Target detection method for aerial image of unmanned aerial vehicle
By improving the neck network of the YOLO model and adopting channel splicing, cross-coupled hourglass structure and attention mechanism, the problems of small target feature loss and insufficient detection accuracy in complex backgrounds in UAV aerial images are solved, thereby improving multi-scale adaptability and detection accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTH CHINA UNIV OF TECH
- Filing Date
- 2025-12-22
- Publication Date
- 2026-04-17
AI Technical Summary
Existing YOLO models suffer from problems such as loss of subtle features of small targets and insufficient detection accuracy in complex backgrounds when processing drone aerial images, especially when they are difficult to generalize effectively under conditions of drastic scale changes.
The neck network of the YOLO model is improved by adopting a feature fusion based on channel splicing, a cross-coupled hourglass structure, and an enhanced feature fusion network based on attention mechanism to enhance the fusion capability of multi-scale features. Feature interaction is improved through recursive learning of the hourglass network and attention mechanism.
It improves the accuracy of target detection in UAV aerial images, especially the detection effect of small targets and complex backgrounds, and enhances the model's adaptability to multi-scale changes.
Smart Images
Figure CN121883804A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of target detection technology, specifically relating to a target detection method for drone aerial images. Background Technology
[0002] With their unique maneuverability and wide field of view, drones have been widely used in agricultural monitoring, urban planning, infrastructure inspection, public safety, and disaster response. The resulting large volume of aerial images has created an urgent need for efficient and accurate target detection technologies. However, the unique characteristics of drone aerial images present serious challenges to target detection tasks that are not present in traditional natural scene images.
[0003] First, drones typically operate at high altitudes, resulting in targets of interest appearing as "small targets" in the images, characterized by small pixel size, blurred appearance features, and a lack of texture information. Second, drone aerial images often feature complex backgrounds, such as urban roads, mountains, and jungles, and different weather conditions can also affect the background. Drone aerial images also contain targets with drastically different scales, such as pedestrians in the distance and vehicles in the foreground. Furthermore, variations in drone flight altitude cause the same type of target to exhibit a continuous and wide-ranging scale distribution in the image, potentially ranging from large targets in the foreground to extremely small targets in the background. This scale sensitivity necessitates that the detection model possess a certain degree of multi-scale adaptability.
[0004] YOLO models extract high-level semantic features from images through deep convolutional downsampling and effectively aggregate multi-scale contextual information using feature pyramid structures, improving the detection accuracy of targets at normal scales. However, they do not adequately address the core challenges in drone aerial imagery. Deep downsampling leads to significant loss of high-resolution details of small targets. Although feature pyramids perform geometric restoration of feature map scale through upsampling, this process cannot reconstruct the subtle features and textures lost in deep networks, resulting in weak representation capabilities for small targets. Furthermore, the feature pyramid structure's generalization ability for drastic scale changes and complex backgrounds remains insufficient. In summary, existing YOLO structures have significant shortcomings in handling the unique characteristics of dense small targets and drastically changing imaging conditions in drone images. Summary of the Invention
[0005] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and to propose a target detection method for drone aerial images.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] A target detection method for drone aerial images includes the following steps:
[0008] S1. Acquire aerial images taken by the drone;
[0009] S2. Construct an object detection model based on YOLO11. The model includes a backbone network, a neck network, and a head network connected in sequence. Improve the neck network of the object detection model to obtain an improved object detection model.
[0010] S3. Input the drone aerial images into the improved target detection model for training;
[0011] S4. Evaluate and test the trained target detection model;
[0012] S5. The trained target detection model is used for target detection in UAV aerial images. The UAV aerial images are preprocessed and then input into the target detection model to obtain the detection results.
