Small object detection in China UAV aerial imagery presents unique challenges due to the high altitude perspective, complex backgrounds, severe occlusions, and extreme variations in object scale. Existing lightweight detection models often suffer from insufficient accuracy when deployed on resource-constrained edge platforms. In this work, we propose an improved YOLOv11n algorithm that incorporates multi-scale attention mechanisms, a novel feature pyramid network, and an enhanced bounding box regression loss to address these limitations. Our method is specifically designed for China UAV applications such as agricultural monitoring, urban surveillance, and disaster assessment, where real-time and precise detection of small objects is critical.
The baseline YOLOv11n model achieves a good balance between speed and parameter count, but its capacity to capture fine-grained details of tiny objects is limited. To overcome this, we first introduce a Multi-Scale Attention Mechanism (MSAM) that replaces the original C3k2 module in the backbone. MSAM simulates the Transformer architecture using efficient convolutional operations, enabling the network to better extract high-frequency features from small objects. Second, we design a Valid Multi-Scale Feature Pyramid Network (VMFPN) to replace the conventional neck structure. VMFPN enhances cross-level multi-scale feature fusion, effectively preventing the loss of small-object information in deeper layers. Finally, we adopt the Inner-ShapeIoU loss function, which combines an inner-box mechanism with shape-aware weighting, improving localization accuracy for irregular or tiny bounding boxes. Experiments on the VisDrone2019 dataset demonstrate that our model achieves a mAP50 of 40.7% and mAP50-95 of 24.4%, outperforming the baseline by 8.7 and 5.2 percentage points, respectively, while only slightly increasing parameters. This work provides an efficient and accurate solution for China UAV edge computing platforms.
1. Introduction
China UAV technology has rapidly advanced, enabling a wide range of applications including precision agriculture, infrastructure inspection, and public safety. However, the aerial perspective introduces unique object detection difficulties: objects often occupy only a few pixels, background clutter is intense, and severe occlusions are common. Traditional deep learning detectors, while achieving high accuracy on standard datasets, struggle with these characteristics. In the context of China UAV operations, real-time detection is essential, yet many high-performance models are too computationally expensive for embedded devices.
One-stage detectors, particularly the YOLO family, offer an excellent trade-off between speed and accuracy. The latest YOLOv11n is designed for lightweight deployment, but its performance on small objects remains suboptimal. To address this, we propose three key improvements that synergistically enhance multi-scale feature representation and localization precision. Our contributions are: (1) a novel MSAM module that mimics Transformer attention via efficient convolutions, boosting detail capture; (2) a VMFPN neck that deeply fuses features across scales while maintaining computational efficiency; (3) a specialized Inner-ShapeIoU loss that better handles the geometric properties of small targets. We validate our approach on the VisDrone2019 benchmark, a representative dataset for China UAV aerial scenes, and show significant gains over both the baseline and larger models such as YOLOv11s.
2. Related Work
Small object detection has been extensively studied in computer vision. Recent approaches for China UAV imagery often modify the backbone or neck of YOLO variants. For example, some works introduce attention modules like CBAM or CA to enhance feature selection, while others design specialized feature pyramids to preserve shallow details. However, many of these methods increase computational cost substantially. Our work builds upon these ideas but focuses on efficiency, using convolution-based attention and a lightweight pyramid structure (VMFPN) that achieves superior fusion without heavy overhead. Additionally, bounding box regression losses have evolved from IoU to CIoU and more recently to Shape-IoU and Inner-IoU. We combine the strengths of both, resulting in Inner-ShapeIoU that is particularly effective for small, elongated objects common in China UAV imagery.
3. Proposed Method
3.1 Multi-Scale Attention Mechanism (MSAM)
The backbone of YOLOv11n uses C3k2 modules for feature extraction. Although efficient, C3k2 lacks sufficient multi-scale context and high-frequency detail enhancement. To address this, we propose MSAM, which replaces the bottleneck inside C3k2. MSAM consists of a convolutional attention sublayer and a feed-forward network (FFN), connected via residual paths. The structure mimics Transformer encoding but uses only convolution operations, making it hardware-friendly.
