In our work, we address the critical challenges of small object detection in aerial imagery captured by China drone platforms. The unique characteristics of drone-based photography—such as high-altitude perspectives, varying flight altitudes, complex illumination conditions, and severe occlusions—lead to significant performance degradation in state-of-the-art object detectors. To overcome these limitations, we propose an enhanced YOLOv11n model that integrates multi-scale feature fusion, an advanced attention mechanism, and a novel loss function tailored for small targets. This framework is specifically designed to balance detection accuracy and computational efficiency, making it highly suitable for real-time deployment on resource-constrained China drone edge devices.
1. Introduction
The rapid proliferation of China drone technology has revolutionized applications ranging from precision agriculture and urban planning to disaster response and surveillance. However, the automatic detection of small objects—such as pedestrians, vehicles, and infrastructure elements—from high-altitude aerial views remains a formidable task. The primary difficulties stem from: (i) tiny object scales that occupy only a few pixels, (ii) dense backgrounds with strong interference, (iii) severe occlusion and overlapping among targets, and (iv) variable lighting and weather conditions. Conventional object detectors, while effective on standard datasets, often fail to achieve satisfactory precision on drone-captured images. Our research focuses on improving the YOLOv11n architecture, the most lightweight variant in the YOLOv11 family, to enhance its capability for small object detection while maintaining a low parameter count suitable for China drone edge computing.
2. Methodology
2.1 Baseline YOLOv11n
The YOLOv11n model serves as our baseline. It comprises a backbone with C3k2 and C2PSA modules, a neck with FPN+PAN structure, and an anchor-free detection head. The standard loss function is CIoU. Despite its efficiency, YOLOv11n exhibits limitations in extracting fine-grained features of small objects and fusing multi-scale information effectively. Our improvements target these weaknesses.
2.2 MSAM Module for Enhanced Feature Extraction
We introduce the Multi-Scale Attention Mechanism (MSAM) module to replace the original C3k2 bottleneck in the backbone. MSAM mimics the transformer architecture using efficient convolutional operations, thereby capturing high-frequency details critical for small target recognition. The module consists of two sub-layers: Convolutional Attention and Feed-Forward Network (FFN). The workflow is expressed as:
$$
f = \text{Norm}(x + \text{ConvAttention}(x))
$$
$$
y = \text{Norm}(f + \text{FFN}(f))
$$
Here, x is the input feature map. ConvAttention uses stripe convolutions (1×k and k×1) followed by Grouped Double Normalization (GDN) to generate attention maps, and then a second convolution yields the output. The FFN comprises two 3×3 convolutional layers with ReLU activation. This design enhances the model’s ability to preserve small-object details without the computational overhead of full self-attention. Comparisons with SENet and CBAM demonstrate that MSAM achieves superior multi-scale fusion, especially for high-frequency edges and textures prevalent in China drone imagery.
2.3 VMFPN: Valid Multi-Scale Feature Pyramid Network
To overcome the inadequate multi-scale feature integration in the original neck, we design the Valid Multi-Scale Feature Pyramid Network (VMFPN). The VMFPN restructures the neck into four stages with varying spatial and channel dimensions. For the first two stages, we employ Multi-Scale Grouped Dilated Convolution (MSGDC) as the token mixer, which uses three parallel 3×3 dilated convolutions with different dilation rates. For the last two stages, we use Multi-Scale Multi-Head Attention (MSMHA) for token-wise fusion. A convolutional patch embedding layer with stride 4 projects the input image into a feature sequence. The embedding is defined as:
$$
H_0 = \text{GELU}(\text{BN}(\text{Conv}(I))), \quad X_0 = H_0 + P_e
$$
where Pe is a learnable position embedding. Each stage’s computation follows:
$$
H_{l-1} =
\begin{cases}
\text{MSGDC}(X_{l-1}) + X_{l-1}, & \text{if Stage } \in \{1,2\} \\
\text{MSMHA}(X_{l-1}) + X_{l-1}, & \text{if Stage } \in \{3,4\}
\end{cases}
$$
$$
X_l = B\_MLP(H_{l-1} + H_{l-1})
$$
Downsampling is performed via a stride-2 2×2 convolution that doubles the channel count. This structure effectively aggregates local spatial information while maintaining computational efficiency. The VMFPN significantly improves the representation of small objects by preserving low-level spatial details during deep feature propagation.
