A target detection method for unmanned aerial vehicle aerial photograph image

Abstract

The application discloses a kind of target detection methods for unmanned aerial vehicle aerial photograph, belong to image data processing technical field, its steps are: the input unmanned aerial vehicle aerial photograph is preprocessed;The image after pre-processing is input into main network and carries out multi-stage feature extraction, and the main network is five-stage progressive feature extraction architecture, and its core component is feature extraction residual module CSP-CGLA;The feature extracted by main network is input into neck network and is fused from top to bottom, to obtain shallow layer fusion feature;Shallow layer fusion feature is progressively down-sampled and fused, and enhanced middle layer fusion feature and deep layer fusion feature are sequentially generated;Output multi-scale enhanced feature is output, for downstream visual task, significantly improve visual perception accuracy and efficiency on unmanned aerial vehicle platform.

Description

Technical Field

[0001] This invention belongs to the field of image data processing technology, specifically relating to a target detection method for drone aerial images. Background Technology

[0002] With the continuous deepening and development of Intelligent Transportation Systems (ITS), the refined management and control of complex traffic scenarios faces unprecedented challenges. To achieve efficient urban traffic management, academia and industry have proposed numerous advanced traffic management strategies in recent years, such as collaborative optimization mechanisms for traffic control and pricing schemes based on real-time traffic conditions, and tradable travel ticket management mechanisms for complex traffic scenarios. The effective operation of these advanced traffic control and dynamic pricing mechanisms heavily relies on the real-time perception and high-precision acquisition of underlying traffic operation data (such as accurate vehicle trajectories, dense congestion flow status, and micro-level individual queue lengths). However, traditional fixed-view traffic monitoring equipment has limitations such as limited field of view, large blind spots, and poor mobility. In contrast, drones, with their wide field of view and flexible maneuverability, have become a core perception method for acquiring high-precision global micro-level traffic state data.

[0003] However, traffic image analysis and processing scenarios based on UAV aerial photography have significant unique characteristics: First, small targets account for a very high proportion, with distant vehicles or pedestrians occupying only a few dozen pixels in the image, resulting in extremely sparse feature information; second, the multi-scale distribution is complex, with the scale of targets in the same traffic scene spanning dozens of times; third, ground background interference is severe, and small-scale traffic participants are easily submerged in complex road textures or building shadows; fourth, real-time requirements are high, as UAV platforms have limited computing power, and traffic control and coordination systems require extremely low latency, necessitating visual models with extremely high computational efficiency.

[0004] To meet the aforementioned perception requirements, current mainstream image feature processing and object detection methods are mainly divided into two categories: convolutional neural networks (CNNs) and end-to-end Transformer-based hierarchical feature extraction. Traditional convolutional networks suffer from limited receptive fields and insufficient long-range dependency modeling capabilities. While the end-to-end Transformer-based hierarchical feature processing architecture significantly improves feature representation capabilities by introducing a global attention mechanism, it still has the following serious shortcomings when processing small-scale target regions in complex traffic images captured by drones: First, existing backbone networks use a uniform convolutional configuration to extract multi-scale features, failing to adaptively adjust the receptive field range according to the resolution and semantic level of different feature layers. Shallow feature maps contain rich spatial details but lack global context, while deep feature maps have strong semantics but low spatial resolution. This feature extraction method results in: in the shallow stage, it is unable to effectively capture the medium-to-long-range spatial dependencies around small target vehicles and pedestrians; in the deep stage, it is limited by a fixed receptive field and finds it difficult to associate with global road network semantic information. Second, the feature fusion of the neck network adopts a fixed weighting or concatenation strategy, without considering the differences in importance of features at different scales and semantic levels under specific inputs. This static fusion method is prone to causing the detailed information of small targets to be covered by high-level abstract features when processing deep semantic features and shallow spatial features, thus reducing the fusion effect.

[0005] Third, existing networks typically build feature pyramids starting from layer P3 (1 / 8 resolution), failing to fully utilize the high-resolution features of shallower layers (such as layer P2, 1 / 4 resolution). This results in the loss of fine-grained information such as the boundaries and textures of small targets in the early stages of feature extraction, making it difficult to recover in subsequent processing and severely limiting the feature representation capabilities of small traffic participants.

[0006] Fourth, traditional stride convolution lacks constraints on high-frequency information during downsampling, introducing high-frequency aliasing and boundary blurring, which distorts the shape features of small targets. Especially in the bottom-up feature transfer path, this accumulated information loss severely affects the discriminative ability of low- and medium-resolution feature maps.

[0007] Fifth, existing methods typically employ convolutional operations with a fixed receptive field at each feature layer, making it difficult to simultaneously capture local details, mid-range spatial relationships, and global semantic information at the same stage. This singular contextual modeling approach limits the network's adaptability to targets of different scales in complex aerial traffic scenes.

[0008] Existing improvement schemes also have obvious limitations: while increasing the resolution of feature maps can retain more information, the computational load and memory consumption increase dramatically, making them unsuitable for resource-constrained UAV platforms; data augmentation methods only optimize at the training level and cannot change the expressive limitations of the network structure itself; directly adding attention modules without optimizing the computational strategy will lead to a significant decrease in inference speed; while improving the feature pyramid topology optimizes information flow, it still adopts a static fusion strategy, making it difficult to dynamically adjust the fusion weights according to the distribution characteristics of targets at different scales and the complexity of the scene.

[0009] Therefore, existing technologies suffer from problems such as difficulty in balancing efficiency and accuracy in small-scale target feature processing scenarios of UAV aerial images, insufficient adaptability of multi-scale feature fusion, and lack of effective utilization of shallow fine-grained features. These underlying visual defects directly restrict the accurate execution of upper-level mechanisms such as "traffic control and pricing scheme optimization" and "travel ticket management". There is an urgent need for an innovative image feature processing method to solve these technical defects. Summary of the Invention

[0010] To address the aforementioned problems in existing technologies, this invention proposes a target detection method for UAV aerial images. This method is rationally designed, overcomes the shortcomings of existing technologies, and achieves good results.