The computation flow is:
$$ f = \text{Norm}(x + \text{ConvAttention}(x)) $$
$$ y = \text{Norm}(f + \text{FFN}(f)) $$
In ConvAttention, the input \(x\) serves as Query, and it is convolved with learnable kernels (Key) using striped convolutions (1×k and k×1) for efficiency. The result is normalized via GDN (Grouped Double Normalization) which applies Softmax in spatial dimensions and grouped L2 norm in channel dimensions. The normalized attention map is then convolved with Value kernels to produce the attention output. The FFN comprises two 3×3 convolutional layers with ReLU activation, providing a larger receptive field than typical linear projections. This design allows MSAM to capture fine-grained details and multi-scale patterns without the heavy computation of standard self-attention.
3.2 Valid Multi-Scale Feature Pyramid Network (VMFPN)
The neck network of YOLOv11n uses a simple FPN+PAN structure, which does not fully exploit cross-level information, especially for small objects. We propose VMFPN, a novel pyramid architecture that integrates a binary hierarchical vision transformer (BHVIT) style module. VMFPN consists of four stages, each with different spatial and channel dimensions. Given input image \(I \in \mathbb{R}^{3\times H\times W}\), a convolutional patch embedding layer projects it into a feature sequence \(X_0 \in \mathbb{R}^{64\times \frac{H}{4}\times \frac{W}{4}}\).
$$ H_0 = \text{GELU}(\text{BN}(\text{Conv}(I))), \quad X_0 = H_0 + P_e $$
where \(P_e\) is a learnable positional embedding. For stages 1 and 2 (high spatial resolution), we use a Multi-Scale Grouped Dilated Convolution (MSGDC) as the token mixer to avoid the high cost of attention. MSGDC consists of three 3×3 grouped dilated convolutions with different dilation rates. For stages 3 and 4, we employ a Multi-Scale Multi-Head Attention (MSMHA) module to perform token-wise feature fusion. Each downsampling layer uses a 2×2 convolution with stride 2, doubling the channel count. This design efficiently aggregates local and global information while keeping the parameter count low. VMFPN significantly boosts multi-scale feature interaction, making small object features survive deeper into the network.
3.3 Inner-ShapeIoU Loss
The baseline YOLOv11n uses CIoU loss, which adds a penalty on center point distance and aspect ratio consistency. However, for small targets, the center distance penalty can dominate, hindering optimization of actual overlap. We adopt Inner-ShapeIoU, which integrates Inner-IoU’s auxiliary bounding box mechanism with Shape-IoU’s shape-aware weighting.
Inner-IoU introduces a ratio-controlled inner box to compute IoU, defined as:
$$ \text{inter} = (\min(b_r^{\text{gt}}, b_r) – \max(b_l^{\text{gt}}, b_l)) \times (\min(b_b^{\text{gt}}, b_b) – \max(b_t^{\text{gt}}, b_t)) $$
$$ \text{union} = (w^{\text{gt}} \times h^{\text{gt}}) \times (\text{ratio})^2 + (w \times h) \times (\text{ratio})^2 – \text{inter} $$
$$ \text{IoU}_{\text{inner}} = \frac{\text{inter}}{\text{union}} $$
Shape-IoU adds shape-based weights to the distance penalty:
$$ w_w = \frac{2 \times (w^{\text{gt}})^{S_{\text{scale}}}}{(w^{\text{gt}})^{S_{\text{scale}}} + (h^{\text{gt}})^{S_{\text{scale}}}}, \quad h_h = \frac{2 \times (h^{\text{gt}})^{S_{\text{scale}}}}{(w^{\text{gt}})^{S_{\text{scale}}} + (h^{\text{gt}})^{S_{\text{scale}}}} $$
$$ \text{distance}_{\text{shape}} = h_h \times (x_c – x_c^{\text{gt}})^2 / c^2 + w_w \times (y_c – y_c^{\text{gt}})^2 / c^2 $$
$$ \Omega_{\text{shape}} = \sum_{t \in \{w,h\}} (1 – e^{-\omega_t})^4 $$
$$ L_{\text{Shape-IoU}} = 1 – \text{IoU} + \text{distance}_{\text{shape}} + 0.5 \times \Omega_{\text{shape}} $$
Our proposed Inner-ShapeIoU simply replaces the IoU term in Shape-IoU with Inner-IoU, i.e.,
$$ L_{\text{Inner-ShapeIoU}} = 1 – \text{IoU}_{\text{inner}} + \text{distance}_{\text{shape}} + 0.5 \times \Omega_{\text{shape}} $$
This combination reduces the negative impact of low-quality samples (through the inner box) while emphasizing geometric shape consistency, leading to more robust regression for small objects in China UAV scenes.
