An unmanned aerial vehicle aerial small target detection method, device, storage medium and product based on improved YOLOv11n
By embedding an adaptive multi-kernel receptive field fusion module and a multi-scale dilated convolution fusion module into the YOLOv11n network, the problems of multi-scale adaptation and feature dilution in small target detection in UAV aerial photography are solved, improving detection accuracy and robustness, and making it suitable for UAV platforms with limited resources.
Patent Information
- Application Number
- CN202610315290.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-16
- Publication Date
- 2026-06-12
AI Technical Summary
The existing YOLOv11n algorithm has several problems in small target detection in UAV aerial photography, such as the fixed receptive field being difficult to adapt to multi-scale targets, small target features being easily diluted, limited attention adjustment ability, and lack of semantic guidance for multi-scale feature interaction.
The adaptive multi-kernel receptive field fusion module (AMK-EFM) and the multi-scale dilated convolution fusion module (MS-DFM) are embedded in the YOLOv11n network. By configuring parallel convolution branches with different convolution kernels and dilation rates, enhanced channel attention, and learnable branch weights, adaptive feature extraction and multi-scale feature fusion are achieved.
It significantly improves the detection capability and robustness of small targets in UAV aerial images, solves the problems of insufficient receptive field adaptation and feature fusion dilution, and meets the application needs of UAV platforms with limited resources.
Smart Images
Figure CN122200437A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer vision and image processing technology, specifically relating to a method, device, storage medium and product for small target detection in UAV aerial photography based on an improved YOLOv11n. Background Technology
[0002] With the rapid iteration and popularization of drone technology, its application depth and breadth in various industries continue to expand: In the field of smart cities, drones can monitor urban traffic flow, investigate illegal buildings, and respond quickly to emergencies; in precision agriculture, they can monitor crop growth, identify pests and diseases early, and locate irrigation areas; in power line inspection scenarios, they can efficiently detect minor damage to transmission lines, insulator defects, and foreign object entanglement in conductors; in security monitoring, they can track suspicious persons and abnormal vehicles in real time over a large area. One of the core requirements of these application scenarios is to quickly and accurately detect various targets from aerial images, and small target detection (usually referring to targets with a pixel ratio of less than 32×32) is a key technology that directly determines the intelligence level and practical value of drone systems.
[0003] The unique characteristics of drone aerial images present multiple technical challenges for small target detection, specifically:
[0004] 1. Extreme target scale: Changes in flight altitude result in huge differences in target size within the same image, ranging from vehicles of tens of pixels to pedestrians of single-digit pixels. The model needs to handle scale changes spanning several orders of magnitude.
[0005] 2. Small, dense targets with weak features: Due to limitations in sensor resolution and shooting distance, the target pixel ratio is extremely small, and the appearance and texture features are limited, making them easily submerged by complex background noise.
[0006] 3. Limited computing platform resources: The UAV's onboard computing unit has strict requirements on model size, power consumption and inference speed, and needs to achieve high-precision detection while being lightweight and efficient.
[0007] Object detection technology has undergone a leapfrog development from traditional methods to deep learning methods. Traditional methods (such as the Viola-Jones algorithm and HOG+SVM) rely on hand-designed features and have extremely poor adaptability to small targets and complex backgrounds, and have been gradually phased out. Deep learning methods are divided into two-stage detection algorithms (such as Faster R-CNN and Mask R-CNN) and single-stage detection algorithms (such as the YOLO series, SSD, and RetinaNet).
[0008] Two-stage algorithms achieve high-precision detection through candidate region generation and fine classification, but they suffer from high computational load and slow inference speed, making it difficult to meet the real-time detection needs of UAVs.
[0009] Single-stage algorithms directly complete target localization and classification through end-to-end regression, which has a significant advantage in speed. Among them, the YOLO (You Only Look Once) series of algorithms has become the mainstream choice for drone scenarios due to its high efficiency.
