A UAV Aerial Target Detection Method Based on Improved YOLOv11n

Abstract

This invention discloses an improved UAV aerial target detection method based on YOLOv11n. This method addresses the problems of low accuracy and high false negative rate in detecting small targets in UAV aerial images. The method includes: inputting the image to be detected into a backbone network containing a dual contextual attention downsampling module for feature extraction; during downsampling, DCAD extracts local detail features and surrounding contextual features in parallel and reweights them through channel attention; inputting multi-level feature maps into a global context path aggregation feature pyramid network for cross-scale fusion; GC-PAFPN adds a direct fusion path across two scales and uses a cross-stage omnidirectional kernel module to capture multi-scale features through parallel multi-morphological convolutional kernels and frequency domain global context modeling; inputting the fused feature maps into a shape and semantic alignment decoupling head; using dynamic deformable convolution to achieve shape alignment and an object probabilistic graph attention mechanism to achieve semantic alignment; and modulating classification confidence with localization quality to output the target category and bounding box.

Description

Technical Field

[0001] This invention relates to a target detection method for UAV aerial photography based on an improved YOLOv11n. Background Technology

[0002] Object detection is a fundamental task in computer vision, aiming to identify the category of objects in an image or video and locate their bounding boxes. In recent years, single-stage detectors, represented by the YOLO (You Only Look Once) series, have been widely used due to their excellent balance between speed and accuracy. However, despite significant progress, small object detection (typically referring to objects that occupy a small area in an image, such as those with a pixel area smaller than 32×32) remains a prominent and persistent technical challenge in this field. Based on advanced models such as YOLO11n, existing techniques still expose the following key bottlenecks when handling small objects:

[0003] 1) Irreversible Information Loss Due to Downsampling: High-resolution shallow feature maps are crucial for small object detection. However, to build deep networks and expand the receptive field, the model must perform multiple downsampling operations. Traditional downsampling operations (such as convolutions or pooling with a stride of 2) are prone to causing irreversible loss of key textures and spatial details of small objects when reducing the feature map size. This results in the signal of small objects becoming extremely weak, or even completely "disappearing," when passed to deep networks.

[0004] 2) Insufficient discriminative power of local receptive fields for small targets: Small targets themselves lack sufficient internal feature information (e.g., only a few pixels), making them easily confused with background noise, texture artifacts, or local parts of other objects. Therefore, accurate classification of small targets heavily relies on their surrounding global scene context (e.g., a "small dot" on the water's surface is likely a boat, and a "small dot" in the sky is likely a bird). Existing models mainly rely on stacking local convolutions (such as 3×3) to perceive the context. This approach is difficult to efficiently and directly establish long-distance dependencies and cannot fully utilize global contextual information to assist in the discrimination of ambiguous small targets, resulting in a large number of false positives and false negatives.

[0005] 3) The task misalignment problem is drastically amplified at the scale of small objects: Object detection comprises two sub-tasks: classification and regression. In standard detection heads, these two tasks typically predict on the same feature point. This design is fatal for small objects because they are extremely sensitive to spatial location. The classification task (identifying the core) and the regression task (localizing the boundary) share the same feature sampling point, making them highly sensitive to spatial location at a scale of a few pixels. The coarse, static alignment methods of traditional detection heads (such as those based on preset anchor points or grid centers) easily cause the feature sampling point to deviate from the small target by only a few pixels, thus sampling background information, leading to classification failure and severe localization drift (a one-pixel offset can cause an IoU drop of more than 50%). The existing task misalignment problem is drastically amplified under the high accuracy requirements of small object detection.

[0006] Therefore, designing a target detection network that can retain key details during downsampling, efficiently integrate global context, and achieve pixel-level precise alignment is a core technical challenge that urgently needs to be addressed to overcome the current performance bottleneck in small target detection. Summary of the Invention

[0007] This invention provides an improved UAV aerial target detection method based on YOLOv11n to address the problems existing in the prior art.