[0011] To achieve the above objectives, the present invention adopts the following technical solution: A target detection method for drone aerial images includes the following steps: S1: Preprocess the input drone aerial images, including resizing, pixel value normalization, and format conversion; S2: Input the preprocessed image into the backbone network for multi-stage feature extraction. The backbone network is a five-stage progressive feature extraction architecture that outputs features at five scales from shallow to deep layers. Its core component is the feature extraction residual module CSP-CGLA. The CSP-CGLA module adopts the following structure: First, the input features are divided into a first-path feature and a second-path feature in the channel dimension after channel adjustment; the first-path feature serves as a cross-stage shortcut, and the second-path feature is input into the CGLABlock processing branch; the working mechanism of CGLABlock includes, in sequence: channel decoupling multi-scale attention feature extraction, convolutional gated nonlinear transformation, and residual connection; finally, the first-path feature, the second-path feature, and the output features of the CGLABlock processing branch are concatenated and fused across stages. S3: Input the features extracted from the backbone network into the neck network for top-down feature fusion to obtain shallow fused features; S4: Perform stepwise downsampling and fusion of shallow fusion features to generate enhanced mid-layer and deep fusion features in sequence; S5: Use the multi-scale enhancement features output from steps S3 and S4 for downstream visual tasks.

[0012] Furthermore, the backbone network is a five-stage progressive feature extraction architecture, with each stage downsampling through a basic convolutional module with a stride of 2; The first stage involves downsampling by a factor of 2 to output the features of layer P1. The second stage involves 4x downsampling, followed by a feature extraction residual module CSP-CGLA, which outputs the P2 layer features. The third stage involves 8x downsampling, followed by one CSP-CGLA, to output the P3 layer features; In the fourth stage, the sample is downsampled by 16 times, and then passed through one CSP-CGLA to output the P4 layer features; In the fifth stage, the sample is downsampled by 32 times, and then passed through 3 cascaded CSP-CGLA to output the P5 layer features; The size of the feature maps output at each stage decreases sequentially, while the number of channels gradually increases.

[0013] Furthermore, let the characteristic of the first branch be denoted as The second branch characteristic is The channel decoupling multi-scale attention feature extraction is implemented using the CDMSA mechanism, specifically: ... Divided into four paths according to channel dimension Point attention, local attention, mid-range attention, and stage-adaptive global attention are applied separately, and the resulting four attention enhancement features are concatenated along the channel dimension to form a multi-scale attention feature. ; The formula for calculating point attention is: ; in, To The output features obtained after applying point attention (PA) This indicates a 1×1 convolution operation, used to expand the number of channels in the input features to 4 times; BN indicates a batch normalization operation. This represents the activation function. This represents a 1×1 convolution operation for dimensionality reduction, used to restore the number of feature channels to the original dimension; This represents element-wise multiplication. For the Sigmoid function; The formula for calculating local attention is: ; in, To The output features obtained after applying local attention (LA) This indicates that a 3×3 convolution kernel is used for the convolution operation; The formula for calculating mid-range attention is: ; in, To Output features obtained after applying mid-range attention MRA This indicates a composite downsampling operation. This represents a convolution operation in four different directions: the horizontal direction, the vertical direction, and two diagonal directions after a horizontal-vertical coordinate transformation. Indicates an upsampling operation; The phase-adaptive global attention mechanism is as follows: a downsampling convolutional attention mechanism is used for the downsampled features in the second and third phases, a downsampling multi-head self-attention mechanism is used for the downsampled features in the fourth phase, and a standard multi-head self-attention mechanism is used for the downsampled features in the fifth phase.

[0014] Furthermore, the convolutional gated nonlinear transformation of the CSP-CGLA module, through the CGLU module, applies the multi-scale attention features. Adaptive channel modulation is performed, specifically by first applying a 1×1 convolution to... The number of channels is expanded to twice the hidden dimension, and the expanded features are evenly divided along the channel dimension to obtain two sub-feature tensors. and The output features are calculated using the following formula. : ; Where GELU represents the Gaussian error linear unit activation function. This indicates a 3×3 depthwise convolution operation used to generate a gate signal; Furthermore, the residual connection and cross-stage fusion of the CSP-CGLA module are specifically as follows: Residual connection: As an identity mapping preserved, with Features after batch normalization and DropPath regularization are summed through residual connections to obtain block output features. The expression is: ; Cross-stage final fusion: combining the first splitting features Second branch characteristics and block output features Full concatenation is performed along the channel dimension, and the resulting features are fused using a 1×1 convolution. The expression is: in, This is a fusion convolution operation implemented using 1×1 convolution, used for feature fusion and dimension adjustment.

[0015] Further, in step S3, the deepest feature map P5 output by the backbone network is obtained, and single-scale global feature interaction processing is performed on it. Specifically, it is first compressed to 256 channels through 1×1 convolution, then mapped to a low-dimensional embedding space through linear projection, and then input into a multi-head self-attention module to perform global context modeling on the feature sequence to capture long-distance semantic dependencies between pixels; finally, the feature channels are adjusted through 1×1 convolution to output the top-level fusion feature Y5 rich in global semantic context, thus obtaining the high-level enhanced feature Y5 rich in global semantic information. Y5 is upsampled by a factor of two nearest neighbors, and the P4 layer features, aligned with the channel, are used as two inputs to the multi-feature modulation module CAF for fusion. CAF achieves content-aware feature fusion by calculating channel-level adaptive weights, and its mathematical expression is: ; in, For the first Channel-level weights of road features For the first Path input features, The number of input features; Weight vector Calculations are performed using a multilayer sensing network: ; in, Indicates global average pooling. and This is the weight matrix of a multilayer perceptron (MLP). For the hidden layer dimension, For the reduction rate, The number of channels for the input feature. For the hidden layer feature dimension, This is an operation that reshapes the feature vector dimension to an n×C form; Will Input three cascaded RepC3 residual enhancement modules, reduce the dimensionality to 128 channels through 1×1 convolution, and obtain the Y4 feature.

[0016] Further, in step S3, the features of layer P2 are downsampled by a GCConv unit with a stride of 2, adjusting the resolution to 1 / 8 of the original image. During the model training phase, the GCConv module adopts a bidirectional expansion structure with vertical and horizontal dimensions, including a 3×3 convolution branch, a 1×1 convolution branch, and an identity mapping branch. The 3×3 and 1×1 convolution branches both use a vertical multi-convolution structure, consisting of multiple convolutional layers connected in series with no activation function between layers. The outputs of each branch are added element-wise and then activated by SiLU. During the model inference phase, the parameters of the vertically connected convolutional kernels are first merged, and then the convolutional kernels and batch normalization parameters of the parallel branches are equivalently fused into the weights and biases of a single 3×3 convolutional layer through structural reparameterization. The Y4 feature is upsampled by a factor of 2 using bilinear interpolation. The upsampled Y4 feature, the downsampled P2 feature, and the P3 layer feature are aligned along the channel dimension and then input into the multi-feature modulation module (CAF) for fusion. The fused features are then input into the multi-scale context extraction module. In this module, the fused features are divided into two parts, denoted as... and ,right After dimensionality reduction, the input consists of four parallel branches: three dilated convolution branches with different dilation rates and one global context extraction branch, resulting in four features. These four features are concatenated along the channel dimension and then initially fused using multi-scale fusion convolution. The fused features are then combined with the features processed by skip-path convolution. The data is then concatenated again along the channel dimension, and finally fused through a convolutional process to obtain the features. The expression is: ; ; ; in, , , , and All of these are convolution operations implemented using 1×1 convolution, and are used for input feature dimensionality reduction, skip path channel alignment, final feature cross-channel integration, multi-scale feature preliminary fusion, and global branch dimensionality reduction, respectively. This indicates a dilated convolution operation, using a 3×3 convolution with a dilation rate of [value missing]. ; After deep feature extraction using three RepC3 residual blocks, shallow fusion feature F3 is obtained.