[0010] While the YOLO series of algorithms has undergone continuous optimization in terms of lightweight design, detection accuracy, and speed, it still has the following core shortcomings when it comes to detecting small targets in drone aerial photography:
[0011] 1. Fixed receptive fields are difficult to adapt to multi-scale targets;
[0012] 2. Details of small targets are easily diluted during feature fusion;
[0013] 3. The attention mechanism has limited modulatory capacity in lightweight models;
[0014] 4. Multi-scale feature interactions lack semantic guidance.
[0015] The closest existing technology is the YOLOv11n object detection algorithm, whose core architecture and workflow are as follows:
[0016] (1) Network structure composition
[0017] Backbone feature extraction network: It extracts features from the input image through basic structures such as standard convolution and C3 module, and uses convolution kernels of fixed size to complete multi-scale feature capture. Features at each level are passed through a fixed process.
[0018] Neck fusion network: Construct a feature pyramid structure, and achieve the fusion of feature maps at different levels through cross-layer feature concatenation and basic convolution operations, thereby completing multi-scale feature interaction;
[0019] Head detection network: Receives multi-scale feature maps output from the neck, and completes target bounding box coordinate prediction and category determination through regression and classification branches, respectively, and finally outputs the detection results.
[0020] (2) Core working principle
[0021] With "lightweight + end-to-end" as its core, it controls the number of parameters by simplifying the number of network layers and reducing the number of convolutional kernels. At the same time, it relies on cross-layer fusion of feature pyramids to improve the multi-scale target adaptation capability. Ultimately, it seeks a balance between "detection speed" and "accuracy", and is suitable for general small target detection scenarios.
[0022] The existing YOLOv11n algorithm has significant shortcomings in small target detection tasks for UAV aerial photography, as follows:
[0023] Fixed receptive fields are difficult to adapt to multi-scale targets: The backbone network adopts a fixed-size convolutional kernel and a single receptive field design, which cannot simultaneously take into account the local details of small targets and the global semantic features of large targets. When facing scenarios with extreme target scales in drone aerial photography, the feature extraction is not adaptable enough.
[0024] In feature fusion, details of small targets are easily diluted: The feature fusion process of the neck network lacks targeted design. When the shallow high-resolution details of small targets are fused with deep low-resolution semantic features, they are easily diluted by the latter, causing the information of small targets to gradually decay during feature propagation.
[0025] Attention mechanisms have limited adaptability in lightweight models: Traditional efficient channel attention mechanisms lack dynamic adjustment mechanisms and cannot adaptively adjust the attention intensity based on scene information such as the density and distribution of small targets in the input image, resulting in poor screening and enhancement effects for small target features.
[0026] Multi-scale feature interaction lacks semantic guidance: Multi-scale feature fusion relies solely on simple cross-layer concatenation and convolutional transformation, without establishing a feature priority mechanism for small target detection. This results in a lack of directionality in the fusion of local detail features and global semantic features, failing to highlight key information about small targets. Summary of the Invention
[0027] Purpose of the invention: To address the shortcomings of existing YOLOv11n algorithms in small target detection in UAV aerial photography, such as insufficient receptive field adaptation, easy loss of small target features, limited attention adjustment, and lack of guidance in multi-scale fusion, this invention proposes a method, device, storage medium, and product for small target detection in UAV aerial photography based on an improved YOLOv11n. By fusing an adaptive multi-kernel receptive field enhancement module and a multi-scale dilated convolution fusion module, small target detection in UAV aerial photography can be achieved.
[0028] Technical solution: In the first aspect, this invention proposes a method for detecting small targets in UAV aerial photography based on an improved YOLOv11n, including:
[0029] Acquire drone aerial images;
[0030] The drone aerial images are input into a trained improved YOLOv11n network to obtain target detection results;
[0031] The improved YOLOv11n network is obtained by embedding an adaptive multi-kernel receptive field fusion module into the backbone network of the existing YOLOv11n network framework and a multi-scale dilated convolution fusion module into the neck network.
[0032] The adaptive multi-kernel receptive field fusion module achieves adaptive, multi-level feature extraction by configuring parallel convolutional branches with different convolutional kernels and dilation rates, combined with enhanced channel attention units and learnable branch weights; the enhanced channel attention units are located at the output of each convolutional branch; the feature maps of each convolutional branch after passing through the enhanced channel attention units are multiplied by their corresponding learnable branch weights and then weighted and fused.