4. Experiments
4.1 Dataset and Settings
We evaluate our method on the VisDrone2019 dataset, which contains 6,471 training images and 1,610 test images captured from China UAV platforms. The images cover diverse urban and rural scenes with 10 object categories including pedestrian, car, bicycle, and van. All experiments are performed on an NVIDIA GeForce RTX 4090 GPU with 24 GB memory. We use PyTorch 2.0.0, CUDA 12.9, and Python 3.10.16. Images are resized to 640×640. Training uses SGD optimizer for 200 epochs, batch size 8, initial learning rate 0.01, weight decay 0.0005, and patience 20 for early stopping. Key hyperparameters are summarized in the table below.
| Parameter | Value |
|---|---|
| Epochs | 200 |
| Batch-size | 8 |
| Works | 8 |
| Optimizer | SGD |
| Patience | 20 |
| Initial learning rate | 0.01 |
| Final learning rate | 0.01 |
| Weight-Decay | 0.0005 |
4.2 Ablation Study
We perform ablation experiments to verify the contribution of each proposed module. The baseline YOLOv11n is denoted as Base. We add MSAM (A), VMFPN (B), and Inner-ShapeIoU (C) individually and in combination. Results are shown in the table below. All experiments run under identical conditions.
| Method | A | B | C | Params (M) | P (%) | R (%) | mAP50 (%) | mAP50-95 (%) | FLOPs (G) |
|---|---|---|---|---|---|---|---|---|---|
| YOLOv11n | 2.67 | 43.3 | 33.4 | 32.0 | 19.2 | 6.3 | |||
| (1) | √ | 2.82 | 44.7 | 33.7 | 33.8 | 19.8 | 6.7 | ||
| (2) | √ | 2.50 | 49.4 | 38.9 | 39.8 | 23.7 | 13.1 | ||
| (3) | √ | 2.59 | 42.9 | 32.6 | 32.0 | 18.8 | 6.5 | ||
| (4) | √ | √ | 2.82 | 44.1 | 34.4 | 33.9 | 19.8 | 6.7 | |
| (5) | √ | √ | 2.72 | 50.3 | 39.7 | 40.6 | 24.3 | 13.4 | |
| (6) | √ | √ | 2.50 | 50.1 | 39.3 | 40.1 | 24.1 | 13.1 | |
| (7) Ours | √ | √ | √ | 2.72 | 50.0 | 39.6 | 40.7 | 24.4 | 13.4 |
The baseline achieves mAP50 of 32.0%. Adding MSAM alone increases mAP50 by 1.8 points, confirming its effectiveness in capturing details. VMFPN alone provides the largest gain: +7.8 points in mAP50 and +4.5 points in mAP50-95, while even reducing parameters from 2.67M to 2.50M. This demonstrates the superiority of VMFPN in multi-scale feature fusion. Inner-ShapeIoU alone does not improve mAP, but when combined with VMFPN (row 6) it yields a slight boost (+0.3 points mAP50). The full model (row 7) achieves the best overall accuracy: 40.7% mAP50 and 24.4% mAP50-95, with only 2.72M parameters and 13.4 GFLOPs. The synergy of all three components is evident.