[0017] Furthermore, in S4, the bottom-up feature re-fusion uses a spatially weighted convolutional downsampling unit wConv2d, specifically by: using learnable convolutional kernel weights... With predefined two-dimensional spatial weight template Element-wise multiplication yields the spatially weighted convolution kernel weights. Downsampling is achieved by performing convolution with a stride of 2; First, the F3 feature is downsampled by a factor of 2 using a wConv2d spatially weighted convolution downsampling unit. The downsampled feature is then fused with the channel-aligned Y4 feature using a CAF module. The fused feature is then enhanced by three cascaded RepC3 residual blocks to obtain the mid-level fused feature R4. R4 is then downsampled using another wConv2d unit. The downsampled feature is then fused with the Y5 feature using a CAF module and enhanced by three cascaded RepC3 residual blocks to obtain the deep fused feature R5.

[0018] Furthermore, the backbone network, neck network, and Transformer-based decoder are constructed based on the RT-DETR target detection framework; the backbone network is used to replace the feature extraction network in the original RT-DETR framework; the top-down and bottom-up feature fusion processes in steps S3 and S4 together constitute the improved neck network; the Transformer-based decoder is the original RT-DETR decoder, outputting the bounding boxes and class prediction probabilities of each target in the UAV aerial image.

[0019] The beneficial technical effects of this invention are as follows: (1) By introducing channel decoupling multi-scale attention into the backbone network and dynamically selecting the optimal global modeling method according to the feature resolution at different network stages, the effective aggregation of multi-scale spatial features is achieved; at the same time, a convolutional gated nonlinear transformation submodule is introduced, and the aggregated features are filtered and nonlinearly enhanced through an adaptive gated modulation mechanism, which improves the expressive power of the backbone features. While fully extracting fine-grained features of small targets, the number of parameters and computational complexity are effectively controlled, achieving a balance between accuracy and efficiency. (2) By setting up a multi-feature modulation module at the key fusion node of the neck network, the content adaptive fusion of features at different scales and semantic levels was realized, which enhanced the collaborative expression and semantic consistency of cross-scale features. (3) By designing a shallow feature enhancement branch, high-resolution shallow features are introduced into the neck network for fusion after being processed by reparameterized convolution. This fully preserves fine-grained information such as the boundary and texture of small targets, and significantly improves the feature expression quality of small-scale target regions with minimal parameter cost. (4) By introducing spatially weighted convolutional downsampling units in the bottom-up path, spatial weight constraints are applied to the convolutional kernel, which reduces high-frequency aliasing and information loss during the downsampling process and improves the quality and stability of medium and low resolution feature maps. (5) By setting up a multi-scale context enhancement module and using parallel multi-scale dilated convolution groups and global context extraction branches, effective modeling of context information of different ranges such as local, mid-range and global was achieved, which enhanced the network’s ability to understand complex aerial photography scenes. (6) Compared with the original end-to-end RT-DETR target detection model, this invention effectively overcomes the shortcomings of the original model in extracting features of extremely small targets under complex ground backgrounds of UAVs, and the fact that detailed information is easily covered by high-level semantics in deep layers. The reconstructed backbone network and encoder of this invention not only significantly improve the recall rate and localization accuracy of the model for small targets of UAVs, but also perfectly supports the end-to-end characteristics of RT-DETR, maintains real-time inference capabilities, and is suitable for deployment on resource-constrained UAV platforms. Attached Figure Description

[0020] Figure 1 This is an overall flowchart of an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the CSP-CGLA module network in an embodiment of the present invention; Among them, (a) is the overall flowchart of the CSP-CGLA module; (b) is the flowchart of CGLABlock in the CSP-CGLA module; and (c) is the flowchart of CDMSA in the CSP-CGLA module. Figure 3 This is a schematic diagram of the CAF module network in an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of the GCConv module network structure in an embodiment of the present invention.

[0023] Figure 5 This is a schematic diagram of the network structure of the MSCE module in an embodiment of the present invention.

[0024] Figure 6 This is a comparison chart showing the results of UAV image detection using the method of the present invention and the detection using the RT-DETR model in an embodiment of the present invention; Wherein, (a) is the original input image and the region of interest annotation; (b) is a magnified view of the local details of the region of interest; (c) is the detection result of the RT-DETR model; and (d) is the detection result of the method of the present invention. Detailed Implementation

[0025] Before proceeding with the specific implementation, it should be noted that, in order to address the issues of low accuracy and lack of global context in small target detection using traditional convolutional networks in UAV aerial images, the optimal embodiment of this invention is based on the end-to-end target detection framework RT-DETR. This invention specifically reconstructs the backbone feature extraction network and cross-scale feature fusion encoder of the original RT-DETR to enhance the feature representation of small-scale targets, while retaining its original Transformer decoder for efficient target prediction.

[0026] The specific embodiments of the present invention will be further described below with reference to specific examples: A target detection method for drone aerial images, such as Figure 1 As shown, the process includes the following steps: S1: Preprocessing the input drone aerial images, including resizing, pixel value normalization, and format conversion; Specifically, the image is first resized to 640×640 pixels using a proportional scaling method, maintaining the original aspect ratio, and the shorter sides are filled with gray to match the target size. Then, the pixel values ​​are normalized from [0,255] to the [0,1] range. Finally, the image is converted to a tensor in batch size × number of channels × height × width (NCHW) format. This preprocessing strategy avoids image stretching and distortion, ensuring the complete preservation of the target's geometric features.