[0033] The multi-scale dilated convolution fusion module employs a multi-scale dilated convolution and fixed weight fusion strategy to perform context-aware fusion of feature maps from different levels.
[0034] The method of embedding a multi-scale dilated convolutional fusion module in the neck network includes:
[0035] On each lateral connection and downsampling path of the feature pyramid constructed by the neck network, after feature concatenation and basic convolution operations, a multi-scale dilated convolution fusion module is inserted to obtain a multi-scale predicted feature map.
[0036] The detection head directly receives and processes the multi-scale predicted feature map enhanced by the multi-scale dilated convolution fusion module.
[0037] Furthermore, based on the number of layers in the backbone network, the backbone network is divided into shallow, medium, and deep regions; the adaptive multi-core receptive field fusion module is located in one layer of the shallow region, one layer of the medium region, and one layer of the deep region of the backbone network.
[0038] Furthermore, for each adaptive multi-kernel receptive field fusion module, a set of multiple parallel convolutional branches with different kernel sizes and dilation rates are predefined based on the depth position of the adaptive multi-kernel receptive field fusion module in the backbone network.
[0039] The feature map input to the adaptive multi-kernel receptive field fusion module is simultaneously fed into multiple parallel convolutional branches. Each convolutional branch uses standard convolution operations to capture feature information from local details to global context.
[0040] Furthermore, the enhanced channel attention unit introduces learnable gating parameters on the basis of the existing channel attention mechanism, constrains its range through the Sigmoid function, and multiplies it with the initial weights channel by channel to adaptively obtain the attention intensity of each branch feature map.
[0041] Furthermore, the learnable branch weights are a set of learnable scalar weight vectors. ,in, This represents the number of branches in the parallel convolution;
[0042] The scalar weight vector is obtained by optimizing the gradient descent algorithm during model training.
[0043] Furthermore, the multi-scale dilated convolution fusion module employs a multi-scale dilated convolution and fixed-weight fusion strategy to perform context-aware fusion of feature maps from different levels, including:
[0044] The input feature map is channel-adjusted and then dimensionality-reduced using 1×1 convolution to obtain the adjusted feature map.
[0045] The adjusted feature map is fed into a parallel group consisting of multiple depth-separable dilated convolution branches, each with a different dilation rate.
[0046] The output feature maps of each depthwise separable dilated convolutional branch are weighted and fused according to their respective fixed weight coefficients, wherein the depthwise separable dilated convolutional branch with the smallest dilation rate is given the highest weight; the fixed weight coefficients are set based on prior knowledge of the target detection task.
[0047] The weighted and fused feature map is then channel-integrated through a 1×1 convolution, and finally added to the original input feature map of the multi-scale dilated convolution fusion module through residual connections to obtain the final output.
[0048] Secondly, the present invention provides an electronic device, the electronic device comprising:
[0049] At least one processor;
[0050] and a memory communicatively connected to the at least one processor;
[0051] The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to execute a UAV aerial small target detection method based on the improved YOLOv11n.
[0052] Thirdly, the present invention proposes a computer-readable storage medium storing computer instructions for causing a processor to execute a method for detecting small targets in UAV aerial photography based on an improved YOLOv11n.
[0053] Fourthly, the present invention proposes a computer program product, which includes a computer program that, when executed by a processor, implements a method for detecting small targets in UAV aerial photography based on an improved YOLOv11n.
[0054] Beneficial Effects: This invention constructs an improved YOLO object detection network, embedding an adaptive multi-kernel receptive field fusion module in its backbone network and a multi-scale dilated convolution fusion module in its neck network. The AMK-EFM module achieves adaptive, multi-level feature extraction by configuring parallel branches with different convolution kernels and dilation rates, combined with enhanced channel attention and learnable branch weights. The MS-DFM module effectively enhances the expressive power of small target features through multi-scale dilated convolution and a fixed-weight fusion strategy for small targets. The two modules work together to significantly improve the model's detection capability and robustness for multi-scale, small targets in UAV aerial images. Compared with existing technologies, this invention has the following advantages:
[0055] (1) The AMK-EFM module of the present invention achieves dynamic adaptation of receptive field and effective fusion of multi-scale features through parallel multi-branch structure and learnable weights, thereby improving the adaptive feature extraction capability of multi-scale targets and solving the adaptation problem of fixed receptive field.