[0008] The technical solutions adopted in this invention are as follows:

[0009] A method for drone aerial target detection based on an improved YOLOv11n includes the following steps:

[0010] S1: The image to be detected is input into a backbone network containing a dual contextual attention downsampling module for feature extraction to obtain multi-level feature maps; the dual contextual attention downsampling module first downsamples the input features through convolution to obtain basic features, and then inputs the basic features into local branches and peripheral branches respectively to extract local detail features and peripheral context features in parallel. After fusing the local detail features and the peripheral context features, global context reweighting is performed through a channel attention mechanism;

[0011] S2: Input the multi-level feature map into the global context path aggregation feature pyramid network for cross-scale feature fusion to obtain a fused feature map; the global context path aggregation feature pyramid network adds a direct fusion path across two scales on the basis of adjacent scale fusion, and adopts a cross-stage omnidirectional kernel module at the fusion node. The cross-stage omnidirectional kernel module captures multi-scale features through parallel multi-morphological convolution kernels and frequency domain global context modeling.

[0012] S3: Input the fused feature map into the shape and semantic alignment decoupling head for target detection, and output the target category and bounding box; the shape and semantic alignment decoupling head includes a classification branch and a regression branch. The regression branch achieves shape alignment and outputs localization quality through dynamic deformable convolution. The classification branch achieves semantic alignment and outputs category score through object probabilistic graph attention mechanism. The category score is modulated using the localization quality to obtain the final classification confidence.

[0013] Furthermore, in S1, the dual contextual attention downsampling module downsamples the input features using a 3×3 convolution with a stride of 2.

[0014] Furthermore, in S1, the local branches use 3×3 depth convolution, and the peripheral branches use 3×3 dilated depth convolution to expand the receptive field.

[0015] Furthermore, in S1, the channel attention mechanism includes: generating channel descriptors through global average pooling, learning the channel dependencies through two fully connected layers, and obtaining channel weights through the Sigmoid function for reweighting.

[0016] Furthermore, in S2, the cross-stage omnidirectional kernel module divides the input features into a direct-connection preservation branch and a transformation enhancement branch. The transformation enhancement branch is enhanced by the parallel multimorphic convolution kernel and frequency domain global context modeling, and then fused with the direct-connection preservation branch.

[0017] Furthermore, the parallel multimorphic convolution kernel includes parallel 1×1 convolution kernels, 31×1 convolution kernels, and 1×31 convolution kernels.

[0018] Furthermore, in S2, the frequency domain global context modeling is achieved through Fast Fourier Transform and Inverse Fast Fourier Transform.

[0019] Further, in S3, modulating the category score using the positioning quality includes: mapping the category score using a Sigmoid function, mapping the positioning quality using a lightweight function and then again using a Sigmoid function, and multiplying the mapping results element-wise to obtain the final classification confidence. The category score With the positioning quality Satisfying the relation:

[0020] ,

[0021] in, This represents the Sigmoid function. This indicates a lightweight mapping.

[0022] Furthermore, in S1, the image to be detected is an aerial image taken by a drone, and the size of the image to be detected is uniformly adjusted to 800×800 pixels before input.

[0023] The present invention has the following beneficial effects:

[0024] (1) By using the dual context attention downsampling module (DCAD), local detail features and surrounding context features are extracted in parallel during the downsampling stage, and global context reweighting is performed through the channel attention mechanism. This can reduce the resolution of the feature map while retaining the key texture information of the target, suppressing background interference, and alleviating the attenuation of small target features during the transmission process in the deep network.

[0025] (2) Global Context Path Aggregation Feature Pyramid Network (GC-PAFPN) shortens the path of high-level semantic features to shallow-level resolution features by adding a direct fusion path across two scales, thereby reducing the loss of semantic information. Combined with the parallel multimorphic convolution kernel of the cross-stage omnidirectional kernel module (CSPOmniKernel) and frequency domain global context modeling, it can effectively capture long-distance spatial dependencies and improve the model's ability to distinguish fuzzy small targets in complex scenes.

[0026] (3) Shape and semantic alignment decoupling head (SADH) achieves shape alignment through dynamic deformable convolution, semantic alignment through object probabilistic graph attention mechanism, and modulates classification confidence by using localization quality, so that classification confidence can reflect localization reliability, alleviate the problem of misalignment between classification and regression tasks in traditional detection heads, and reduce localization drift caused by sampling point deviation.