[0027] S2: Input the preprocessed image into the backbone network for multi-stage feature extraction. The backbone network is used to replace the basic network in the original RT-DETR framework. It adopts a five-stage progressive feature extraction architecture, and its core component is the feature extraction residual module CSP-CGLA. The backbone network outputs features at five scales from shallow to deep, denoted as P1 layer features to P5 layer features; Specifically, the backbone network is a five-stage progressive feature extraction architecture. Each stage downsamples through a basic convolutional module with a stride of 2. The basic convolutional module consists of convolutional layers, batch normalization layers, and SiLU activation function layers connected sequentially. During the model inference stage, the batch normalization parameters can be fused into the convolutional layers to reduce the number of computational layers. The first stage involves downsampling by a factor of 2 to output the features of layer P1. The second stage involves 4x downsampling, followed by a feature extraction residual module CSP-CGLA, which outputs P2 layer features (stage=0). The third stage involves 8x downsampling, followed by one CSP-CGLA, to output the P3 layer features (stage=1). In the fourth stage, the sample is downsampled by 16 times, and then passed through one CSP-CGLA to output the P4 layer features (stage=2). The fifth stage involves 32x downsampling, followed by three cascaded CSP-CGLA layers to output the P5 layer features (stage=3). The size of the feature maps output at each stage decreases sequentially, while the number of channels gradually increases to enhance the feature representation capability.

[0028] The CSP-CGLA module inherits from the cross-stage partial network structure, such as Figure 2 As shown in (a), the following structure is adopted: First, the input features are divided into a first branch feature and a second branch feature in the channel dimension after channel adjustment; the first branch feature serves as a cross-stage shortcut, and the second branch feature is input into the CGLABlock processing branch; as shown in (a). Figure 2As shown in (b), the working mechanism of CGLABlock includes, in sequence: channel decoupling multi-scale attention feature extraction, convolutional gated nonlinear transformation and residual connection; finally, the first branch feature, the second branch feature and the output feature of the CGLABlock processing branch are spliced ​​and fused across stages. Let the characteristic of the first branch be: The second branch characteristic is Channel decoupling and multi-scale attention feature extraction are implemented using the CDMSA mechanism, such as... Figure 2 As shown in (c), specifically: Divided into four paths according to channel dimension The number of channels per route is Point attention (PA), local attention (LA), mid-range attention (MRA), and stage-adaptive global attention (GA) are applied separately, and the resulting four attention enhancement features are concatenated along the channel dimension to form a multi-scale attention feature. ; The first path employs point attention (PA) to model the cross-channel feature relationships at a single spatial location using 1×1 convolutions. This branch first expands the number of channels to four times, activates them with SiLU, and then compresses them back to the original number of channels. The input features are then modulated using a Sigmoid gating function, calculated as follows: ; in, To The output features obtained after applying point attention (PA) This indicates a 1×1 convolution operation, used to expand the number of channels in the input features to 4 times; BN indicates a batch normalization operation. This represents the activation function. This represents a 1×1 convolution operation for dimensionality reduction, used to restore the number of feature channels to the original dimension; This represents element-wise multiplication. For the Sigmoid function; The second approach employs Local Attention (LA), using 3×3 convolutions to extract fine-grained texture and edge information from the local spatial neighborhood. The calculation formula is as follows: ; in, To The output features obtained after applying local attention (LA) This indicates that a 3×3 convolution kernel is used for the convolution operation; The third approach employs Mid-Range Attention (MRA), extracting a medium-range spatial context through multi-scale pooling and irregular convolution. This branch first performs 3×3 max pooling and 3x blur pooling downsampling on the input, then applies convolutions in the horizontal, vertical, and diagonal directions after horizontal-vertical coordinate transformation. The results in all four directions are summed and upsampled to the original resolution. The input features are then modulated using a sigmoid gate: The calculation formula is as follows: ; in, To Output features obtained after applying mid-range attention MRA This indicates a composite downsampling operation, which consists of a cascaded 3×3 max pooling and a 3x fuzzy pooling operation, used to retain salient features while resisting aliasing. This represents a convolution operation in four different directions: the horizontal direction, the vertical direction, and two diagonal directions after a horizontal-vertical coordinate transformation. This indicates an upsampling operation.

[0029] The fourth approach employs stage-adaptive global attention, dynamically selecting the implementation method based on the network's current stage. For features downsampled in the second and third stages, a downsampling convolutional attention mechanism is used: first, downsampling is performed through a max-pooling layer, and the index is recorded; then, local and spatial features are extracted via parallel convolutional branches, fused to generate weights for feature weighting; finally, the recorded index is used to upsample through a max-unpooling layer to restore resolution. For features downsampled in the fourth stage, a downsampling multi-head self-attention mechanism is used: first, downsampling is performed through a max-pooling layer, and the index is recorded; then, the data is input into a multi-head self-attention module for global modeling; finally, the recorded index is used to upsample through a max-unpooling layer to restore resolution, further enhancing semantic modeling capabilities. For features downsampled in the fifth stage, a standard multi-head self-attention mechanism is used to fully capture global dependencies.

[0030] Four attention enhancement features are spliced ​​together along the channel dimension to form multi-scale attention features: ; This stage employs a channel decoupling strategy to divide features into four parallel processing paths, enabling the network to simultaneously focus on feature patterns at four different scales: point-level, local, mid-range, and global. This provides multi-granular feature representations for small-scale target regions. The stage adaptive mechanism dynamically adjusts the calculation method of global attention based on the feature map resolution. Downsampling attention is used in shallow layers to reduce computational load, while full attention is used in deeper layers to enhance semantic modeling, achieving a dynamic balance between accuracy and efficiency.

[0031] Convolutional gated nonlinear transformation, through the CGLU (Convolutional Gated Linear Unit) module, focuses on multi-scale attention features. Adaptive channel modulation is performed, specifically by first applying a 1×1 convolution to... The number of channels is expanded to twice the hidden dimension, and the expanded features are evenly divided along the channel dimension to obtain two sub-feature tensors. and One path generates a gated signal through a 3×3 depthwise convolution and GELU activation, while the other path directly serves as the value feature. The two paths are multiplied element-wise and then compressed back to the original number of channels through a 1×1 convolution. The output feature is calculated using the following formula. : ; Where GELU represents the Gaussian error linear unit activation function. This indicates a 3×3 depthwise convolution operation used to generate a gating signal.

[0032] This stage utilizes the gating mechanism of CGLU to adaptively modulate multi-scale attention features. After extracting spatial features through depthwise separable convolution, the information flow is dynamically controlled by gating signals, achieving nonlinear enhancement and redundancy suppression of the CDMSA output. This solves the problem of insufficient expressive power of a simple linear attention mechanism. This enables the network to adaptively enhance useful features and suppress redundant information, making it particularly suitable for processing small-scale target region features in complex aerial photography scenes.