[0013] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0014] 1. This invention makes key improvements to the neck network of the YOLO model. First, a feature fusion scheme based on channel splicing is designed to retain richer information during feature fusion. Second, a cross-coupled hourglass structure is designed. This structure utilizes the inherent downsampling and upsampling paths of the hourglass network to recursively learn shallow detail features and deep semantic features, and achieves effective fusion of multi-scale features through cross-coupling, thereby alleviating the problem of losing subtle features of small targets in deep networks. Finally, to further enhance the information interaction between different feature layers, an enhanced feature fusion network based on an attention mechanism is introduced. Through these improvements, this invention enhances the accuracy of the model in detecting drone aerial images. Attached Figure Description
[0015] Figure 1 This is a flowchart of the method of the present invention.
[0016] Figure 2 This is a schematic diagram of the structure of the improved target detection model in this invention.
[0017] Figure 3 This is a schematic diagram of the feature fusion module based on channel splicing in this invention.
[0018] Figure 4 This is a schematic diagram of the cross-coupled hourglass structure of the present invention.
[0019] Figure 5 This is a schematic diagram of the attention-based enhanced feature fusion network of the present invention. Detailed Implementation
[0020] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.
[0021] Examples; such as Figure 1 As shown, a target detection method for drone aerial images includes the following steps:
[0022] S1. Acquire aerial images taken by the drone;
[0023] S2. Construct an object detection model based on YOLO11. The model includes a backbone network, a neck network, and a head network connected in sequence. Improve the neck network of the object detection model to obtain an improved object detection model.
[0024] S3. Input the drone aerial images into the improved target detection model for training, including:
[0025] Set training parameters, including number of iterations, batch size, learning rate, optimizer, and learning rate decay strategy;
[0026] Configure data augmentation parameters, including hsv_h, hsv_s, hsv_v, flipud, fliplr, mosaic, and mixup;
[0027] The improved target detection model was trained using training parameters, data augmentation parameters, and a training set.
[0028] S4. Evaluate and test the trained target detection model;
[0029] S5. The trained target detection model is used for target detection in UAV aerial images. The UAV aerial images are preprocessed and then input into the target detection model to obtain the detection results.
[0030] In this embodiment, as Figure 2 As shown, the target detection model includes:
[0031] 1) Backbone Network: Based on an improved feature extraction network of CSPDarknet, YOLO11 uses an improved C3K2 feature extraction module, introducing multi-scale convolutional kernels to expand the receptive field. It also divides the input features into two parts, extracts features separately, and finally fuses them through 1×1 convolution. YOLO11 obtains the C2PSA module by introducing PSA (Point Spatial Attention) into the standard C2f module, enhancing feature extraction capabilities based on the attention mechanism.
[0032] 2) Neck Network: The feature fusion based on tensor addition is replaced with feature fusion based on channel concatenation to reduce information loss during feature fusion; the feature pyramid structure is replaced with a cross-coupled hourglass structure based on hourglass network, which utilizes the natural downsampling and upsampling process of hourglass network to learn the semantic information of feature map at different scales. This mode is very different from the traditional pyramid structure. It helps the model to better learn the feature representation of the target in the dimensions of scale, background and other dimensions by increasing the horizontal depth of the network; an enhanced multi-scale feature fusion module is designed. Through coupling-decoupling, the feature map of all layers of information is first obtained, and then decoupling is achieved by convolution and residual connection, so that each feature layer of the output contains both the fused features and the unique features of its corresponding scale.
[0033] 3) Detection heads: Three detection heads are used to predict the bounding box and category at three different scales.
[0034] In this embodiment, as Figure 2 As shown, the improved target detection neural network model includes the following specific components in its neck network:
[0035] The feature fusion module (XFusion) based on channel concatenation uses average pooling or linear interpolation to ensure that the input features are consistent in width and height, and concatenates the input features in channels to retain richer semantic information.
[0036] The Cross-Coupled Hourglass (C2H) module is used to learn and fuse shallow and deep features from the backbone network through its built-in recursive downsampling and upsampling paths.