[0056] (2) The MS-DFM module of the present invention adopts a lightweight multi-scale context extraction and fixed weighting strategy to explicitly enhance the details of small targets and avoid them being diluted during feature fusion.
[0057] (3) Lightweight design and performance balance: Both modules of the present invention adopt lightweight structures (such as depthwise separable convolution), which achieve significant performance improvement with limited computational overhead, making them suitable for deployment on resource-constrained platforms such as UAVs;
[0058] (4) The AMK-EFM of the present invention provides a multi-scale foundation from the source of feature extraction, and the MS-DFM performs targeted enhancement in the fusion stage, forming a collaborative link of "extraction → protection → enhancement". Without significantly increasing the number of parameters, it improves the detection accuracy and robustness of small targets and meets the application requirements of UAVs with limited onboard resources. Attached Figure Description
[0059] Figure 1 This is a flowchart of the detection method of the present invention;
[0060] Figure 2 A diagram of the ADF-YOLOv11 network architecture;
[0061] Figure 3 A diagram of the AMK-EFM module structure;
[0062] Figure 4 Here is a structural diagram of the MS-DFM module;
[0063] Figure 5 This is a schematic diagram of the enhanced channel attention submodule structure;
[0064] Figure 6A comparison chart of mAP curves before and after improvement. Detailed Implementation
[0065] The technical solution of the present invention will now be further described in conjunction with the accompanying drawings and embodiments.
[0066] This invention improves the YOLOv11n network through innovative modular design and collaborative integration optimization. The core focus is on the technical chain of adaptive feature extraction, small target feature enhancement, and scientific experimental verification. The specific steps are as follows:
[0067] Step 1: Dataset preparation and standardization. This includes:
[0068] The embodiments of this invention use the publicly available VisDrone2019 drone aerial photography dataset, which covers a variety of typical aerial photography scenarios such as smart cities, agriculture, and inspection. It has a rich variety of target categories and is authoritative in the industry and representative of the scenarios.
[0069] After removing blurry, missing, or incorrectly labeled images from the dataset, high-quality images were retained for model training and evaluation, resulting in a total of 8629 valid images. These were distributed as follows: 6471 images in the training set (75%), 548 images in the validation set (6.4%), and 1610 images in the test set (18.6%), covering more than 10 target categories including pedestrians, vehicles, and non-motorized vehicles, simulating real-world drone aerial photography scenarios.
[0070] Step 2: Based on the original YOLOv11n network framework, add custom enhancement modules to build the improved network (ADF-YOLO). Specifically, this includes:
[0071] Based on the YOLOv11n framework, an adaptive multi-kernel receptive field enhancement module (AMK-EFM) is embedded in the backbone network to replace some standard convolutional or C3 modules, enabling multi-level, adaptive receptive field feature extraction and fusion. A multi-scale dilated convolutional fusion module (MS-DFM) is embedded in the neck network to perform context-aware fusion of feature maps from different levels and enhance small target features. The original detection head is preserved, forming a complete link of feature extraction enhancement, feature fusion enhancement, and target detection.
[0072] Specifically, adaptive multi-kernel receptive field enhancement modules (AMK-EFM) configured with different modes are inserted at the three key locations of the shallow, middle and deep layers of the backbone network to form an adaptive receptive field enhancement link; on each lateral connection and downsampling path of the feature pyramid construction in the neck network, after feature concatenation and basic convolution operation, a multi-scale dilated convolution fusion module (MS-DFM) is inserted to perform context enhancement for small targets on the fused features; the detection head directly receives and processes the multi-scale predicted feature map enhanced by the multi-scale dilated convolution fusion module (MS-DFM).