[0033] Residual connectivity and cross-stage fusion are specifically as follows: Residual connection: As an identity mapping preserved, with Features after batch normalization and DropPath regularization are summed through residual connections to obtain block output features. The expression is: ; Cross-stage final fusion: combining the first splitting features Second branch characteristics and block output features Full concatenation is performed along the channel dimension, and the resulting features are fused using a 1×1 convolution. The expression is: in, This is a fusion convolution operation implemented using 1×1 convolutions, used for feature fusion and dimension adjustment; This stage, through the dual-path design and residual connections of the cross-stage partial network, ensures smooth gradient propagation and achieves full feature fusion, thereby enhancing the network's feature extraction capability and training stability.

[0034] Through the synergistic effect of the three stages described above, the CSP-CGLA module effectively extracts fine-grained features from small-scale target regions. The improved backbone network, by repeatedly applying this module at different depths, progressively extracts multi-scale feature representations from low-level details to high-level semantics, laying the foundation for subsequent feature fusion.

[0035] Steps S3 and S4 together constitute the improvement of the original neck network of RT-DETR in this invention, forming the hybrid feature encoder of this invention.

[0036] S3: Input the features extracted from the backbone network into the neck network for top-down feature fusion, gradually combining high-level semantic features with low-level detail features; First, single-scale global feature interaction processing is performed on the P5 layer features to form the top-level fused feature Y5. Then, after upsampling Y5, it is input together with the features of layer P4 into the multi-feature modulation module CAF for fusion and enhancement to obtain the first intermediate fused feature Y4; Next, the features of layer P2 are downsampled and then input into the CAF module together with the upsampled features of layer Y4 and layer P3 for fusion. After fusion, the features are further processed by the multi-scale context enhancement module MSCE to obtain the shallow fused features F3. Specifically, the P5 layer features are first processed at the top level. In this embodiment, a single-scale global feature interaction unit based on the Transformer architecture is used to implement this processing: the P5 layer features are first compressed to 256 channels through 1×1 convolution, and then mapped to a low-dimensional embedding space through linear projection to optimize computational overhead; subsequently, a multi-head self-attention module is input to perform global context modeling on the feature sequence, capturing long-distance semantic dependencies between pixels; finally, the feature channels are adjusted through 1×1 convolution, outputting a top-level fused feature Y5 rich in global semantic context. This processing injects global contextual information into the top-level features, enhancing semantic understanding capabilities.

[0037] Next, P4 layer feature fusion is performed. Y5 is upsampled by 2 times to its nearest neighbor, and the P4 layer features after channel alignment are used as two inputs. These inputs are then fused into the Channel-Adaptive Fusion (CAF) module to obtain the final feature. ,like Figure 3 As shown, CAF achieves content-aware feature fusion by calculating channel-level adaptive weights, and its mathematical expression is: ; in, For the first Channel-level weights of road features For the first Path input features, The number of input features, This indicates element-wise multiplication in broadcast; Weight vector Calculations are performed using a multilayer sensing network: ; in, Indicates global average pooling. and This is the weight matrix of a multilayer perceptron (MLP). For the hidden layer dimension, For the reduction rate, The number of channels for the input feature. For the hidden layer feature dimension, This is an operation that reshapes the feature vector dimension to an n×C form; Softmax normalizes on the feature dimension (dim=1) to ensure that each channel The weights satisfy CAF employs channel-level weights, allowing different channels to select different feature fusion ratios based on semantic characteristics, thereby enhancing the adaptability and expressive power of feature fusion.

[0038] This embodiment is built on the RT-DETR framework, therefore it uses the RepC3 module from the original framework as the residual enhancement module. Structurally, this module employs a dual-path parallel convolution design: it includes a main branch that concatenates multiple reparameterized convolutional units (RepConv) and a 1×1 convolutional shortcut branch for feature reuse. The outputs of the two branches are fused after element-wise addition. This structure maintains the ability to extract deep features while ensuring inference speed. Three cascaded RepC3 residual enhancement modules are input and reduced to 128 channels via 1×1 convolution to obtain the Y4 feature. The CAF module adaptively adjusts the fusion weights by learning the global semantic information of the features. It can dynamically allocate the fusion ratio of each channel according to the feature content, strengthen features with high semantic relevance while weakening features with low contribution, effectively improving the quality of the fused features.

[0039] Then, shallow feature enhancement and three-way fusion are performed. For the shallow, high-resolution P2 layer features, this embodiment uses the GCConv reparameterized convolution module for enhancement processing.

[0040] Specifically, such as Figure 4As shown, firstly, the features of layer P2 undergo downsampling through GCConv units with a stride of 2, adjusting the resolution to 1 / 8 of the original image. During model training, the GCConv module is constructed as a bidirectional expansion structure, containing parallel 3×3 convolutional branches, 1×1 convolutional branches, and identity mapping branches in the horizontal direction. Each convolutional branch employs a vertical multi-convolutional structure (i.e., composed of two convolutional layers connected in series with no activation functions between layers), enriching the gradient flow and feature representation space of shallow features by increasing the linear transformation depth. The outputs of each branch are element-wise summed and then activated by SiLU. During model inference, structural re-parameterization is used. First, matrix multiplication is used to merge the vertically connected convolutional kernels, and then the convolutional kernels and batch normalization parameters of the parallel branches are equivalently fused into the weights and biases of a single 3×3 convolutional layer, thus transforming the complex training-state network into an efficient single-path inference-state structure.

[0041] This design leverages GCConv's dual expansion capabilities in depth and width during training to enhance shallow feature extraction, while achieving zero-cost inference acceleration through a single-path structure during inference, making it particularly suitable for latency-sensitive UAV edge computing.

[0042] The Y4 feature is upsampled by 2 times through bilinear interpolation. The upsampled Y4 feature, the downsampled P2 feature, and the P3 layer feature are aligned in the channel dimension and then input into the multi-feature modulation module CAF for fusion.

[0043] CAF fusion achieves an organic combination of top-down semantic information, shallow detailed information, and intermediate features, making full use of the feature advantages of different levels.