[0037] Multi Single Multi Accordance (MSMA), an attention-based enhanced feature fusion network, is used to enhance the semantic information interaction between different feature layers, thereby improving the fusion capability of multi-scale features.
[0038] In this embodiment, the feature fusion module based on channel splicing includes a convolutional upsampling module (Conv UpSample, CUS), a convolutional downsampling module (Conv Down Sample, CDS), and a convolutional sampling module (ConvSample).
[0039] like Figure 3 The diagram shown is a structural diagram of the feature fusion module based on channel concatenation in this embodiment. In the diagram, (a) represents convolutional downsampling, and (b) represents convolutional upsampling. The volume theory and function of the feature fusion module based on channel concatenation are as follows:
[0040] Among the basic methods of feature map fusion, there are generally several fundamental and important strategies, including element-wise addition, element-wise addition based on weight coefficients, and channel-based concatenation. The mathematical expression for element-wise addition is as follows:
[0041]
[0042] In this diagram, A and B are feature maps of the same size, and C is the fused result. These represent the indices of the data in terms of channel, height, and width, respectively. This method requires that the feature maps involved in the fusion be completely identical in width, height, and number of channels. It assumes that the feature responses of different feature maps at the same location are complementary, and information fusion is achieved through addition. This method does not increase the number of channels and has high computational efficiency.
[0043] Element-wise addition based on weight coefficients can be viewed as an extension of element-wise addition. It introduces learnable weights to each feature map or its feature channels, and its mathematical formula is shown below:
[0044]
[0045] in, and They are learnable weight coefficients, and they exist. This relationship allows the network to autonomously decide which feature information it relies on more, enabling adaptive selection of features at different scales and thus enhancing the model's expressive power.
[0046] The channel-based fusion method first concatenates the feature maps along the channel dimension, and then uses... Convolution adjusts the number of channels. For two input tensors, the feature map dimensions are respectively... , Let these be P2 and P3; the calculation steps of the convolutional downsampling module include:
[0047] P2 first goes through a 1×1 convolutional layer to change the number of channels. Then, downsampled using average pooling. At this point, P2 processing is complete;
[0048] Concatenate P2 and P3 along the channel dimension to obtain the dimension as follows. Feature map;
[0049] Using a convolution with a kernel size of 1×1, 2×C2 input channels, and C2 output channels, P2 and P3 are fused across channels to obtain a dimension of The output feature map.
[0050] The calculation steps of the convolutional upsampling module include:
[0051] P3 first goes through a 1×1 convolutional layer to change the number of channels. Then upsampled through nearest neighbor interpolation to At this point, P3 processing is complete;
[0052] Concatenate P2 and P3 along the channel dimension to obtain the dimension as follows. Feature map;
[0053] Using a convolution with a kernel size of 1×1, 2×C1 input channels, and C1 output channels, P2 and P3 are fused across channels to obtain a dimension of The output feature map.
[0054] For two input tensors, the feature map dimension is 1. Let these be P2 and P3. The calculation steps of the convolution sampling module include:
[0055] Concatenate P2 and P3 along the channel dimension to obtain the dimension as follows. Feature map;
[0056] Using a convolution with a kernel size of 1×1, 2×C input channels, and C output channels, P2 and P3 are fused across channels to obtain a dimension of The output feature map.
[0057] This approach efficiently integrates and reduces dimensions while preserving the original feature information to the maximum extent. It only requires the feature maps to have the same width and height, and has no limit on the number of channels, thus allowing for flexible fusion of feature maps from different depths and with different numbers of channels. In convolutional downsampling and upsampling, average pooling is used for downsampling, and linear interpolation is used for upsampling, which reduces model parameters compared to directly using convolution and deconvolution.