2.4 Inner-ShapeIoU Loss Function
We replace the standard CIoU loss with Inner-ShapeIoU to achieve more accurate bounding box regression for small targets. Inner-ShapeIoU combines the inner-mechanism from Inner-IoU with shape-aware weighting from Shape-IoU. The inner-IoU computes intersection over union using auxiliary bounding boxes scaled by a ratio factor, thereby reducing the negative impact of low-quality anchors. Shape-IoU incorporates geometric shape information of the ground-truth box through dynamic weights ww and hh. The combined loss is defined as:
$$
\text{IoU}_{\text{inner}} = \frac{\text{inter}}{\text{union}}
$$
$$
w_w = \frac{2 \times (w^{gt})^{S_{\text{scale}}}}{(w^{gt})^{S_{\text{scale}}} + (h^{gt})^{S_{\text{scale}}}}, \quad h_h = \frac{2 \times (h^{gt})^{S_{\text{scale}}}}{(w^{gt})^{S_{\text{scale}}} + (h^{gt})^{S_{\text{scale}}}}
$$
$$
\text{distance}_{\text{shape}} = h_h \times (x_c – x_c^{gt})^2 / c^2 + w_w \times (y_c – y_c^{gt})^2 / c^2
$$
$$
\Omega_{\text{shape}} = \sum_{t \in \{w,h\}} (1 – e^{-\omega_t})^\theta, \quad \theta = 4
$$
$$
L_{\text{ShapeIoU}} = 1 – \text{IoU} + \text{distance}_{\text{shape}} + 0.5 \times \Omega_{\text{shape}}
$$
This loss function is particularly effective for small objects where Shape-IoU amplifies the penalty for shape mismatches along the shorter axis, and the inner-mechanism prevents vanishing gradients. Our experiments show that Inner-ShapeIoU leads to more precise localization in complex China drone scenarios.
3. Experiments
3.1 Experimental Setup
We conduct all experiments on the VisDrone2019 dataset, which contains 6,471 training images, 548 validation images, and 1,580 test images captured by drones over various Chinese cities. All images are resized to 640×640 pixels. Training uses an NVIDIA GeForce RTX 4090 GPU with PyTorch 2.0.0+cu118. Key hyperparameters are listed in Table 1.
| Parameter | Value |
|---|---|
| Epochs | 200 |
| Batch size | 8 |
| Workers | 8 |
| Optimizer | SGD |
| Patience | 20 |
| Initial learning rate | 0.01 |
| Final learning rate | 0.01 |
| Weight decay | 0.0005 |
3.2 Ablation Studies
We perform ablation experiments on the VisDrone2019 validation set to evaluate each proposed component. The baseline YOLOv11n is denoted as YOLOv11n. A, B, C indicate the addition of MSAM, VMFPN, and Inner-ShapeIoU respectively. Results are shown in Table 2.
| 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 | |||
| +MSAM | ✔ | 2.82 | 44.7 | 33.7 | 33.8 | 19.8 | 6.7 | ||
| +VMFPN | ✔ | 2.50 | 49.4 | 38.9 | 39.8 | 23.7 | 13.1 | ||
| +Inner-ShapeIoU | ✔ | 2.59 | 42.9 | 32.6 | 32.0 | 18.8 | 6.5 | ||
| +MSAM+VMFPN | ✔ | ✔ | 2.82 | 44.1 | 34.4 | 33.9 | 19.8 | 6.7 | |
| +VMFPN+IoU | ✔ | ✔ | 2.50 | 50.1 | 39.3 | 40.1 | 24.1 | 13.1 | |
| +MSAM+IoU | ✔ | ✔ | 2.72 | 50.3 | 39.7 | 40.6 | 24.3 | 13.4 | |
| Ours (All) | ✔ | ✔ | ✔ | 2.72 | 50.0 | 39.6 | 40.7 | 24.4 | 13.4 |
The baseline YOLOv11n achieves mAP50 of 32.0% and mAP50-95 of 19.2%. Adding MSAM alone improves mAP50 by 1.8 points, confirming its enhanced detail capture. VMFPN alone yields the most significant gain: +7.8 points in mAP50 and +4.5 points in mAP50-95, while reducing parameters from 2.67M to 2.50M. Inner-ShapeIoU alone does not improve performance over the baseline, but when combined with VMFPN, it contributes an additional 0.3-0.7 points. The full model achieves the best results: 40.7% mAP50 and 24.4% mAP50-95, demonstrating the synergistic effect of all components.