[0073] The specific module design is as follows:
[0074] (1) The establishment of the Adaptive Multi-Kernel Receptive Field Enhancement Module (AMK-EFM) includes:
[0075] Module location: Layer 3 (shallow), Layer 6 (middle), and Layer 13 (deep) of the backbone network;
[0076] Based on the module's depth position in the network (shallow, mid, deep), a set of parallel convolutional path configurations with different kernel sizes and dilation rates are predefined, including:
[0077] Shallow layer (layer 3): Convolution kernel 3×3, 5×5, dilation rate 1;
[0078] Middle layer (6th layer): Increase convolution kernel to 7×7, dilation rate to 2;
[0079] Deep layer (13th layer): Convolution kernels 3×3, 5×5, 7×7, with dilation rates of 2, 2, and 3 respectively;
[0080] Shallow branches focus on local details (such as 3×3 and 5×5 convolutions), while deep branches utilize larger receptive fields to capture global semantics (such as introducing 7×7 convolutions with dilation rates of 2 or 3). Each branch employs standard convolution operations to capture feature information from local details to global context.
[0081] Enhanced Channel Attention Unit: Each branch output is processed by the Enhanced Channel Attention (EnhancedECA) unit. This unit, based on the traditional efficient channel attention mechanism, introduces learnable gating parameters to adaptively adjust the attention intensity of each branch's feature map. The Enhanced Channel Attention Unit generates the final channel attention weights in the following manner. Introduce a learnable gating parameter. Its range is constrained by the Sigmoid function and multiplied channel by channel with the initial weights:
[0082]
[0083] In the formula, g is a learnable gating parameter with an initial value of 1.0, and σ is the Sigmoid function. g can dynamically adjust the attention intensity according to the density of small targets in the input image: when small targets are densely distributed, the gating parameter approaches 1.2, which strengthens the attention to detail channels; when small targets are sparse, the parameter approaches 0.8, which avoids over-focusing on background noise and achieves scene-adaptive adjustment of attention intensity.
[0084] Introduce a set of learnable branch weight parameters In the formula, n is the number of branches, and the weight parameter of this branch is optimized by the gradient descent algorithm during the model training process to adaptively adjust the weight of the first branch. The contribution percentage of each branch's features in the fusion result can be initially set to a uniform distribution, for example, 1 / n. The attention-enhanced feature maps of each branch are multiplied by their corresponding learnable weights and then weighted and fused to achieve adaptive multi-scale feature selection and integration by the network, represented as: Weighted fusion, where the weights are optimized using gradient descent; where, This represents the feature map after multi-scale branch weighted fusion. This represents the total number of parallel convolution branches. Indicates the first Learnable weight parameters for each parallel convolutional branch, This indicates an enhanced channel attention operation, used for the first channel. Output feature map of each branch Perform channel-level feature filtering and enhancement. Indicates the first Each parallel convolutional branch outputs a feature map with a specific receptive field, corresponding to feature information at different scales.
[0085] Residual connection: The weighted multi-branch feature maps are concatenated along the channel dimension to form a composite feature map that integrates multi-scale information. Subsequently, this feature map is processed by 1×1 convolution (channel aggregation), 3×3 group convolution (spatial augmentation), and 1×1 convolution (channel refinement) before being added to the original input to avoid feature loss.
[0086] (2) The establishment of the multi-scale dilated convolution fusion module (MS-DFM) includes:
[0087] Module location: Layers 18, 23, 28, and 33 of the neck network (lateral connections and downsampling paths of the feature pyramid);
[0088] Lightweight dimensionality reduction: The number of channels is reduced to 75% of the input by 1×1 convolution, reducing computational complexity; this step aims to reduce the complexity of subsequent operations while striving to preserve the original details that are sensitive to small targets.
[0089] Multi-scale dilated convolution: 3 parallel depthwise separable convolution branches with dilation rates of 1, 2, and 3 (corresponding to receptive fields of 3×3, 5×5, and 7×7), simultaneously capturing pixel-level local details (dilation=1), neighborhood context (dilation=2), and broader semantic information (dilation=3), i.e., local / intermediate / global features.