[0044] The fused feature input is processed by the Multi-Scale Context Enhancement (MSCE) module, such as... Figure 5 As shown, in the module, the fused features are divided into two parts, denoted as... and ,right After dimensionality reduction, the input has four parallel branches: three dilated convolution branches with different dilation rates and one global context extraction branch. The first branch uses a 3×3 convolution with a dilation rate of 1, and the output... The second branch uses a 3×3 convolution with a dilation rate of 2, and the output... The third branch uses a 3×3 convolution with a dilation rate of 3, and the output... The fourth branch extracts the global context through adaptive average pooling, 1×1 convolution dimensionality reduction, ReLU activation, and bilinear interpolation upsampling, and outputs... After concatenating these four features along the channel dimension, they are initially fused using multi-scale fusion convolution; the fused features are then combined with the features processed by skip-path convolution. The data is then concatenated again along the channel dimension, and finally fused through a convolutional process to obtain the features. The expression is: ; ; ; in, , , , and All of these are convolution operations implemented using 1×1 convolution, and are used for input feature dimensionality reduction, skip path channel alignment, final feature cross-channel integration, multi-scale feature preliminary fusion, and global branch dimensionality reduction, respectively. This indicates a dilated convolution operation, using a 3×3 convolution with a dilation rate of [value missing]. ; The MSCE module constructs continuous receptive field gradients through progressively dilated convolutions, enabling feature points to simultaneously capture local details and scene layout, thus correcting representational gaps caused by ambiguous deep semantic localization and missing shallow semantic details. The global context branch extracts overall semantic information through global pooling, complementing local features. Cross-stage network structures reduce computational redundancy through a flow-sharing mechanism while ensuring feature richness. This module effectively enhances the network's understanding of complex aerial scenes and provides ample contextual support for small-scale target regions.

[0045] After deep feature extraction using three RepC3 residual blocks, shallow fusion feature F3 is obtained.

[0046] S4: The shallow fusion features obtained in step S3 are downsampled and fused step by step to generate enhanced mid-layer fusion features and deep fusion features in sequence. Specifically, the bottom-up feature re-fusion uses the spatially weighted convolutional downsampling unit wConv2d. The core of wConv2d lies in imposing spatial weight constraints on the convolutional kernels, by learning the convolutional kernel weights. With predefined two-dimensional spatial weight template Element-wise multiplication yields a spatially weighted convolution kernel: ; Downsampling is achieved by performing convolution with a stride of 2: This weight template makes the convolution kernel focus more on information in the central region, while the weights at the edge positions gradually decrease. This spatial constraint reduces the aliasing effect and boundary information loss during the downsampling process, improves the quality and stability of the downsampled feature map, and plays an important role in maintaining the integrity of small-scale target region features.

[0047] First, the F3 feature is downsampled by a factor of 2 using a wConv2d spatially weighted convolution downsampling unit. The downsampled feature is then fused with the channel-aligned Y4 feature using a CAF module. The fused feature is then enhanced by three cascaded RepC3 residual blocks to obtain the mid-level fused feature R4. The R4 feature is then downsampled using another wConv2d unit. The downsampled feature is then fused with the Y5 feature using a CAF module and enhanced by three cascaded RepC3 residual blocks to obtain the deep fused feature R5.

[0048] The bottom-up approach, through spatial weighted downsampling and feature refusion, transmits shallow details to deeper layers, forming a bidirectional information flow with the top-down approach. This further enhances the expressive power of multi-scale features, enabling features at all scales to simultaneously acquire semantic and detailed information, thereby improving the overall feature quality.

[0049] After processing in steps S3 and S4, the hybrid feature encoder of this invention uses the final output F3, R4, and R5 three-scale features as multi-scale enhancement features. Among them, the F3 feature corresponds to 1 / 8 spatial resolution, preserving fine-grained boundary and texture information of small-scale target regions; the R4 feature corresponds to 1 / 16 spatial resolution, fusing mid-level semantic information and spatial details; and the R5 feature corresponds to 1 / 32 spatial resolution, containing rich high-level semantics and global context information.

[0050] The enhanced features at the three scales constitute a complete multi-scale feature pyramid optimized for UAV perspective, with the following characteristics: (1) Each scale feature has undergone channel decoupling attention enhancement and adaptive fusion processing, and has rich multi-granularity expression capabilities; (2) The shallow features fully preserve the boundary and texture details of small-scale targets in aerial photography; (3) The deep features contain complete global semantics and contextual information; (4) The features at each scale achieve information complementarity through a bidirectional fusion path.

[0051] S5: Input the above multi-scale enhanced features into the decoder based on the Transformer architecture for target prediction.

[0052] In the optimal embodiment of the present invention, it directly connects to the original Transformer decoding detection head of RT-DETR. The specific process is as follows: First, channel projection and position encoding are performed on the three-scale features of F3, R4, and R5 respectively, and after being flattened in the spatial dimension, they are spliced ​​to form a unified multi-scale feature sequence; then, the encoder selects a preset number of high-confidence feature points to initialize the target query vector, and the dimension of each query vector is set to 256.

[0053] The query vector sequence, used as the query, is input together with the aforementioned unified multi-scale feature sequence into a multi-layer deformable Transformer decoder for iterative refinement. Each decoder layer sequentially optimizes the position and category features of the UAV aerial targets through a self-attention module, a cross-attention module, and a feedforward network. Finally, the query vector output by the decoder is input to the classification head and the regression head, respectively. The classification head outputs the category probability of each target, and the regression head outputs the bounding box coordinates.

[0054] Based on the sparse prediction mechanism of the RT-DETR architecture, the detection head directly outputs the prediction results of the aforementioned 300 targets. During the training and inference phases, this invention utilizes the Hungarian algorithm for a one-to-one matching mechanism in bipartite graphs, significantly simplifying the post-processing of the detection model. Only a reasonable confidence threshold needs to be set for filtering to obtain the final detection results. This mechanism completely avoids the complex non-maximum suppression operation in traditional convolutional object detection algorithms, achieving efficient end-to-end object detection. Furthermore, it should be noted that although the optimal embodiment of this invention is described in detail based on the RT-DETR object detection framework, the improved backbone network (CSP-CGLA) and the high-quality multi-scale enhanced features output by the hybrid feature encoder proposed in this invention have good generalization capabilities. Without departing from the core design concept of this invention, those skilled in the art can use it as a general feature extraction base, applying it to other downstream visual tasks such as semantic segmentation (e.g., accessing a segmentation decoding head), instance segmentation (e.g., accessing a Mask branch), or target tracking from the perspective of UAVs by replacing different task decoding heads. This should also fall within the equivalent protection scope of this invention.

[0055] Experimental verification The experiment in this embodiment was conducted in the following environment: operating system was Windows, graphics card was NVIDIA GeForce RTX 4070, processor was Intel Core i5-13490F with 12GB of video memory, programming language was Python 3.10.18, deep learning framework was PyTorch 2.5.1, and CUDA version was 12.1.