[0027] (4) The method of the present invention, through the synergistic effect of three modules, DCAD, GC-PAFPN and SADH, can effectively improve the detection capability of small targets, dense targets and targets in complex backgrounds in UAV aerial images while maintaining the lightweight model, and provides technical support for real-time target detection on the UAV end. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall architecture of the present invention.

[0029] Figure 2 This is a schematic diagram of the Dual Contextual Attention Downsampling Module (DCAD).

[0030] Figure 3 This is a schematic diagram of the cross-stage omnidirectional kernel module (CSPOmniKernel).

[0031] Figure 4 A schematic diagram of the Shape and Semantic Alignment Decoupling Head (SADH).

[0032] Figure 5 This is a flowchart of the method of the present invention.

[0033] Figure 6 This section compares the performance of different object detection algorithms on the VisDrone2021 dataset.

[0034] Figure 7 This is a comparison of detection performance under complex environments. The left side shows GC-YOLO proposed in this invention, and the right side shows YOLOv11n. Detailed Implementation

[0035] The invention will now be further described with reference to the accompanying drawings.

[0036] The overall framework of the method of this invention is as follows: Figure 5 As shown, it includes four main steps: dataset construction and preprocessing, model construction and training, model validation and evaluation, and testing.

[0037] Figure 1 The overall architecture of the improved target detection network described in this invention is presented. Figure 2 The internal structure of the Dual Contextual Attention Downsampling Module (DCAD) in this invention is shown; Figure 3 The internal structure of the cross-stage omnidirectional kernel module (CSPOmniKernel) in this invention is shown; Figure 4 The internal structure of the Shape and Semantic Alignment Decoupling Head (SADH) in this invention is shown, clearly demonstrating its dual-path asymmetric alignment mechanism.

[0038] The technical solution of this invention mainly includes the following steps:

[0039] Step 1: Dataset Construction and Preprocessing Stage.

[0040] An image dataset containing the target to be detected was obtained. The VisDrone2019 dataset was used during the training phase. The dataset includes thousands of images and corresponding annotation information for each image. The annotation information includes the target's location bounding box and category label. The dataset was then divided into training, validation, and test sets in a ratio of 7:2:1. The images in the dataset were preprocessed by uniformly adjusting the image size to a preset size of 800×800 and converting the annotation information into YOLO format.

[0041] Step 2, Model Building and Training Phase.

[0042] The model includes a backbone network with a dual contextual attention downsampling module (DCAD); a global contextual path aggregation feature pyramid network (GC-PAFPN) with a cross-stage omnidirectional kernel module (CSPOmniKernel); and a novel shape and semantic alignment decoupling head (SADH).

[0043] In YOLOv11n, the backbone network uses 3×3 convolutions with a stride of 2 for downsampling and outputs feature maps of different scales layer by layer. While reducing spatial resolution, this approach inevitably weakens fine-grained information, which is particularly detrimental to small targets that occupy only a few pixels. This can easily lead to their features being overwhelmed by background texture in deeper layers, resulting in missed detections. Furthermore, standard downsampling convolutions have limited receptive fields, lack the ability to model the surrounding context of the target, and apply a consistent processing method to different regions, making it difficult to adaptively adjust based on feature importance.

[0044] To address the characteristics of drone aerial images—namely, small targets with few pixels, weak texture, and complex backgrounds—this invention proposes a Dual Contextual Attention Downsampling Module (DCAD). This module transforms the downsampling process from simple scale compression to context-guided selective feature aggregation, reducing resolution while preserving target-related information and suppressing background interference. The backbone network consists of alternating stacks of DCAD and C3k2 modules, outputting multi-scale feature maps from different layers. The DCAD module replaces traditional convolution to perform the downsampling operation, and its structure is as follows: Figure 2 As shown.

[0045] Input feature map For example, the acquisition of the basic feature map satisfies the following relationship:

[0046] ,

[0047] in, .

[0048] Then this basic feature Simultaneously inputting preset local and peripheral branches, features are extracted in parallel through these two branches. The local branch uses 3×3 depthwise convolution to process the basic features and extract local detail features, while the peripheral branch uses 3×3 dilated depthwise convolution to process the basic features, expanding the receptive field without significantly increasing computational cost and extracting surrounding contextual features. This process satisfies the following relationship:

[0049] .