[0090] Fixed-weighted fusion: The output feature maps of the three branches are weighted and fused according to fixed weight coefficients. The branch corresponding to the smallest receptive field (the branch with the highest dilation rate of 1 has the highest weight, enhancing details of small targets) is given the highest weight. This explicitly enhances the detail features sensitive to small targets during the feature fusion stage. The output feature maps of the three parallel branches... (corresponding to expansion rates 1, 2, and 3 respectively), according to a fixed weighted coefficient vector. Weighted fusion is represented as: The fixed weight coefficient vector is set based on prior knowledge of the small object detection task, aiming to explicitly strengthen the features most sensitive to local details.
[0091] Residual connection: After the fused features are integrated by 1×1 convolution, they are added to the original input feature map of the module through residual connection to form the final output of the module, thus avoiding the loss of small target feature information during the fusion process.
[0092] Two modules are integrated into the YOLOv11n framework to form a complete enhancement chain. AMK-EFM modules with different configurations are embedded at key depth locations in the backbone network to construct an adaptive receptive field expansion path. After each cross-layer feature concatenation and basic transformation in the neck network, an MS-DFM module is embedded to refine the fused features for small targets. Through this design, the AMK-EFM module provides a rich and scale-complete feature base from the source, while the MS-DFM module performs targeted filtering and enhancement of these features in subsequent fusion stages. The two modules work together to ensure the effective transmission and enhancement of small target information during the network's forward propagation.
[0093] Step 3: Ablation Experiment
[0094] To accurately evaluate the independent technical contributions of the Adaptive Multi-Kernel Receptive Field Enhancement Module (AMK-EFM) and the Multi-Scale Dilated Convolutional Fusion Module (MS-DFM) to the model's detection performance, as well as the synergistic effect of combining the two modules, the following four models were designed for ablation experiments: the original YOLOv11n (without any improvements); adding the AMK-EFM module alone; adding the MS-DFM module alone; and introducing both the AMK-EFM and MS-DFM modules simultaneously. Details are shown in Table 1, and the experimental configurations are shown in Table 2.
[0095] Table 1 Ablation Experiment Grouping
[0096]
[0097] Table 2 Experimental Configuration
[0098]
[0099] After training, all models were evaluated on the same test set using the following common metrics in the object detection field for objective evaluation: Precision: measures the proportion of samples predicted as positive that are actually positive, reflecting the accuracy of the model's detection results; Recall: measures the proportion of samples that are actually positive that are correctly predicted, reflecting the model's ability to capture targets; mAP@0.5: the average precision when the Intersection over Union (IoU) threshold is set to 0.5, comprehensively reflecting the model's overall detection accuracy for targets of different categories and scales.
[0100] To further verify the industry competitiveness of the improved model of this invention and to eliminate the limitation of "only having advantages within its own improvement system", we selected the current mainstream lightweight target detection models of the same scale for horizontal comparison. All experiments strictly followed the principle of "fairness and consistency" to ensure that the comparison results are objective and credible.
[0101] Focusing on the core requirements of lightweight and high real-time performance, three generations of lightweight models from the YOLO series were selected: YOLOv10n, YOLOv11n (i.e., the benchmark model for the ablation experiment of this invention), and YOLOv12n. All three models are low-parameter, high-speed solutions widely used in the field of target detection, which are completely consistent with the design positioning of ADF-YOLO of this invention and have direct comparative significance.
[0102] All comparison models used the experimental configuration and training hyperparameters (learning rate, number of training epochs, batch size, etc.) in Table 2, and were trained and tested based on the VisDrone2019 dataset preprocessed in step 1. The evaluation metrics were uniformly precision, recall, and mAP@0.5 to ensure that only the model structure was different. The experimental results are shown in Table 3 below:
[0103] Table 3 Ablation Experiment
[0104]
[0105] The experimental results clearly show that both core modules have independent positive technical value, and embedding either module alone can significantly improve model performance. After the two modules are integrated, the improvement in mAP@0.5 (4.2 percentage points) is close to the sum of the improvement of the two individual modules, which fully verifies the complementary design logic of "multi-scale feature extraction + small target feature enhancement" and achieves a synergistic effect of "1+1>2".