[0058] like Figure 4 The diagram shown is a schematic of the cross-coupled hourglass module of the present invention; the specific principle and function of the cross-coupled hourglass module are as follows:
[0059] The Hourglass Network, first proposed by Alejandro Newell, Kaiyu Yang, and Jia Deng of the University of Michigan, was initially designed to solve the task of human pose estimation. Its design motivation stems from a core challenge: human pose estimation requires understanding both local details (such as hand and facial features) and global structure (such as limb connections and torso orientation). Early convolutional neural network methods often struggled to effectively handle spatial relationships at different scales. The Hourglass Network mimics how humans observe pose (first looking at the whole, then focusing on the details, and finally combining the whole for correction), designing a repetitive top-down (feature extraction, resolution reduction) and bottom-up (information propagation, resolution restoration) processing flow. It is named for its symmetrical topology, which resembles an hourglass in visualization. The computation of a single hourglass module can be viewed as a symmetrical process of encoding (downsampling) and decoding (upsampling). Let the recursion depth of the hourglass be... (For example ), then the first class( The calculations from 1 to N involve:
[0060] Encoding path (from top to bottom):
[0061] 1) Residual block processing: At each resolution level, the input features first pass through residual blocks. Processing, i.e. .in , represents the module's input. This indicates the output of the previous level;
[0062] 2) Process the features of the current level Saved for subsequent feature fusion at the same scale as the decoding path, i.e., skip connections, denoted as... ;
[0063] 3) First utilize The convolution completes the operation on The feature space is transformed, and then downsampling is performed using average pooling to prepare for the next stage. ;
[0064] Bottom bottleneck layer:
[0065] 4) When the encoding reaches the minimum resolution, i.e., the... When using residual blocks at level 1 Perform feature transformation. ;
[0066] Decoding path (from bottom to top):
[0067] 5) For The recursive depth is first determined by the residual blocks in the decoding path. Then, the lower-resolution features of the previous stage output are upsampled using nearest neighbor interpolation. ;
[0068] 6) Features of the same scale stored in the encoding path via skip connections To perform element-wise addition, that is: .
[0069] Within a single hourglass network, low-resolution layers possess a large receptive field, excelling at capturing global structure and large targets; while high-resolution layers are adept at locating details and small targets. By fusing features at different scales through skip connections, the network can simultaneously utilize local and global information for judgment. The proposed cross-coupled hourglass structure further enhances this. On one hand, it achieves multi-scale feature fusion by cross-coupling the output of the first stage of the hourglass structure through convolutional downsampling and upsampling modules; on the other hand, subsequent hourglass networks can perform higher-order feature learning based on the previously fused features. Through repeated bottom-up and top-down processing, it can continuously integrate and validate global contextual information, further enhancing the model's generalization ability in the face of background interference.
[0070] In this embodiment, the cross-coupled hourglass module implements recursive feature learning through multiple hourglass modules. Each hourglass module contains a symmetrical downsampling sequence and an upsampling sequence, and its calculation process is as follows:
[0071]
[0072]
[0073]
[0074]
[0075] in, Representing the Level downsampling blocks, including convolution and pooling; Representing the Level upsampling blocks, including nearest neighbor upsampling and convolution; for Convolution adjusts the number of channels in the corresponding layer to match the number of decoding branches; Representing the Level skip features, taken from ; Representing the Feature map after sub-fusion; For the final output, the spatial resolution is kept consistent with the input.
[0076] The cross-coupled hourglass module also includes a cross-coupled pathway for fusing multi-scale information between different hourglass modules;
[0077] Object detection models contain feature maps at multiple scales, making The representative size is Feature maps The representative size is The feature map, using the XFusion module, cross-couples the output of the first-stage hourglass module as the input of the second-stage hourglass module. The calculation process is as follows:
[0078]
[0079]
[0080] in, and This represents the output of the two hourglass modules in the first stage. and This represents the result of cross-coupling, as well as the inputs to the two hourglass modules in the second stage; This represents a convolution operation with a kernel size of 1×1, used to change the number of feature channels; This represents concatenating two tensors, specifically concatenating them along a channel; Upsample represents nearest neighbor interpolation upsampling; AvgPool represents average pooling downsampling.
