Robust unmanned aerial vehicle detection method based on dynamic feature fusion and context attention

Abstract

The invention relates to a robust unmanned aerial vehicle detection method based on dynamic feature fusion and context attention, and belongs to the technical field of image processing. Aiming at the problems of small target feature loss, semantic gap, background noise interference and the like caused by a fixed convolution kernel scale, one-way feature fusion and a static attention mechanism in an existing unmanned aerial vehicle aerial image target detection method, the method comprises the following steps: constructing a detection model comprising a backbone network, a neck network and a detection head network; a feature rearrangement and extraction module is designed in the backbone network to enhance feature learning, an enhanced double-flow feature fusion pyramid is designed in the neck network to optimize multi-scale feature fusion, and a dynamic multi-scale context attention mechanism is designed in the detection head network to suppress irrelevant background noise. The method effectively improves the accuracy and robustness of small target detection, and achieves a clearer and more stable detection effect in a complex environment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing, and relates to a robust unmanned aerial vehicle detection method based on dynamic feature fusion and context attention. BACKGROUND

[0002] Unmanned aerial vehicle aerial image target detection is an important research direction in the fields of computer vision and remote sensing technology. With the development of deep learning technology, target detection methods based on convolutional neural networks have made significant progress. Among them, the YOLO model and similar architectures exhibit good performance in small target detection accuracy through multi-scale feature extraction and the application of attention mechanisms. Attention mechanisms such as channel attention and spatial attention can guide the model to focus on key image regions, thereby enhancing the saliency of small target features.

[0003] However, existing methods still have several key defects. Traditional convolution operations rely on fixed-size convolution kernels and cannot dynamically adapt to the varying size of small targets in unmanned aerial vehicle images, resulting in the loss of fine-grained information such as texture and contour during extraction. The standard Feature Pyramid Network (FPN) as a common feature fusion structure can integrate multi-scale information, but its one-way fusion method limits the full interaction of shallow detail features and deep semantic features, causing insufficient fusion of local details and high-level semantic information. In addition, most attention mechanisms use static designs and cannot effectively suppress complex background noise in unmanned aerial vehicle aerial images, thereby affecting the detection clarity of small targets in cluttered environments.

[0004] These problems jointly constrain the accuracy and robustness of small target detection in unmanned aerial vehicle images, and there is an urgent need for a new method that can dynamically adapt to scale changes, optimize feature fusion, and enhance context attention. SUMMARY

[0005] Therefore, the purpose of the present application is to provide a robust unmanned aerial vehicle detection method based on dynamic feature fusion and context attention.

[0006] To achieve the above purpose, the present application provides the following technical solutions: A robust unmanned aerial vehicle detection method based on dynamic feature fusion and context attention, comprising the following steps: Step 1: Constructing an unmanned aerial vehicle remote sensing image data training set; Step 2: Constructing an unmanned aerial vehicle remote sensing small target detection network model, the network model comprising a backbone network, a detection neck network, and a detection head network; Step three, design a feature rearrangement and extraction module (FREM) in the backbone network to enhance feature learning through spatial-to-depth conversion; Step four, design an enhanced dual-flow feature fusion pyramid (EDFFP) in the neck network to optimize the multi-scale feature fusion process; Step five, design a dynamic multi-scale context attention mechanism (DMCA) in the head network to suppress irrelevant background noise; Step six, implement detection on the unmanned aerial vehicle remote sensing dataset.

[0007] Further, in step one, constructing the unmanned aerial vehicle remote sensing image data training set specifically includes: using the VISDRONE dataset and the VEDAI dataset for multi-scale training, setting the training batch size to 300, the image training size from 640 to 512, and applying a data enhancement strategy of mixing, label smoothing, and random affine transformation; the image size is 640 when testing.

[0008] Further, the processing process of the feature rearrangement and extraction module FREM includes: inputting the feature map obtaining the output feature map through convolution operation The mathematical expression is:

[0009]

[0010] wherein, represents a standard convolution operation with a kernel size of 1x1, represents a standard convolution operation with a kernel size of 5x5, is a feature map splicing operation, is a spatial-to-depth rearrangement operation, represents a block size and is set to 2, refers to the C2f-MSRM module; For an input feature map dimension of , , , , the output feature map dimension is , , , , wherein is the batch size, is the number of channels, and respectively represent the height and width of the input feature map.