[0106] To verify the competitiveness of this invention among models of similar scale, the current mainstream lightweight object detection models (YOLOv10n, YOLOv11n, YOLOv12n) were selected as comparison objects. All models were tested on the same test set and under the same evaluation criteria (precision, recall, mAP@0.5). The comparison results are shown in Table 4 below:
[0107] Table 4 Comparative Experiments
[0108]
[0109] The comparative results show that the ADF-YOLO model of this invention comprehensively outperforms models of the same scale in core detection metrics—mAP@0.5 is 2.4 percentage points higher than YOLOv12n, precision is 2.9 percentage points higher, and recall is 2.5 percentage points higher. It also addresses the three core requirements of "detection accuracy, lightweight design, and real-time performance," better suited to the resource constraints of UAV onboard computing units, and possesses stronger practical application value. This demonstrates the positive technical contributions of the AMK-EFM module and the MS-DFM module, as well as the synergistic effect of their integration, providing quantitative experimental support for the innovation and effectiveness of the overall technical solution.
[0110] This invention innovatively introduces an Adaptive Multi-Kernel Receptive Field Enhancement Module (AMK-EFM) and a Multi-Scale Dilated Convolutional Fusion Module (MS-DFM) into the YOLOv11n framework. This specifically addresses four core shortcomings of existing technologies in UAV aerial target detection: "fixed receptive fields are difficult to adapt to multi-scale targets, small target details are easily diluted, attention adjustment capabilities are limited, and multi-scale feature interactions lack semantic guidance." The specific technical effects are as follows:
[0111] I. Overcoming the limitations of fixed receptive fields and enhancing multi-scale adaptation capabilities
[0112] To address the limitation of fixed receptive fields in adapting to multi-scale targets, the AMK-EFM module achieves continuous receptive field coverage from local to global through dynamic branch configuration, enhanced channel attention, and learnable weight fusion. The module pre-defines differentiated convolutional kernels and dilation rates based on network depth, and dynamically adjusts attention using a learnable gating parameter g to adapt to extreme scale differences, ranging from tens of pixels for vehicles to single-digit pixels for pedestrians.
[0113] II. Strengthen the protection of small target features to avoid information dilution during fusion.
[0114] To address the pain point of "small target details being easily diluted", the MS-DFM module adopts a triple design of "lightweight dimensionality reduction + fixed weighted fusion + residual connection". It retains fine-grained features through 1×1 convolution, strengthens the small target sensitive branch with weights of v=[0.5,0.3,0.2], and avoids feature loss through residual connection.
[0115] III. Optimize attention regulation to improve feature selection quality
[0116] To address the limitation of "limited attention adjustment capability," the AMK-EFM's built-in EnhancedECA unit introduces a learnable gating parameter g, enabling scene-adaptive adjustment. It enhances details when small targets are densely packed and suppresses background noise when they are sparse. In complex backgrounds, the recognition accuracy of small targets is improved by 15% compared to the baseline, compensating for the rigidity of traditional attention mechanisms.
[0117] IV. Establish integrated semantic guidance to enhance the effectiveness of multi-scale interactions.
[0118] To address the deficiency of "lack of semantic guidance in multi-scale feature interactions," the two modules form a collaborative closed loop: AMK-EFM provides a multi-scale feature foundation of "details + semantics," while MS-DFM establishes small target priorities through fixed weighting, achieving targeted optimization of "extraction → preservation → enhancement." This significantly improves the model's robustness in complex scenarios.
[0119] In summary, this invention effectively solves the problems of insufficient receptive field adaptation and feature fusion dilution in small target detection in UAV aerial photography through the collaborative design of AMK-EFM and MS-DFM modules. It improves detection accuracy while maintaining the lightweight characteristics of the model, making it suitable for practical application scenarios such as UAV inspection and smart city monitoring.