[0081] In this embodiment, the calculation steps of the cross-coupled hourglass module include:
[0082] Using two hourglass modules with identical structures but different depths, each receiving feature layers of dimensions P3 and P4 as input, the resulting outputs have the same dimensions as the input data. These are represented as follows: and ;
[0083] Will and As input, XFusion modules are cross-coupled to obtain and :
[0084]
[0085]
[0086] Feature learning is performed again using two hourglass modules corresponding to L2 and L3 respectively. After the calculation is completed, two outputs of the cross-coupled hourglass module are obtained. and .
[0087] like Figure 5 The diagram shows the enhanced feature fusion network based on the attention mechanism of this invention. Figure (a) shows the fusion of three inputs, and (b) shows the feature fusion of two inputs. Although the cross-coupled hourglass module, designed based on an hourglass network and cross-coupling structure, can effectively extract and fuse semantic information at different levels within the module, its external feature maps at various scales still lack direct information interaction. To break this isolation and achieve deep fusion of global multi-scale features, an enhanced feature fusion network based on the attention mechanism is designed. Its core operation is an efficient "coupling-decoupling" mechanism:
[0088] 1) At the channel dimension, features from all levels are stitched together at the L2 level to form a fused feature that integrates full-scale information;
[0089] 2) A 1×1 convolutional layer is used to actively fuse and reduce the dimensionality of the concatenated features. The role of this convolutional layer is to dynamically learn and establish the dependencies between features at different levels, achieving effective coupling of cross-scale information;
[0090] 3) The coupled feature maps are then decoupled and restored to the scale of the original layers through specific convolutional layers. Through this process, the feature maps output by each layer are no longer isolated, but contain multi-scale contextual information from all layers that has been filtered and enhanced.
[0091] In this embodiment, the attention-based enhanced feature fusion network fuses the input features by coupling feature maps of different scales on the channels, and then decouples them through convolution.
[0092] There are feature maps at three scales, namely , and The corresponding feature map dimension is , , First, transform the data dimensions to At the corresponding level, and actively coupled on the channel:
[0093]
[0094]
[0095]
[0096]
[0097] This results in the simultaneous inclusion of , and Fusion characteristics of information at three scales Then, the decoupling phase begins:
[0098]
[0099]
[0100]
[0101] In this embodiment, the attention mechanism is specifically implemented as a CBAM module, including:
[0102] The channel attention mechanism module first performs global average pooling and global max pooling on the input feature map to obtain two different spatial context descriptors. Then, these two descriptors are fed into a shared multilayer perceptron (MLP), which consists of two fully connected layers with a ReLU activation function in between. The output feature vector of the shared MLP is summed element-wise and then processed by a sigmoid activation function to generate the final channel attention weight map. This weight map is multiplied channel-wise with the original input feature map to complete the feature relabeling in the channel dimension.
[0103] The spatial attention mechanism module takes the channel-attention-enhanced feature map as input. First, average pooling and max pooling are performed along the channel dimension to obtain two two-dimensional feature maps. These two feature maps are concatenated along the channel dimension to form a dual-channel feature descriptor. Then, a convolutional layer is used to reduce the number of channels to 1, and a sigmoid activation function is applied to generate a spatial attention weight map. This weight map is multiplied spatially with the input feature map, thereby emphasizing information-rich regions and suppressing irrelevant backgrounds in the spatial dimension.
[0104] The sequential integration and output module; the CBAM module sequentially cascades the channel attention submodule and the spatial attention submodule, that is, it first performs channel attention and then performs spatial attention; the feature map after adaptive refinement by this two-level attention mechanism is the final output of the module, which can be directly fed into subsequent network layers.