[0050] Then, the local detail features obtained from the local branches are... Peripheral context features obtained from surrounding branches The splicing and fusion are performed along the channel dimension to obtain the splicing features. The concatenated features are then processed by batch normalization and SiLU activation function, followed by feature fusion and channel compression through 1×1 convolution to obtain the fused features. Subsequently, the fused features are subjected to global context reweighting through a channel attention mechanism. This mechanism first processes the fused features using global average pooling to generate corresponding channel descriptors. These descriptors are then input into two fully connected layers, which learn the dependencies between channels. The outputs of the fully connected layers are then mapped using a sigmoid function to obtain the channel weights for each channel. Finally, these channel weights are multiplied channel-by-channel and applied to the fused features, completing the global context reweighting of the fused features. This process satisfies the following relationship:

[0051] ,

[0052] in, This is the channel descriptor obtained from global average pooling. , It consists of two fully connected layers. For activation function, For channel weights, For the Sigmoid function, This is the final output feature of the DCAD module. After the above processing, the backbone network outputs multi-level feature maps, completing the feature extraction process.

[0053] By jointly modeling local, surrounding, and global contextual information during the downsampling stage, DCAD can effectively alleviate the problem of small target features decaying with scale compression, providing a more discriminative feature representation for subsequent feature fusion and detection.

[0054] Drone aerial images often feature targets with wide scales, a high proportion of small targets, and complex backgrounds. The effectiveness of multi-scale feature fusion directly impacts detection performance. While traditional FPN / PAN methods can achieve top-down and bottom-up information transfer, the injection of high-level semantics into shallow layers often involves long paths, leading to insufficient semantic information for small targets during fusion. Furthermore, the backflow of shallow details to deeper layers is prone to gradual attenuation, affecting the localization stability of dense targets. To address this, this invention proposes a GC-PAFPN structure. This structure enhances inter-layer information interaction by constructing additional cross-scale fusion paths and introduces a CSPOmniKernel module at the fusion node to improve fusion quality. A schematic diagram of the CSPOmniKernel module is shown in Figure 3.

[0055] After obtaining the multi-level feature maps, they are input into the Global Context Path Aggregation Feature Pyramid Network for cross-scale feature fusion processing to obtain the fused feature map. This Global Context Path Aggregation Feature Pyramid Network adds a direct fusion path across two scales to the traditional feature pyramid adjacent-scale fusion. Let the backbone output multi-scale features be {C3, C4, C5}, and their fusion form satisfies the following relationship:

[0056] ,

[0057] Where T(⋅) is a channel-aligned 1×1 convolution, and U(⋅) is the upsampling operator. For aggregation operations, this fusion method allows high-level semantic features to participate more directly in the feature construction of high-resolution layers, reducing information loss during semantic transmission. Simultaneously, a cross-stage omnidirectional kernel module is used at the fusion node of the network for feature integration. When performing feature processing, the cross-stage omnidirectional kernel module first divides the input multi-level features into direct connection preservation branches and transformation enhancement branches. The direct connection preservation branch directly preserves the features, while the transformation enhancement branch performs feature enhancement processing. Specifically, it captures multi-scale features through parallel multi-morphological convolutional kernels and frequency domain global context modeling to achieve feature enhancement. This parallel multi-morphological convolutional kernel package... The system includes parallel 1×1, 31×1, and 1×31 convolutional kernels. Through parallel processing of multi-morphological convolutional kernels, it achieves multi-scale feature extraction. Frequency domain global context modeling is performed by fast Fourier transform and inverse fast Fourier transform to model the global context of features in the frequency domain, thereby capturing long-distance spatial dependencies. After the transform enhancement branch completes feature enhancement in the above manner, the enhanced features are fused with the features directly retained by the direct connection retention branch. Then, the global context path aggregation feature pyramid network completes the cross-scale fusion of features at each scale and outputs a fused feature map, thus completing the cross-scale feature fusion processing of multi-level feature maps.