[0011] Further, the processing procedure of the C2f-MSRM module includes: inputting the feature map obtaining the output feature map through convolution and splicing operations , and the mathematical expression is:

[0012] wherein, MSRM represents a multi-scale residual module; the processing procedure of the MSRM module includes: inputting the feature map obtaining the output feature map through multi-branch convolution and attention mechanism , and the mathematical expression is:

[0013] wherein, , . , , respectively represent depth separable convolution operations with kernel sizes of 3x3, 5x5 and 7x7, SE is an SE attention mechanism.

[0014] Further, the enhanced dual-flow feature fusion pyramid (EDFFP) includes a top-down path and a bottom-up path; the mathematical expression of the processing procedure of the top-down path is:

[0015]

[0016]

[0017] the mathematical expression of the processing procedure of the bottom-up path is:

[0018]

[0019]

[0020] wherein, , , , is the input feature map, CARAFE represents a CARAFE up-sampling operator, represents a standard convolution operation with a kernel size of 3x3 and a stride of 2, indicates a C2f-MSRM module.

[0021] Further, the processing procedure of the dynamic multi-scale context attention mechanism DMCA comprises: grouped feature map obtained by double-branch processing , and the mathematical expression is:

[0022]

[0023]

[0024] wherein the first branch calculates: , , and is decomposed into and ; the second branch calculates: ; the fusion weight is calculated as: , , ; the parameters are defined as: represents a grouped normalization operation, is a Sigmoid function, represents an element-wise multiplication operation, is a Softmax function, is an adaptive global average pooling operation, is a shape reshaping operation, , , represents a dynamic channel weighting coefficient.

[0025] Further, in step six, when implementing detection on the unmanned aerial vehicle remote sensing dataset, the VISDRONE dataset and the VEDAI dataset are used for evaluation, and the detection indicators include the average precision intersection over union threshold of 50.

[0026] Further, the backbone network, the detection neck network and the detection head network are constructed based on the YOLOv8n framework, the feature rearrangement and extraction module FREM replaces the traditional convolution layer, the enhanced dual-flow feature fusion pyramid EDFFP replaces the traditional feature pyramid network, and the dynamic multi-scale context attention mechanism DMCA is integrated in the detection head network.

[0027] Further, the CARAFE upsampling operator is used to aggregate context information and replace the traditional upsampling method.

[0028] Further, the dynamic channel weighting coefficient 、 and Adaptive allocation is learned by a model to optimize multi-scale feature fusion.

[0029] The beneficial effects of the present application are: First, at the feature extraction level, the present application effectively overcomes the limitations of the fixed size convolution kernel of the traditional convolution operation. Through the unique design of the feature rearrangement and extraction module, efficient capture of small target fine-grained features is realized, significantly improving the perception ability of micro texture and contour features. This module realizes intelligent conversion from space to depth, while maintaining feature richness and expanding the network receptive field, so that the model can better adapt to the challenge of small target scale variability.

[0030] Secondly, in terms of feature fusion, the present application completely improves the deficiency of one-way fusion of the traditional feature pyramid network. The enhanced double-flow feature fusion pyramid realizes the full fusion of shallow detail features and deep semantic features through bidirectional interaction path and content-aware up-sampling mechanism. This fusion method not only eliminates the semantic gap, but also significantly enhances the model's representation ability for multi-scale targets, making the feature expression of small targets in complex background more complete and clear.

[0031] In the noise suppression link, the dynamic multi-scale context attention mechanism proposed by the present application breaks through the limitations of traditional static attention mechanism. This mechanism can intelligently distinguish target features and background interference through multi-branch convolution structure and dynamic weight distribution strategy, effectively suppressing irrelevant noise in complex environments. This dynamic adaptability makes the model show stronger robustness when facing diversified aerial scenes, significantly improving the discrimination of small targets and background.

[0032] In addition, the innovative design of the overall architecture of the present application brings comprehensive improvement of detection performance. The model maintains high inference efficiency while showing more stable detection consistency, and is more accurate and reliable in identifying various small targets. Especially in complex scenes with target dense overlap, the present application can effectively reduce the missed detection phenomenon and improve the positioning accuracy.