[0105] In this embodiment, the computation steps of the attention-based enhanced feature fusion network include:
[0106] use Convolutional changes The number of channels, making it consistent with The number of channels is consistent, and then the size is also made consistent through average pooling. Same, obtained Using the CBAM module for The process involves extracting spatial and channel features based on an attention mechanism to obtain... ;
[0107] use The convolution on the channel will and Coupling, to obtain ;
[0108] Will and As input to the convolutional upsampling module, the output of layer P3 is obtained. ,Will and As input to the convolutional sampling module, the output of layer P4 is obtained. .
[0109] It should also be noted that, in this specification, terms such as "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0110] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for target detection for aerial images of unmanned aerial vehicles, characterized in that, Includes the following steps: S1. Acquire aerial images taken by the drone; S2. Construct an object detection model based on YOLO11. The model includes a backbone network, a neck network, and a head network connected in sequence. The neck network of the target detection model is improved to obtain the improved target detection model; S3. Input the drone aerial images into the improved target detection model for training; S4. Evaluate and test the trained target detection model; S5. The trained target detection model is used for target detection in UAV aerial images. The UAV aerial images are preprocessed and then input into the target detection model to obtain the detection results. 2.The method of claim 1, wherein, The improved target detection neural network model includes the following neck network: The feature fusion module XFusion, which is based on channel concatenation, uses average pooling or linear interpolation to keep the input features consistent in width and height, and concatenates the input features in channels to retain richer semantic information. The cross-coupled hourglass module is used to learn and fuse shallow and deep features from the backbone network through its built-in recursive downsampling and upsampling paths. An attention-based enhanced feature fusion network is used to enhance the semantic information interaction between different feature layers, thereby improving the fusion capability of multi-scale features.
3. The target detection method for UAV aerial images according to claim 2, characterized in that, The feature fusion module XFusion, based on channel concatenation, includes a convolutional upsampling module (CUS), a convolutional downsampling module (CDS), and a convolutional sampling module (CS). For the two input tensors, the feature map dimensions are respectively... , Let these be P2 and P3; the calculation steps of the convolutional downsampling module include: P2 first goes through a 1×1 convolutional layer to change the number of channels. Then, downsampled using average pooling. At this point, P2 processing is complete; Concatenate P2 and P3 along the channel dimension to obtain the dimension as follows. Feature map; Using a convolution with a kernel size of 1×1, 2×C2 input channels, and C2 output channels, P2 and P3 are fused across channels to obtain a dimension of The output feature map; The calculation steps of the convolutional upsampling module include: P3 first goes through a 1×1 convolutional layer to change the number of channels. Then upsampled through nearest neighbor interpolation to At this point, P3 processing is complete; Concatenate P2 and P3 along the channel dimension to obtain the dimension as follows. Feature map; Using a convolution with a kernel size of 1×1, 2×C1 input channels, and C1 output channels, P2 and P3 are fused across channels to obtain a dimension of The output feature map; For two input tensors, the feature map dimension is 1. Let these be P2 and P3. The calculation steps of the convolution sampling module include: Concatenate P2 and P3 along the channel dimension to obtain the dimension as follows. Feature map; Using a convolution with a kernel size of 1×1, 2×C input channels, and C output channels, P2 and P3 are fused across channels to obtain a dimension of The output feature map.
4. The target detection method for UAV aerial images according to claim 3, characterized in that, The cross-coupled hourglass module implements recursive feature learning through multiple hourglass modules. Each hourglass module contains a symmetrical downsampling sequence and an upsampling sequence, and its calculation process is as follows: in, Representing the Level downsampling blocks, including convolution and pooling; Representing the Level upsampling blocks, including nearest neighbor upsampling and convolution; for Convolution adjusts the number of channels in the corresponding layer to match the number of decoding branches; Representing the Level skip features, taken from ; Representing the Feature map after sub-fusion; For the final output, the spatial resolution is kept consistent with the input.