4.3 Comparison with State-of-the-Art
We compare our model with several mainstream detection methods, including YOLOv5n, YOLOv8n, Faster R-CNN, Drone-YOLO, YOLOv10s, YOLOv11n, and YOLOv11s. All models are trained and tested on the same VisDrone2019 dataset. Results are listed below.
| Method | Params (M) | FLOPs (G) | P (%) | R (%) | mAP50 (%) | mAP50-95 (%) |
|---|---|---|---|---|---|---|
| YOLOv5n | 5.03 | 7.2 | 44.2 | 33.7 | 32.5 | 18.1 |
| YOLOv8n | 5.97 | 8.9 | 44.7 | 33.9 | 32.7 | 18.7 |
| Faster R-CNN | 41.0 | 207.0 | 45.5 | 32.2 | 33.1 | 16.0 |
| Drone-YOLO | 16.4 | 19.0 | 52.8 | 41.5 | 43.0 | 26.0 |
| YOLOv10s | 8.04 | 24.5 | 50.5 | 38.3 | 39.1 | 23.5 |
| YOLOv11n | 2.58 | 6.3 | 43.8 | 33.3 | 33.3 | 19.4 |
| YOLOv11s | 9.41 | 21.3 | 50.5 | 38.7 | 39.3 | 23.6 |
| Ours (YOLOv11n-ours) | 2.72 | 13.4 | 50.0 | 39.6 | 40.7 | 24.4 |
Our proposed model, with only 2.72M parameters, achieves higher mAP50 (40.7%) than YOLOv11s (39.3%) and YOLOv10s (39.1%), while having about 71% fewer parameters than YOLOv11s. It also surpasses YOLOv5n and YOLOv8n by a large margin. Drone-YOLO achieves the highest mAP (43.0%) but uses 16.4M parameters, 6 times more than ours. Considering the trade-off between accuracy and efficiency, our model is particularly suitable for China UAV edge devices where computational resources are limited. We also note that our model maintains competitive recall (39.6%) and precision (50.0%), demonstrating robust detection capabilities.
To further illustrate the effectiveness, we provide a visual comparison of detection results between the baseline YOLOv11n and our improved model. The figure below shows that our model successfully detects many small pedestrians and vehicles that are missed by the baseline, especially in crowded or distant regions. The improvements in feature extraction and multi-scale fusion are clearly reflected in the reduced false negatives.
4.4 Category-wise Performance
We analyze the per-class mAP50 improvement of our model over the baseline. The most significant gains are observed in small categories such as “pedestrian” (from 27.1% to 35.2%), “bicycle” (from 14.3% to 20.8%), and “motor” (from 23.5% to 30.1%). This confirms that our multi-scale attention and feature pyramid are particularly beneficial for detecting tiny objects. The detailed breakdown is shown in the table below.
| Category | Baseline (%) | Ours (%) | Improvement (pp) |
|---|---|---|---|
| pedestrian | 27.1 | 35.2 | +8.1 |
| people | 19.8 | 25.6 | +5.8 |
| bicycle | 14.3 | 20.8 | +6.5 |
| car | 65.4 | 68.9 | +3.5 |
| van | 41.2 | 45.0 | +3.8 |
| truck | 36.7 | 40.1 | +3.4 |
| tricycle | 22.5 | 27.3 | +4.8 |
| awning-tricycle | 13.9 | 18.6 | +4.7 |
| bus | 56.8 | 59.2 | +2.4 |
| motor | 23.5 | 30.1 | +6.6 |
5. Conclusion
In this paper, we presented an improved YOLOv11n model tailored for small object detection in China UAV aerial imagery. By replacing the backbone’s C3k2 module with MSAM, we enhanced the network’s sensitivity to fine details. The proposed VMFPN neck effectively fused multi-scale features, preserving critical information for tiny objects. The Inner-ShapeIoU loss provided more accurate bounding box regression, especially for irregularly shaped small targets. Experimental results on the VisDrone2019 dataset demonstrated that our model achieved a mAP50 of 40.7% and mAP50-95 of 24.4%, outperforming the baseline YOLOv11n by 8.7 and 5.2 percentage points, respectively, while only increasing parameters slightly. Moreover, our model surpassed the larger YOLOv11s in accuracy, making it an ideal candidate for real-time deployment on resource-constrained China UAV edge computing platforms. Future work will explore model pruning and knowledge distillation to further reduce computational cost without sacrificing performance.