Claims
1. A method for detecting small targets in UAV aerial photography based on an improved YOLOv11n, characterized in that: include: Acquire drone aerial images; The drone aerial images are input into a trained improved YOLOv11n network to obtain target detection results; The improved YOLOv11n network is obtained by embedding an adaptive multi-kernel receptive field fusion module into the backbone network of the existing YOLOv11n network framework and a multi-scale dilated convolution fusion module into the neck network. The adaptive multi-kernel receptive field fusion module achieves adaptive, multi-level feature extraction by configuring parallel convolutional branches with different convolutional kernels and dilation rates, combined with enhanced channel attention units and learnable branch weights; the enhanced channel attention units are located at the output of each convolutional branch; the feature maps of each convolutional branch after passing through the enhanced channel attention units are multiplied by their corresponding learnable branch weights and then weighted and fused. The multi-scale dilated convolution fusion module employs a multi-scale dilated convolution and fixed weight fusion strategy to perform context-aware fusion of feature maps from different levels. The method of embedding a multi-scale dilated convolutional fusion module in the neck network includes: On each lateral connection and downsampling path of the feature pyramid constructed by the neck network, after feature concatenation and basic convolution operations, a multi-scale dilated convolution fusion module is inserted to obtain a multi-scale predicted feature map. The detection head directly receives and processes the multi-scale predicted feature map enhanced by the multi-scale dilated convolution fusion module.
2. The method for detecting small targets in UAV aerial photography based on improved YOLOv11n according to claim 1, characterized in that: Based on the number of layers in the backbone network, the backbone network is divided into shallow, medium, and deep regions; the adaptive multi-core receptive field fusion module is located in one layer of the shallow region, one layer of the medium region, and one layer of the deep region of the backbone network.
3. The method for detecting small targets in UAV aerial photography based on the improved YOLOv11n according to claim 2, characterized in that: For each adaptive multi-kernel receptive field fusion module, a set of multiple parallel convolutional branches with different kernel sizes and dilation rates are predefined based on the depth position of the adaptive multi-kernel receptive field fusion module in the backbone network. The feature map input to the adaptive multi-kernel receptive field fusion module is simultaneously fed into multiple parallel convolutional branches. Each convolutional branch uses standard convolution operations to capture feature information from local details to global context.
4. The method for detecting small targets in UAV aerial photography based on improved YOLOv11n according to claim 3, characterized in that: The enhanced channel attention unit introduces learnable gating parameters based on the existing channel attention mechanism. The range of gating parameters is constrained by the Sigmoid function and multiplied with the initial weights channel by channel to adaptively obtain the attention intensity of each branch feature map.
5. The method for detecting small targets in UAV aerial photography based on the improved YOLOv11n according to claim 4, characterized in that: The learnable branch weights are a set of learnable scalar weight vectors. ,in, This represents the number of branches in the parallel convolution; The scalar weight vector is obtained by optimizing the gradient descent algorithm during model training.
6. A method for detecting small targets in UAV aerial photography based on an improved YOLOv11n, as described in claim 1, is characterized in that: The multi-scale dilated convolution fusion module employs a multi-scale dilated convolution and fixed-weight fusion strategy to perform context-aware fusion of feature maps from different levels, including: The input feature map is channel-adjusted and then dimensionality-reduced using 1×1 convolution to obtain the adjusted feature map. The adjusted feature map is fed into a parallel group consisting of multiple depth-separable dilated convolution branches, each with a different dilation rate. The output feature maps of each depthwise separable dilated convolutional branch are weighted and fused according to their respective fixed weight coefficients, wherein the depthwise separable dilated convolutional branch with the smallest dilation rate is given the highest weight; the fixed weight coefficients are set based on prior knowledge of the target detection task. The weighted and fused feature map is then channel-integrated through a 1×1 convolution, and finally added to the original input feature map of the multi-scale dilated convolution fusion module through residual connections to obtain the final output.
7. An electronic device, characterized in that, The electronic device includes: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor, which is then executed by the at least one processor to enable the at least one processor to perform the UAV aerial small target detection method based on the improved YOLOv11n as described in any one of claims 1-6.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the UAV aerial photography small target detection method based on the improved YOLOv11n as described in any one of claims 1-6.
9. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the UAV aerial photography small target detection method based on the improved YOLOv11n as described in any one of claims 1-6.