[0058] After obtaining the fused feature map, it is input into the shape and semantic alignment decoupling head for target detection. This decoupling head outputs the category and bounding box of the target in the image to be detected. The shape and semantic alignment decoupling head contains independent classification and regression branches, which process the fused feature map in parallel. Let the multi-scale features output by the neck be... SADH first uses lightweight Stem to perform feature shaping to obtain Subsequently, two decoupled branches are used to process the features. The regression branch extracts features from the fused feature map through dynamic deformable convolution, allowing the receptive field to adaptively fit the actual contour of the target, thus achieving shape alignment. Simultaneously, after completing shape alignment, the regression branch outputs the corresponding target localization quality. The localization quality is used to characterize the reliability of target localization. The classification branch extracts features from the fused feature map through an object probabilistic graph attention mechanism, allowing the classifier to focus on the core region of the target with the highest discriminative power, thus achieving semantic alignment of the target. After completing semantic alignment, the classification branch outputs the corresponding target's category score. The category score is used to characterize the probability of the target belonging to a category. After obtaining the localization quality output from the regression branch and the category score output from the classification branch, the category score is modulated using the localization quality to obtain the final classification confidence. The specific process of modulation is as follows: first, the category score is directly mapped using the Sigmoid function; then, the localization quality is first lightly mapped and then mapped again using the Sigmoid function; finally, the results of the two Sigmoid function mappings are multiplied element-wise to obtain the final classification confidence. Category Score With positioning quality Satisfying the relation:

[0059] ,

[0060] in, This represents the Sigmoid function. The lightweight mapping, shape and semantic alignment decoupling head, based on the final classification confidence and the target location information output by the regression branch, outputs the specific category and corresponding bounding box of the target in the drone aerial image to be detected, thus completing the drone aerial target detection process for a single image.

[0061] Step 3: Model validation and evaluation phase.

[0062] To verify the effectiveness of the improved method of this invention, a dedicated experimental platform was built in this embodiment, and ablation experiments and comparative experiments were designed. The experimental platform was a GPU server running CentOS Linux 7 (Core) system. The server was equipped with two NVIDIA GeForce RTX 3080 Ti (12 GB) graphics cards, a 12th Gen Intel(R) Core(TM) i9-12900K CPU, and 128 GB of memory. On the software side, CUDA 11.8 and PyTorch 2.2.2 deep learning frameworks were used.

[0063] The model was trained using the SGD optimizer with a momentum of 0.937 and a weight decay of 5 × 10⁻⁶. -4The initial learning rate was set to 0.01, and cosine annealing was used for decay. Training lasted 300 epochs. To ensure consistency in comparisons, each model was trained from scratch. No additional techniques such as multi-scale testing were used during the testing phase. The performance metrics reported in the experiments are based on the final model results after reparameterization. COCO-style evaluation metrics were used to quantitatively assess model performance, primarily reporting mAP50 / % (IoU=0.50) and mAP50-95 / % (IoU=0.50-0.95), supplemented with AP50 and AP75 to reflect the localization quality under different IoU thresholds. Simultaneously, the performance of tiny and small targets was statistically analyzed based on target pixel area: targets with an area less than 16² were categorized as tiny targets, with the corresponding metric being AP_tiny; targets with an area between 16² and 32² were categorized as small targets, with the corresponding metric being AP_small. Furthermore, Params(M) and GFLOPs were also provided to measure the model's parameter size and computational cost.

[0064] In this embodiment, an ablation experiment was first designed based on YOLOv11n as the baseline model to verify the effectiveness of each improved module. The results of the ablation experiment are shown in Table 1.

[0065] Table 1 Ablation Experiment Results

[0066]

[0067] As shown in Table 1, the YOLOv11n baseline model achieves mAP50% and mAP50–95% of 38.5% and 23.1% respectively on the VisDrone2021 dataset. Its detection capability for tiny and small-scale targets is relatively limited (AP_tiny=6.2, AP_small=17.2), indicating that relying solely on the original detection head and feature fusion structure is insufficient to meet the detection needs of small-scale, densely distributed targets in aerial photography scenarios.