[0056] To verify the effectiveness of the proposed multi-scale feature extraction and target detection method based on the improved RT-DETR framework, the publicly available UAV aerial image dataset VisDrone2019 was selected for experiments. This dataset contains 6471 training images, 548 validation images, and 1610 test images, covering 10 target categories: pedestrians, people, cars, vans, buses, trucks, bicycles, motorcycles, sunshade tricycles, and tricycles. This dataset features a high proportion of small targets, large scale variations, and complex backgrounds, making it an ideal choice for evaluating the performance of feature enhancement methods in UAV aerial photography scenarios.

[0057] This experiment uses the COCO evaluation system to assess the feature enhancement effect, including: (Average precision of IoU thresholds from 0.5 to 0.95) (Average accuracy when IoU threshold is 0.5) (Average accuracy of small targets, area) (pixels) (Average accuracy of medium-sized targets) area (pixels) (Average accuracy of large targets, area) The efficiency of the model was evaluated using pixels (pixels), Params, and GFLOPs. The method of this invention was compared with several mainstream object detection algorithms, and the experimental results are shown in Table 1.

[0058] Table 1. Comparison of the method of this invention with other methods on the VisDrone2019 dataset. In the object detection application verification, the feature enhancement method proposed in this invention was integrated into the RT-DETR architecture for experimentation. Experimental results show that the method of this invention achieves optimal performance across all metrics. The performance index reached 0.217, representing a 2.0 percentage point improvement over the baseline RT-DETR-R18 and a 1.4 percentage point improvement over the second-best performing YOLOv11-M. The indicator reached 0.377, an improvement of 2.5 percentage points compared to the baseline. (Regarding the smaller target indicator...) The accuracy reached 0.121, an improvement of 1.1 percentage points compared to the baseline, significantly outperforming all other comparative methods, fully validating the effectiveness of the proposed method for enhancing features in small-scale target regions. and The metrics reached 0.318 and 0.445 respectively, both optimal, indicating that the method of this invention enhances the feature representation ability of small targets while also improving the feature representation ability of medium and large targets. In terms of model efficiency, the method of this invention has only 13.765M parameters and a computational cost of only 48.4G, which are 30.8% and 15.1% lower than the baseline RT-DETR-R18, respectively. This verifies that the method of this invention maintains good computational efficiency while improving feature representation ability, making it suitable for deployment on resource-constrained UAV platforms. Figure 6 Images (a) through (d) demonstrate a visual comparison of detection in different scenarios. Compared to Figure 6 The baseline model in (c) shows obvious missed detections and localization bias. Figure 6 The method of the present invention in (d) is more accurate in detecting dense and small target areas, significantly reducing the false negative rate. This visualization result is consistent with the quantitative data in Table 1, intuitively verifying the effectiveness of the method of the present invention.

[0059] This invention constructs a backbone feature extraction network for small-scale target regions by deeply integrating a channel-decoupled multi-scale attention mechanism with convolutional gating nonlinear transformation, achieving effective extraction of multi-granular features. A stage-adaptive strategy dynamically selects the global attention implementation method at different network depths, achieving a balance between accuracy and efficiency. By introducing a shallow feature enhancement branch and designing an adaptive feature fusion path, high-resolution fine-grained information is fully utilized. A multi-scale context enhancement module is embedded in the neck network, enabling effective modeling of contextual information at different ranges. A spatially weighted convolutional downsampling strategy is introduced to mitigate information loss during downsampling. The synergistic effect of these modules results in output multi-scale enhanced features with excellent expressive power, demonstrating significant performance advantages in UAV aerial image processing tasks while maintaining low computational overhead, making it suitable for resource-constrained UAV platform deployments.

[0060] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A target detection method for drone aerial images, characterized in that, Includes the following steps: S1: Preprocess the input drone aerial images, including resizing, pixel value normalization, and format conversion; S2: Input the preprocessed image into the backbone network for multi-stage feature extraction. The backbone network is a five-stage progressive feature extraction architecture that outputs features at five scales from shallow to deep layers. Its core component is the feature extraction residual module CSP-CGLA. The CSP-CGLA module adopts the following structure: First, the input features are divided into a first branch feature and a second branch feature in the channel dimension after channel adjustment; the first branch feature serves as a cross-stage shortcut, and the second branch feature is input into the CGLABlock processing branch; The working mechanism of CGLABlock includes, in sequence: channel decoupling multi-scale attention feature extraction, convolutional gated nonlinear transformation and residual connection; finally, the first branch feature, the second branch feature and the output feature of the CGLABlock processing branch are spliced ​​and fused across stages. S3: Input the features extracted from the backbone network into the neck network to perform top-down cross-scale feature fusion to obtain shallow fused features; S4: Perform bottom-up downsampling and fusion on the shallow fusion features to generate enhanced mid-layer and deep fusion features in sequence; S5: Combine the shallow, medium and deep fused features output from steps S3 and S4 as multi-scale enhancement features, input them into the decoder based on the Transformer architecture, and output the target detection results of the UAV aerial image.

2. The target detection method for UAV aerial images according to claim 1, characterized in that, The backbone network is a five-stage progressive feature extraction architecture, with each stage using a basic convolutional module with a stride of 2 for downsampling; The first stage involves downsampling by a factor of 2 to output the features of layer P1. The second stage involves 4x downsampling, followed by a feature extraction residual module CSP-CGLA, which outputs the P2 layer features. The third stage involves 8x downsampling, followed by one CSP-CGLA, to output the P3 layer features; In the fourth stage, the sample is downsampled by 16 times, and then passed through one CSP-CGLA to output the P4 layer features; In the fifth stage, the sample is downsampled by 32 times, and then passed through 3 cascaded CSP-CGLA to output the P5 layer features; The size of the feature maps output at each stage decreases sequentially, while the number of channels gradually increases.

3. The target detection method for UAV aerial images according to claim 1, characterized in that, Let the characteristic of the first branch be: The second branch characteristic is The channel decoupling multi-scale attention feature extraction is implemented using the CDMSA mechanism, specifically: ... Divided into four paths according to channel dimension Point attention, local attention, mid-range attention, and stage-adaptive global attention are applied separately, and the resulting four attention enhancement features are concatenated along the channel dimension to form a multi-scale attention feature. ; The formula for calculating point attention is: ； in, To The output features obtained after applying point attention (PA) This indicates a 1×1 convolution operation, used to expand the number of channels in the input features to 4 times; BN indicates a batch normalization operation. This represents the activation function. This represents a 1×1 convolution operation for dimensionality reduction, used to restore the number of feature channels to the original dimension; This represents element-wise multiplication. For the Sigmoid function; The formula for calculating local attention is: ； in, To The output features obtained after applying local attention (LA) This indicates that a 3×3 convolution kernel is used for the convolution operation; The formula for calculating mid-range attention is: ； in, To Output features obtained after applying mid-range attention MRA This indicates a composite downsampling operation. This represents a convolution operation in four different directions: the horizontal direction, the vertical direction, and two diagonal directions after a horizontal-vertical coordinate transformation. Indicates an upsampling operation; The phase-adaptive global attention mechanism is as follows: a downsampling convolutional attention mechanism is used for the downsampled features in the second and third phases, a downsampling multi-head self-attention mechanism is used for the downsampled features in the fourth phase, and a standard multi-head self-attention mechanism is used for the downsampled features in the fifth phase.