[0033] Other advantages, objects and features of the present application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following or can be learned by practice of the present application. The objects and other advantages of the present application can be realized and attained by the structure particularly pointed out in the description below. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to make the purpose, technical scheme and advantages of the present application clearer, the preferred detailed description of the present application will be given below in combination with the drawings, in which: Figure 1 The overall framework diagram of the unmanned aerial vehicle remote sensing small target detection network model proposed in the application is shown in the figure; Figure 2 The feature rearrangement and extraction module FREM proposed in the application is shown in the figure; Figure 3 The C2f-MSRM module and the MSRM module proposed in the application are shown in the figure; Figure 4 The enhanced double-flow feature fusion pyramid EDFFP proposed in the application is shown in the figure; Figure 5 The dynamic multi-scale context attention mechanism DMCA proposed in the application is shown in the figure; Figure 6 The comparison results of detection on the VISDRONE dataset with other detectors are shown in the figure; Figure 7 The comparison results of detection on the VEDAI dataset with other detectors are shown in the figure. DETAILED DESCRIPTION

[0035] The embodiments of the application are described below through specific concrete examples, and other advantages and effects of the application can be easily understood by those skilled in the art from the disclosure of the specification. The application can also be implemented or applied through other different specific embodiments, and various modifications or changes can be made to the details in the specification based on different views and applications without departing from the spirit of the application. It should be noted that the drawings provided in the following examples only illustrate the basic concept of the application in a schematic manner, and the following examples and features in the examples can be combined with each other without conflict.

[0036] The drawings are only used for illustrative description, and the representation is only a schematic diagram, not a physical diagram, and cannot be understood as a limitation on the application; in order to better illustrate the embodiments of the application, some components in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; it is understandable for those skilled in the art that some known structures and their descriptions in the drawings may be omitted.

[0037] The same or similar reference numerals in the drawings of the embodiments of the application correspond to the same or similar components; in the description of the application, it should be understood that if the terms "upper", "lower", "left", "right", "front", "back" and the like indicate the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore the terms describing the positional relationship in the drawings are only used for illustrative description, and cannot be understood as a limitation on the application, and for those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0038] I. Constructing a UAV remote sensing image data training set to facilitate effective learning of the model: The trained image data set adopts VISDRONE and VEDAI data sets, the training batch size of each iteration is set to 300, multi-scale training is carried out on the two data sets, including image sizes from 640 to 512. In order to enhance the representation ability of the model under limited instances, a combination of mixing, label smoothing and random affine transformation is adopted. The size of the image test is 640.

[0039] II. Constructing a UAV remote sensing small target detection network model, including a backbone network, a detection neck network and a detection head network: The convolutional neural network designed in the application is as shown in Figure 1 The input image is first extracted by the backbone network to obtain multi-scale features. In the network, the outputs of the 3rd, 4th and 5th stages are represented as P3, P4 and P5 respectively. These outputs can be used as the basis for subsequent feature fusion. In the feature pyramid of the detection neck network, the feature maps in P3, P4 and P5 are used as inputs to generate output feature maps in the pyramid structure for subsequent processing by the detection head network. Figure 1 In the network, the Feature Rearrangement and Extraction Module (FREM) is designed to enhance small target feature learning. In addition, the Enhanced Dual-Flow Feature Fusion Pyramid (EDFFP) is used to optimize the multi-scale feature fusion process. The Dynamic Multi-scale Context Attention (DMCA) can effectively suppress irrelevant background noise. P3, P4 and P5 represent the features extracted by the backbone network for feature fusion.

[0040] III. Designing a feature rearrangement and extraction module (FREM) in the backbone network to enhance feature learning through spatial-to-depth conversion: Small targets in UAV images have a low pixel ratio, which can easily lead to the loss of key features such as texture and contour. At the same time, the feature extraction capability of existing backbone networks is limited, and the extracted features often lack semantic information and have limited receptive fields, making it difficult to achieve accurate detection of small targets. To solve this problem, the application proposes a feature rearrangement and extraction module (FREM) as the core module for small target detection. This module uses lightweight convolution, spatial-to-depth conversion and a designed bottleneck structure to extract rich small target feature information and expand the receptive field of the network.