[0068] After introducing only the SADH decoupled detection head, the overall model performance was significantly improved, with mAP50% increasing to 42.3% and mAP50–95% increasing to 25.6%. Meanwhile, AP_tiny and AP_small increased to 7.3 and 19.9, respectively. Furthermore, the number of model parameters decreased from 2.584M to 2.2M, but the computational cost only increased slightly, indicating that SADH possesses excellent lightweight characteristics while improving the consistency between classification and regression.

[0069] By further introducing GC-PAFPN and DCAD, the model's ability to perform in multi-scale feature interaction and small target feature preservation is significantly enhanced, achieving a high level of detection performance in our ablation experiments. mAP50% and mAP50–95% are improved to 46.4% and 28.5%, respectively, with AP_tiny and AP_small reaching 8.1 and 23.1, respectively. This result validates the effectiveness of context-guided downsampling and enhanced cross-scale fusion structures in aerial small target detection.

[0070] When SADH, GC-PAFPN, and DCAD modules are used simultaneously, the model achieves optimal overall performance, with mAP50% and mAP50–95% reaching 46.4% and 28.7%, respectively. Meanwhile, AP50 / AP75 improves to 44.4 / 27.5, and the tiny and small objective metrics further increase to AP_tiny=8.2 and AP_small=23.4. While maintaining a computational cost of 18.7 GFLOPs, the number of parameters decreases from 4.269M to 4.02M, indicating that SADH effectively suppresses structural redundancy in the complete model, achieving a more reasonable trade-off between accuracy and complexity.

[0071] To verify the effectiveness of the improved algorithm proposed in this invention, the algorithm of this invention was compared and analyzed with other common algorithms in this field. The results are shown in Table 2.

[0072] Table 2. Comparison of experimental results of different algorithms on the public dataset VisDrone2021

[0073]

[0074] As shown in Table 2, under the same training and testing settings, the method of this invention achieves the best results in both the mAP50 / % and mAP50–95 / % core metrics, reaching 46.4% and 28.7% respectively, outperforming the mainstream detectors in comparison. Compared with YOLO10s (34.9% / 20.7%), the method of this invention improves by 11.5 and 8.0 percentage points in mAP50 / % and mAP50–95 / % respectively; compared with the lightweight baseline YOLO11n (38.5% / 23.1%), the corresponding improvements are 7.9 and 5.6 percentage points, indicating that the proposed improvements can bring stable and significant accuracy gains in aerial target detection tasks.

[0075] In terms of model complexity, the method of this invention has 4.02M parameters, far lower than heavier models such as Faster R-CNN (136.9M) and RetinaNet (32.7M), and also less than YOLO5s / YOLO8s / YOLO9s / YOLO10s (7.0–11.1M). The computational cost is 18.7 GFLOPs, significantly lower than methods such as SSD and RetinaNet (87.9–93.7 GFLOPs), and also lower than YOLO8s and YOLO9s (28.5 / 26.7 GFLOPs). Although the computational cost is increased compared to the lighter YOLO11n (6.3 GFLOPs) and EfficientDet (4.8 GFLOPs), it is traded for a more significant improvement in accuracy, demonstrating a superior accuracy-efficiency trade-off. Figure 6 It can be further observed that the method of the present invention maintains a leading position in both mAP50 / % and mAP50–95 / % and its advantage is more obvious under the more stringent IoU evaluation.

[0076] Step 4: Testing phase.

[0077] To intuitively evaluate the actual detection performance of various methods in UAV aerial photography scenarios, in addition to quantitative index comparisons, this section further provides a visual analysis of the results. The detection outputs of different algorithms in typical complex scenarios, such as dense targets, high-brightness interference, and nighttime environments, are shown in Figure 7. The left side represents the proposed GC-YOLO, and the right side represents YOLOv11.