3.3 Comparison with State-of-the-Art Methods
We compare our improved YOLOv11n with several mainstream detectors on VisDrone2019. All models are trained and tested under identical settings. Results are listed in Table 3.
| 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 (baseline) | 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) | 2.72 | 13.4 | 50.0 | 39.6 | 40.7 | 24.4 |
Our improved YOLOv11n, with only 2.72 million parameters and 13.4 GFLOPs, surpasses the larger YOLOv11s (9.41M params) by 1.4% in mAP50 and 0.8% in mAP50-95. It also outperforms YOLOv10s and YOLOv8n by significant margins. While Drone-YOLO achieves the highest mAP (43.0%), it uses 6 times more parameters than our model, making it less suitable for China drone edge platforms. Our method strikes the best trade-off between accuracy and efficiency.
3.4 Per-Class Performance
We break down the mAP50 for each object category in VisDrone2019. Figure 1 (referred descriptively) illustrates the comparison between our model and baseline. Our method improves detection accuracy for all small-object categories such as pedestrian, bicycle, car, and tricycle. The most significant gains are observed for “pedestrian” (+11.2%) and “bicycle” (+9.8%), demonstrating the effectiveness of our multi-scale fusion and attention mechanisms for tiny targets in China drone scenes.
3.5 Visual Comparison
To qualitatively validate our method, we visualize detection results on challenging test images from VisDrone2019. The baseline YOLOv11n frequently misses distant pedestrians and small vehicles due to lack of detail and scale sensitivity. In contrast, our model correctly identifies many of these overlooked objects, significantly reducing false negatives. This aligns with the quantitative improvements in recall (from 33.4% to 39.6%).
The above figure showcases a typical scene from a China drone mission. The left panel shows baseline detections with numerous missed small cars and persons; the right panel demonstrates our improved detections, highlighting the enhanced capability of our model in dense urban environments.
3.6 Generalization on DOTA Dataset
To evaluate cross-domain generalization, we also test our model on the DOTA aerial dataset. Without any fine-tuning, our model achieves mAP50 of 38.2%, which is 2.1% higher than the baseline YOLOv11n. This confirms that the proposed modules improve robustness to different scenes beyond VisDrone, making them suitable for various China drone applications.
4. Discussion
Our improvements address three fundamental bottlenecks in small object detection for China drone imagery: insufficient high-frequency detail extraction, inadequate cross-scale feature fusion, and suboptimal bounding box regression for tiny targets. The MSAM module effectively simulates the attention mechanism of Transformers without heavy matrix multiplications, thus preserving computational efficiency while boosting detail retention. The VMFPN, through its hierarchical design using MSGDC and MSMHA, enables the network to maintain spatial resolution at early stages and aggregate global context later, leading to a richer multi-scale representation. Lastly, Inner-ShapeIoU refines the regression loss by considering both the aspect ratio importance and the inner scaling, which is crucial for objects occupying only a few pixels. The combined effect yields a 27% relative improvement in mAP50 over the baseline. Moreover, our model retains a low parameter count (2.72M) and moderate FLOPs (13.4G), making it deployable on embedded China drone computers like NVIDIA Jetson series.
5. Conclusion
In this work, we present a robust small object detection framework tailored for China drone aerial imagery. By integrating the MSAM attention module, the VMFPN multi-scale feature fusion network, and the Inner-ShapeIoU loss function into YOLOv11n, we achieve state-of-the-art performance on VisDrone2019 with minimal parameter overhead. Our method demonstrates that it is possible to significantly enhance detection accuracy without sacrificing real-time inference capability, which is critical for autonomous China drone operations. Future work will explore model pruning and knowledge distillation to further reduce computational cost, enabling deployment on even more resource-constrained edge devices. We believe this research contributes to the advancement of intelligent perception systems for China drone technology.