[0041] The overall structure of the FREM is as shown in Figure 2The feature extraction process is described as follows: first, the input feature map is extracted by a lightweight module; then, the extracted features are rearranged from the spatial dimension to the depth dimension; finally, to enhance the multi-scale feature extraction capability, the rearranged feature map is processed by the C2f-MSRM module to generate the final output feature map. This process can be represented as:

[0042]

[0043] wherein, and represent the standard convolution operation with kernel size and , respectively. is the output feature map of convolution, is the output feature map of convolution, is the input feature map. represents the output feature map of the GhostConv module, is the feature map splicing operation. is the final output feature map. is the spatial-to-depth rearrangement operation, which increases the number of channels and reduces the spatial dimension to improve small target detection capability. represents the block size of the module, which is set to 2 in this paper. For an input feature map with dimensions , , , , the output feature map has dimensions , , , , where is the batch size, is the number of channels, and are the height and width of the input feature map, respectively. is the C2f-MSRM module.

[0044] To enhance the nonlinear representation capability of the model and solve the problem of limited receptive field in small target detection of traditional convolutional networks, a multi-scale residual structure is introduced in the C2f module. As shown in Figure 3 , the C2f-MSRM module combines the C2f module with the MSRM module. In this structure, the original bottleneck module is replaced by the MSRM module, and three MSRM modules are used to extract more rich gradient information, thereby better processing complex image features. Wherein represents the MSRM module. The specific process is as follows:

[0045] As shown in FIG. 1, the multi-scale residual structure process of MSRM is as follows: the input feature map is divided into a residual branch and a backbone branch. In the backbone branch, first, a convolution operation is performed to change the feature map size; then the branch is divided into three depth separable convolution branches: the first branch uses a depth separable convolution for feature extraction, the second and third branches use depth separable convolutions respectively; next, the outputs of the three branches are spliced along the channel dimension, and dimension reduction is realized through a convolution module; then, the output feature maps of the backbone branch and the residual branch are spliced along the channel dimension; finally, the SE attention mechanism is used to enhance the feature representation capability. Among them,,, and respectively represent the depth separable convolution operation with the kernel size of,,, is the feature map after the convolution, is the fused feature map. The core multi-scale feature fusion process can be represented as: Figure 3

[0046]

[0047] The mathematical expression of MSRM can be represented as:

[0048] Among them, is the SE attention mechanism.

[0049] Four, an enhanced double-flow feature fusion pyramid (EDFFP) is designed in the neck network to optimize the multi-scale feature fusion process and solve the inherent information loss problem of the traditional neck network: ​​​​​​​​​​​​​​​​The feature maps extracted by backbone network contain different levels of information: shallow features have high resolution and rich details but weak semantics, while deep features have strong semantics but low resolution. PANet, as a bidirectional feature fusion structure, transmits deep semantic information to shallow layers through a top-down path and transmits shallow detail information to deep layers through a bottom-up path, making the feature representation more abundant. However, relying on deep features for fusion may lead to the loss of small target information when the deep features are insufficient. In addition, although the bidirectional information flow improves the interaction of hierarchical features, relying solely on simple additive fusion may still produce a semantic gap. To address these challenges, this paper proposes a new neck structure EDFFP, which is shown in Figure 4 .

[0050] In EDFFP, the CARAFE up-sampling operator replaces the traditional up-sampling method, which can aggregate contextual information and capture more rich semantic information. At the same time, EDFFP realizes the comprehensive integration of multi-scale feature information, enhancing the fusion effect of shallow and deep feature information. In addition, this structure replaces the original C2f module with the C2f-MSRM module, further improving the multi-scale feature extraction capability. The above improvements effectively solve the problems of semantic gap and small target information loss in UAV images.

