response adjustment capabilities; (3) Introducing the above features into the constructed scale dynamic fusion module to perform cross-level and cross-receptive field information adaptation and aggregation; (4) The fused features are input into the detection head module, and the dual-path attention mechanism is used to enhance the response accuracy of the localization path and the classification path respectively; (5) Constructing task training sets for various remote sensing scenarios and adopting joint optimization strategies for end-to-end training and evaluation of detection models; (6) Deploy the trained model to a heterogeneous computing platform to achieve real-time reasoning and target detection tasks for resource-constrained terminals.

2. The method according to claim 1, characterized in that The step (2) is specifically: The backbone network integrates a spatial-channel attention submodule with a soft gating adjustment mechanism, which can adaptively allocate attention resources according to the response of the target area and enhance the saliency modeling capability.

3. The method according to claim 1, characterized in that The step (3) is specifically: The scale dynamic fusion module contains a multi-scale perception structure and a direction-adjustable convolution unit, which dynamically adjusts the fusion path according to the scale and density of targets of different categories. It has the adaptive collaborative ability to enhance semantic expression in shallow features and retain detailed information in deep features.

4. The method according to claim 1, wherein The step (4) is specifically: The detection head includes a spatial attention enhancement mechanism that strengthens boundary modeling in high-density regions through global constraints and introduces a context-aware fusion channel in the category prediction process.

5. The method according to claim 1, wherein The steps (5) and (6) are specifically: The target detection model is highly decoupled and structurally configurable, compatible with multiple mainstream detection frameworks such as YOLO and Transformer, and supports lightweight deployment in low-power computing environments. The model can adapt to multi-source remote sensing data and perform parallel computing and rapid deployment through heterogeneous hardware platforms (including but not limited to NVIDIA Jetson, Huawei Ascend, ARM CPU clusters, etc.).

6. A remote sensing image target detection method implementing any one of the methods described in claims 1-5, characterized in that: An end-to-end detection architecture consisting of multiple functional modules: It uses the data input and preprocessing modules to normalize the scale and optimize the structure of the original remote sensing images to generate a standard image tensor that is suitable for the model; the dynamic feature extraction module completes the construction of multi-level semantic information and combines it with the spatial sensitivity mechanism to enhance the expression of key areas; then, the adaptive fusion and attention guidance module selectively reconstructs the multi-scale features and aligns them with the global context to improve the cross-layer semantic consistency capability; finally, the decision and reasoning module completes the generation of the target bounding box and the category confidence estimation, and relies on a high-performance inference engine and deployment platform to achieve rapid response of the model in different remote sensing application scenarios.

Images