5. The target detection method for UAV aerial images according to claim 4, characterized in that, The cross-coupled hourglass module also includes a cross-coupled pathway for fusing multi-scale information between different hourglass modules; Object detection models contain feature maps at multiple scales, making The representative size is Feature maps The representative size is The feature map, using the XFusion module, cross-couples the output of the first-stage hourglass module as the input of the second-stage hourglass module. The calculation process is as follows: in, and This represents the output of the two hourglass modules in the first stage. and This represents the result of cross-coupling, as well as the inputs to the two hourglass modules in the second stage; This represents a convolution operation with a kernel size of 1×1, used to change the number of feature channels; This represents concatenating two tensors, specifically concatenating them along a channel; Upsample represents nearest neighbor interpolation upsampling; AvgPool represents average pooling downsampling.
6. The target detection method for UAV aerial images according to claim 5, characterized in that, The calculation steps for the cross-coupled hourglass module include: Using two hourglass modules with identical structures but different depths, each receiving feature layers of dimensions P3 and P4 as input, the resulting outputs have the same dimensions as the input data. These are represented as follows: and ; Will and As input, XFusion modules are cross-coupled to obtain and : Feature learning is performed again using two hourglass modules corresponding to P3 and P4 respectively. After the calculation is completed, two outputs of the cross-coupled hourglass module are obtained. and .
7. A target detection method for UAV aerial images according to claim 6, characterized in that, Attention-based enhanced feature fusion networks fuse input features by coupling feature maps of different scales on the channel, and then decouple them through convolution. There are feature maps at three scales, namely , and The corresponding feature map dimension is , , First, transform the data dimensions to At the corresponding level, and actively coupled on the channel: This results in the simultaneous inclusion of , and Fusion characteristics of information at three scales Then, the decoupling phase begins:
8. The target detection method for UAV aerial images according to claim 7, characterized in that, The attention mechanism is specifically implemented as a CBAM module, including: The channel attention mechanism module first performs global average pooling and global max pooling on the input feature map to obtain two different spatial context descriptors. These two descriptors are then fed into a shared multilayer perceptron (MLP), which consists of two fully connected layers with a ReLU activation function in between. The output feature vector of the shared MLP is element-wise summed and then processed by a sigmoid activation function to generate the final channel attention weight map. This weight map is then multiplied channel-wise with the original input feature map to complete the feature relabeling along the channel dimension. The spatial attention mechanism module takes the channel-attention-enhanced feature map as input. First, average pooling and max pooling are performed along the channel dimension to obtain two two-dimensional feature maps. These two feature maps are concatenated along the channel dimension to form a dual-channel feature descriptor. Then, a convolutional layer is used to reduce the number of channels to 1, and a sigmoid activation function is applied to generate a spatial attention weight map. This weight map is multiplied spatially with the input feature map, thereby emphasizing information-rich regions and suppressing irrelevant backgrounds in the spatial dimension. Sequential integration and output module: The CBAM module sequentially cascades the channel attention submodule and the spatial attention submodule, that is, it first performs channel attention and then performs spatial attention; the feature map after adaptive refinement by this two-level attention mechanism is the output of the module, which can be directly fed into subsequent network layers.
9. A target detection method for UAV aerial images according to claim 8, characterized in that, The computational steps of an attention-based enhanced feature fusion network include: use Convolutional changes The number of channels, making it consistent with The number of channels is consistent, and then the size is also made consistent through average pooling. Same, obtained Using the CBAM module for The process involves extracting spatial and channel features based on an attention mechanism to obtain... ; use The convolution on the channel will and Coupling, to obtain ; Will and As input to the convolutional upsampling module, the output of layer P3 is obtained. ,Will and As input to the convolutional sampling module, the output of layer P4 is obtained. .
10. The target detection method for UAV aerial images according to claim 1, characterized in that, Step S3 includes the following steps: Set training parameters, including number of iterations, batch size, learning rate, optimizer, and learning rate decay strategy; Set data augmentation parameters; The improved object detection model was trained using training parameters, data augmentation parameters, and a training set.