4. The target detection method for UAV aerial images according to claim 3, characterized in that, The convolutional gating nonlinear transformation of the CSP-CGLA module, through the CGLU module, applies the multi-scale attention features. Adaptive channel modulation is performed, specifically by first applying a 1×1 convolution to... The number of channels is expanded to twice the hidden dimension, and the expanded features are evenly divided along the channel dimension to obtain two sub-feature tensors. and The output features are calculated using the following formula. : ； Where GELU represents the Gaussian error linear unit activation function. This indicates a 3×3 depthwise convolution operation used to generate a gating signal.

5. A target detection method for UAV aerial images according to claim 4, characterized in that, The residual connections and cross-stage fusion of the CSP-CGLA module are as follows: Residual connection: As an identity mapping preserved, with Features after batch normalization and DropPath regularization are summed through residual connections to obtain block output features. The expression is: ； Cross-stage final fusion: combining the first splitting features Second branch characteristics and block output features Full concatenation is performed along the channel dimension, and the resulting features are fused using a 1×1 convolution. The expression is: ； in, This is a fusion convolution operation implemented using 1×1 convolution, used for feature fusion and dimension adjustment.

6. The target detection method for UAV aerial images according to claim 5, characterized in that, In step S3, the deepest feature map P5 output by the backbone network is obtained, and single-scale global feature interaction processing is performed on it. Specifically, it is first compressed to 256 channels through 1×1 convolution, then mapped to a low-dimensional embedding space through linear projection, and then input into a multi-head self-attention module to perform global context modeling on the feature sequence to capture long-distance semantic dependencies between pixels; finally, the feature channels are adjusted through 1×1 convolution to output the top-level fusion feature Y5 rich in global semantic context, thus obtaining the high-level enhanced feature Y5 rich in global semantic information. Y5 is upsampled by a factor of two nearest neighbors, and the P4 layer features, aligned with the channel, are used as two inputs to the multi-feature modulation module CAF for fusion. CAF achieves content-aware feature fusion by calculating channel-level adaptive weights, and its mathematical expression is: ； in, For the first Channel-level weights of road features For the first Path input features, The number of input features; Weight vector Calculations are performed using a multilayer sensing network: ； in, Indicates global average pooling. and This is the weight matrix of a multilayer perceptron (MLP). For the hidden layer dimension, For the reduction rate, The number of channels for the input feature. For the hidden layer feature dimension, This is an operation that reshapes the feature vector dimension to an n×C form; Will Input three cascaded RepC3 residual enhancement modules, reduce the dimensionality to 128 channels through 1×1 convolution, and obtain the Y4 feature.

7. A target detection method for UAV aerial images according to claim 6, characterized in that, In step S3, the features of layer P2 are downsampled by GCConv units with a stride of 2, adjusting the resolution to 1 / 8 of the original image. During model training, the GCConv module adopts a bidirectional expansion structure with vertical and horizontal dimensions, including 3×3 convolutional branches, 1×1 convolutional branches, and identity mapping branches. The 3×3 and 1×1 convolutional branches both use vertical multi-convolutional structures, consisting of multiple convolutional layers connected in series with no activation functions between layers. The outputs of each branch are added element-wise and then activated by SiLU. During model inference, the parameters of the vertically connected convolutional kernels are first merged, and then the convolutional kernels and batch normalization parameters of the parallel branches are equivalently fused into the weights and biases of a single 3×3 convolutional layer through structural reparameterization. The Y4 feature is upsampled by a factor of 2 using bilinear interpolation. The upsampled Y4 feature, the downsampled P2 feature, and the P3 layer feature are aligned along the channel dimension and then input into the multi-feature modulation module (CAF) for fusion. The fused features are then input into the multi-scale context extraction module. In this module, the fused features are divided into two parts, denoted as... and ,right After dimensionality reduction, the input consists of four parallel branches: three dilated convolution branches with different dilation rates and one global context extraction branch, resulting in four features. These four features are concatenated along the channel dimension and then initially fused using multi-scale fusion convolution. The fused features are then combined with the features processed by skip-path convolution. The data is then concatenated again along the channel dimension, and finally fused through a convolutional process to obtain the features. The expression is: ； ； ； in, , , , and All of these are convolution operations implemented using 1×1 convolution, and are used for input feature dimensionality reduction, skip path channel alignment, final feature cross-channel integration, multi-scale feature preliminary fusion, and global branch dimensionality reduction, respectively. This indicates a dilated convolution operation, using a 3×3 convolution with a dilation rate of [value missing]. ; After deep feature extraction using three RepC3 residual blocks, shallow fusion feature F3 is obtained.

8. A target detection method for UAV aerial images according to claim 1, characterized in that, In S4, the bottom-up feature re-fusion uses the spatially weighted convolutional downsampling unit wConv2d, specifically by: [the use of] learnable convolutional kernel weights... With predefined two-dimensional spatial weight template Element-wise multiplication yields the spatially weighted convolution kernel weights. Downsampling is achieved by performing convolution with a stride of 2; First, the F3 feature is downsampled by a factor of 2 using a wConv2d spatially weighted convolution downsampling unit. The downsampled feature is then fused with the channel-aligned Y4 feature using a CAF module. The fused feature is then enhanced by three cascaded RepC3 residual blocks to obtain the mid-level fused feature R4. R4 is then downsampled using another wConv2d unit. The downsampled feature is then fused with the Y5 feature using a CAF module and enhanced by three cascaded RepC3 residual blocks to obtain the deep fused feature R5.

9. A target detection method for UAV aerial images according to claim 1, characterized in that, The backbone network, neck network, and Transformer-based decoder are constructed based on the RT-DETR object detection framework; the backbone network is used to replace the feature extraction network in the original RT-DETR framework; the top-down and bottom-up feature fusion processes in steps S3 and S4 together constitute the improved neck network. The decoder based on the Transformer architecture is the original RT-DETR decoder, which outputs the bounding boxes and class prediction probabilities of each target in the UAV aerial image.

Images