[0051] The top-down feature information fusion process is as follows: the input feature map includes ( ), ( ), ( ), and ( ). First, , the CARAFE operator (denoted as ) is used for up-sampling, and the obtained feature map is combined with after concatenation (Concat) operation, and then processed by C2f-MSRM module to generate . Next, , the CARAFE operator is used for up-sampling to the same size as , and then processed by concatenation and C2f-MSRM module to obtain . Repeat the same process to generate . The corresponding calculation process can be summarized as:

[0052]

[0053]

[0054] The bottom-up feature fusion process is similar to the top-down approach, the main difference being the use of a step size of 2. Convolution performs downsampling. First, After downsampling, the obtained feature map and and The data is merged through a splicing operation and then processed by the C2f-MSRM module to generate the final product. The process is in Repeat the above to generate and The output of the EDFFP structure , and It will be transmitted to the DMCA module to extract channel and spatial information. The calculation process can be described as follows:

[0055]

[0056]

[0057] Compared to PANet, EDFFP optimizes the fusion strategy of shallow and deep feature information and enhances the fusion effect of multi-scale semantic features of small objects. This improvement provides a more effective feature map input for subsequent attention mechanisms.

[0058] 5. A dynamic multi-scale contextual attention mechanism (DMCA) is designed in the detection head network to effectively suppress irrelevant background noise, thereby significantly improving detection accuracy in complex and cluttered environments: After the backbone network and neck structure complete feature extraction of the small target, the small target is sufficiently represented in the feature map. Compared with the backbone network, modeling the global contextual relationship between the small target and the background can suppress invalid background interference and enhance the feature level's ability to distinguish between the small target and the background. Traditional attention mechanisms can only capture local information, resulting in insufficient contextual relevance. Inspired by EMA, this paper proposes the DMCA module, which introduces a multi-branch convolutional structure and integrates a dynamic channel weight mechanism. Figure 5 As shown, DMCA first groups the input feature maps, and then, in order to fuse multi-scale spatial information, it combines the grouped feature maps... Two parallel branches are fed in to simultaneously capture local and global feature representations. In the first branch, the module captures global information along the vertical dimension. Capture along the horizontal dimension Then, information from the two dimensions is integrated and passed through Convolution processing yields feature maps .Will Decomposed into two parallel one-dimensional feature encoding vectors and After nonlinear Sigmoid function processing, the attention weight of each dimension is obtained. Finally, the module calculates the fusion feature and records it as The specific calculation process is as follows:

[0059]

[0060]

[0061] wherein denotes the group normalization operation, is the Sigmoid function, denotes the element-wise multiplication operation.

[0062] In the multi-scale convolution branch, convolution kernels with sizes , and are used to extract features with different receptive fields. The model assigns weights to the feature maps output by each convolution branch through dynamic channel coefficients. This process can be described as:

[0063]

[0064] wherein , and denote the dynamic channel weighting coefficients, is the output of the multi-scale convolution branch.

[0065] Finally, the model calculates the fusion weight using the output feature maps . This weight then guides the group feature map to restore its original shape. The specific calculation is as follows:

[0066]

[0067]

[0068]

[0069] wherein denotes the Softmax function, is the adaptive global average pooling operation, represents the shape reshaping operation on the feature map.

[0070] Six. Implementation of detection on unmanned aerial vehicle remote sensing dataset: ​Based on the method, experiments are performed on the challenging VISDRONE and VEDAI datasets to evaluate the effectiveness of the detection method proposed in the application and different detection networks on the VISDRONE and VEDAI datasets. As shown in Tables 1 and 2, the method of the application obtains excellent detection performance.

[0071] Table 1 Detection results of different detectors on VISDRONE

[0072] Table 2 Detection results of different detectors on VEDAI

[0073] Figure 6 For VISDRONE dataset, the comparison results of detection with other detectors are as follows: Figure 7 For VEDAI dataset, the comparison results of detection with other detectors are as follows.

[0074] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the application and are not limiting. Although the application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the application can be modified or replaced equivalently without departing from the purpose and scope of the technical solutions, and they should be covered in the scope of the claims of the application.

Claims

1. A robust UAV detection method based on dynamic feature fusion and contextual attention, characterized in that: Includes the following steps: Step 1: Construct a training set of UAV remote sensing image data; Step 2: Construct a UAV remote sensing small target detection network model, which includes a backbone network, a detection neck network, and a detection head network; Step 3: Design the Feature Rearrangement and Extraction (FREM) module in the backbone network to enhance feature learning through spatial-to-depth transformation; Step 4: In the detection neck network, an enhanced dual-stream feature fusion pyramid (EDFFP) is designed to optimize the multi-scale feature fusion process. Step 5: Design a dynamic multi-scale contextual attention mechanism (DMCA) in the detection head network to suppress irrelevant background noise; Step 6: Perform detection on the UAV remote sensing dataset.