[0078] As shown in Figure 7, in bright daytime aerial photography scenarios, GC-YOLO detects distant small targets more effectively. Compared to YOLOv11n, which falsely detects awning-tricycles in locally highly reflective areas, GC-YOLO tends to output reasonable "car" and "van" targets, indicating that the model effectively suppresses background interference and retains fine-grained target features more fully. In low-light scenarios such as nighttime roads, GC-YOLO can detect more pedestrians and trucks, mitigating the missed detection phenomenon and demonstrating its advantages in cross-scale representation and contextual supplementation. In dense traffic scenarios in urban blocks, GC-YOLO separates targets more clearly in overlapping and occluded areas, detecting more motors. Furthermore, compared to the baseline model, it has fewer duplicate and invalid boxes and more consistent category labeling, exhibiting better overall performance in detecting dense small targets. These improvements are consistent with the structural optimization directions of this invention in terms of downsampling information preservation, multi-scale fusion, and prediction consistency.

[0079] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A method for UAV aerial target detection based on an improved YOLOv11n, characterized in that: Includes the following steps: S1: The image to be detected is input into a backbone network containing a dual contextual attention downsampling module for feature extraction to obtain multi-level feature maps; the dual contextual attention downsampling module first downsamples the input features through convolution to obtain basic features, and then inputs the basic features into local branches and peripheral branches respectively to extract local detail features and peripheral context features in parallel. After fusing the local detail features and the peripheral context features, global context reweighting is performed through a channel attention mechanism; S2: Input the multi-level feature map into the global context path aggregation feature pyramid network for cross-scale feature fusion to obtain a fused feature map; the global context path aggregation feature pyramid network adds a direct fusion path across two scales on the basis of adjacent scale fusion, and adopts a cross-stage omnidirectional kernel module at the fusion node. The cross-stage omnidirectional kernel module captures multi-scale features through parallel multi-morphological convolution kernels and frequency domain global context modeling. S3: Input the fused feature map into the shape and semantic alignment decoupling head for target detection, and output the target category and bounding box; the shape and semantic alignment decoupling head includes a classification branch and a regression branch. The regression branch achieves shape alignment and outputs localization quality through dynamic deformable convolution. The classification branch achieves semantic alignment and outputs category score through object probabilistic graph attention mechanism. The category score is modulated using the localization quality to obtain the final classification confidence.

2. The UAV aerial target detection method based on the improved YOLOv11n as described in claim 1, characterized in that: In S1, the dual contextual attention downsampling module downsamples the input features using a 3×3 convolution with a stride of 2.

3. The UAV aerial target detection method based on the improved YOLOv11n as described in claim 1 or 2, characterized in that: In S1, the local branches use 3×3 depth convolution, and the peripheral branches use 3×3 dilated depth convolution to expand the receptive field.

4. The UAV aerial target detection method based on the improved YOLOv11n as described in claim 1, characterized in that: In S1, the channel attention mechanism includes: generating channel descriptors through global average pooling, learning the channel dependencies through two fully connected layers, and obtaining channel weights through the Sigmoid function for reweighting.

5. The UAV aerial target detection method based on the improved YOLOv11n as described in claim 1, characterized in that: In S2, the cross-stage omnidirectional kernel module divides the input features into a direct connection preservation branch and a transformation enhancement branch. The transformation enhancement branch enhances the features through the parallel multimorphic convolution kernel and frequency domain global context modeling, and then merges with the direct connection preservation branch.

6. The UAV aerial target detection method based on the improved YOLOv11n as described in claim 5, characterized in that: The parallel multimorphic convolution kernels include parallel 1×1 convolution kernels, 31×1 convolution kernels, and 1×31 convolution kernels.

7. The UAV aerial target detection method based on the improved YOLOv11n as described in claim 1 or 5, characterized in that: In S2, the frequency domain global context modeling is achieved through Fast Fourier Transform and Inverse Fast Fourier Transform.

8. The UAV aerial target detection method based on the improved YOLOv11n as described in claim 1, characterized in that: In S3, modulating the category score using the positioning quality includes: mapping the category score using a Sigmoid function, mapping the positioning quality using a lightweight function and then again using a Sigmoid function, and multiplying the mapping results element-wise to obtain the final classification confidence score. The category score With the positioning quality Satisfying the relation: ， in, This represents the Sigmoid function. This indicates a lightweight mapping.

9. The UAV aerial target detection method based on the improved YOLOv11n as described in claim 1, characterized in that: In S1, the image to be detected is an aerial image taken by a drone. Before input, the size of the image to be detected is uniformly adjusted to 800×800 pixels.

Images