2. The robust UAV detection method based on dynamic feature fusion and contextual attention according to claim 1, characterized in that: In step one, constructing the UAV remote sensing image data training set specifically includes: using the VISDRONE dataset and the VEDAI dataset for multi-scale training, setting the training batch size to 300, and the image training size from 640 to 512, and applying data augmentation strategies such as mixing, label smoothing, and random affine transformation; the image size during testing is 640.

3. The robust UAV detection method based on dynamic feature fusion and contextual attention according to claim 1, characterized in that: The processing steps of the Feature Rearrangement and Extraction (FREM) module include: inputting feature maps. The output feature map is obtained through convolution operation. Its mathematical expression is: in, This represents a standard convolution operation with a kernel size of 1×1. This represents a standard convolution operation with a kernel size of 5×5. This is a feature map stitching operation. This refers to the space-to-depth rearrangement operation. This represents the block size and is set to 2. Refers to the C2f-MSRM module; For the input feature map of dimension ( , , , The output feature map has a dimension of (). , , , ),in For batch size, For the number of channels, and These represent the height and width of the input feature map, respectively.

4. The robust UAV detection method based on dynamic feature fusion and contextual attention according to claim 3, characterized in that: The processing steps of the C2f-MSRM module include: inputting feature maps. The output feature map is obtained through convolution and concatenation operations. Its mathematical expression is: in, This represents the Multiscale Residual Module (MSRM). The processing procedure of the MSRM module includes: input feature map The output feature map is obtained through multi-branch convolution and attention mechanism. Its mathematical expression is: in, , . , , These represent depthwise separable convolution operations with kernel sizes of 3×3, 5×5, and 7×7, respectively. It is the SE attention mechanism.

5. The robust UAV detection method based on dynamic feature fusion and contextual attention according to claim 1, characterized in that: The Enhanced Dual-Stream Feature Fusion Pyramid (EDFFP) includes a top-down path and a bottom-up path; The mathematical expression for the top-down path processing is: The mathematical expression for the bottom-up path processing is: in, , , , For the input feature map, This indicates the CARAFE upsampling operator. This represents a standard convolution operation with a kernel size of 3×3 and a stride of 2. Refers to the C2f-MSRM module.

6. The robust UAV detection method based on dynamic feature fusion and contextual attention according to claim 1, characterized in that: The processing steps of the Dynamic Multi-Scale Contextual Attention (DMCA) mechanism include: Grouped feature maps The output feature map is obtained through a two-branch process. Its mathematical expression is: The first branch is calculated as follows: , and decomposed into and Second branch calculation: ; Fusion weights calculate: , , Parameter definition: This indicates a grouping normalization operation. For the Sigmoid function, This represents element-wise multiplication. This represents the Softmax function. To enable adaptive global average pooling, For shape reshaping operations, , , This represents the dynamic channel weighting coefficient.

7. The robust UAV detection method based on dynamic feature fusion and contextual attention according to claim 1, characterized in that: In step six, when performing detection on the UAV remote sensing dataset, the VISDRONE dataset and VEDAI dataset are used for evaluation, and the detection index includes the mean value when the average precision cross-union ratio threshold is 50.

8. The robust UAV detection method based on dynamic feature fusion and contextual attention according to claim 1, characterized in that: The backbone network, detection neck network, and detection head network are constructed based on the YOLOv8n framework. The feature rearrangement and extraction module FREM replaces the traditional convolutional layer, the enhanced dual-stream feature fusion pyramid EDFFP replaces the traditional feature pyramid network, and the dynamic multi-scale contextual attention mechanism DMCA is integrated into the detection head network.

9. The robust UAV detection method based on dynamic feature fusion and contextual attention according to claim 5, characterized in that: The CARAFE upsampling operator is used to aggregate contextual information, replacing traditional upsampling methods.

10. The robust UAV detection method based on dynamic feature fusion and contextual attention according to claim 6, characterized in that: The dynamic channel weighting coefficient , and The model learns adaptive allocation to optimize multi-scale feature fusion.

Images