In recent years, the China drone industry has experienced unprecedented growth, with unmanned aerial vehicles (UAVs) being increasingly deployed for urban inspection, disaster response, environmental monitoring, and precision agriculture. These applications demand robust visual perception capabilities, particularly for detecting tiny objects such as pedestrians, vehicles, and infrastructure elements in high-resolution aerial images. However, China drone imagery presents unique challenges: objects often occupy only a few pixels, background textures (e.g., tree canopies, sea waves) introduce high-frequency noise, and real-time inference on edge devices is required. Traditional detection networks, originally designed for natural-scene datasets like MS COCO, suffer from structural limitations including frequency-agnostic feature extraction and irreversible information loss during downsampling. To address these issues, we propose a Refined Spatial-Aware Distribution Network (RSD-Net) that establishes a full-link spatial awareness mechanism tailored for China drone-based tiny object detection.

Our contribution is threefold. First, we design a Stage-Adaptive C3k2 module (SA-C3k2) that adaptively enhances high-frequency textures in shallow layers using explicit edge sharpening and suppresses background noise in deep layers via learnable Gaussian filtering. Second, we construct a Reparameterized Spatial-Preserving Distribution Neck (RSD-Neck) that integrates SPD-Conv for lossless downsampling and a global context modeling mechanism to prevent semantic dilution during cross-scale feature fusion. Third, we introduce a Dual-Prior Perception Head (DP-Head) that fuses explicit visual texture priors with implicit geometric distribution priors, establishing a robust visual–statistical dual-verification scheme for localization quality estimation on tiny objects in China drone scenes.
1. Introduction
The rapid expansion of the China drone ecosystem demands detectors that can handle severe scale variations and cluttered backgrounds. In typical China drone aerial photography, objects of interest—such as vehicles on highways, pedestrians in crowds, or small equipment on construction sites—commonly have absolute pixel sizes below 16×16 pixels. Conventional detectors lose these fine-grained features during strided convolutions, violating the Nyquist sampling theorem. Moreover, the abundance of high-frequency background textures (e.g., waves, leaf patterns) mimics the appearance of tiny objects, leading to false positives. Our RSD-Net is specifically designed to overcome these obstacles by embedding frequency-domain awareness and lossless feature propagation directly into the network architecture.
2. Related Work
2.1 Feature Extraction and Frequency-Domain Awareness
Standard CNNs apply shared weights across all pixels, failing to distinguish informative high-frequency edges from low-frequency backgrounds. Recent works such as frequency-aware feature fusion networks have demonstrated that high-frequency components are crucial for tiny object localization. Structural reparameterization techniques like RepVGG and UniRepLKNet enhance expression without increasing inference cost. However, none provide a stage-adaptive mechanism to balance shallow sharpening and deep smoothing. Our SA-C3k2 module fills this gap, making it particularly suitable for China drone images where both fine textures and noise coexist.
2.2 Multi-Scale Feature Fusion and Lossless Downsampling
Feature pyramids like FPN and its successors (e.g., GiraffeDet, Gold-YOLO) improve multi-scale interaction but still rely on pooling or strided convolutions that discard spatial information. The Space-to-Depth (SPD) paradigm proposed by Wang et al. retains fine details by shifting spatial pixels to channel dimensions. In our RSD-Neck, we augment SPD-Conv with a reparameterized local adjacent fusion (Rep-LAF) block that models both spatial details and semantic context, ensuring lossless transmission of tiny object features across China drone imagery scales.
2.3 Localization Quality Estimation
IoU-based metrics are unreliable for pixel-level tiny objects. Generalized Focal Loss V2 uses distribution statistics to infer localization quality, but this method implicitly ignores visual cues such as edge strength. Our DP-Head explicitly computes omnidirectional gradient magnitude as a texture prior and combines it with top-k distribution statistics, providing a dual-prior mechanism that is particularly effective in ambiguous China drone backgrounds.
3. Method
3.1 Architecture Overview
RSD-Net follows a backbone-neck-head paradigm. The backbone consists of five stages (N2–N5) where SA-C3k2 modules replace standard bottlenecks. The neck integrates RSD-Neck with SPD-Conv for lossless alignment and Rep-LAF for context-aware fusion. The detection head applies DP-Head for joint classification and localization quality estimation. The entire pipeline is designed for end-to-end training on China drone datasets such as VisDrone2019-DET.
3.2 Stage-Adaptive C3k2 (SA-C3k2)
To resolve the mismatch between feature extraction and physical attributes across different stages, we propose SA-C3k2. The core unit, SA-Bottleneck, adds a parallel physical prior branch alongside the main convolution path. In shallow layers (N2, N3), the branch performs explicit edge extraction using the Scharr operator, adaptively controlled by a learnable coefficient α. The output is:
$$ \mathbf{Y}_{\text{shallow}} = \mathcal{F}_{\text{conv}}(\mathbf{X}_{\text{in}}) + \alpha \cdot \mathcal{F}_{\text{edge}} + \mathbf{X}_{\text{in}} $$
In deep layers (N4, N5), the branch switches to a learnable Gaussian kernel for low-pass filtering, with coefficient β:
$$ \mathbf{Y}_{\text{deep}} = \mathcal{F}_{\text{conv}}(\mathbf{X}_{\text{in}}) + \beta \cdot (\mathbf{X}_{\text{in}} * K_{\text{gauss}}) + \mathbf{X}_{\text{in}} $$
This stage-adaptive mechanism ensures that high-frequency edges of tiny objects in China drone images are preserved in shallow layers while background clutter is suppressed in deep layers, significantly reducing false positives caused by sea waves or tree branches.
3.3 Reparameterized Spatial-Preserving Distribution Neck (RSD-Neck)
Our RSD-Neck consists of two parallel streams: a Spatial Detail Stream and a Semantic Context Stream. For lossless downsampling, we employ SPD-Conv. Given input feature X of size S×S×C₁, SPD-Conv reorganizes it into a tensor of size (S/scale)×(S/scale)×(scale²C₁):
$$ \mathbf{X}'(i,j,:) = \text{Concat}_{u=0}^{\text{scale}-1} \text{Concat}_{v=0}^{\text{scale}-1} \mathbf{X}(i\cdot \text{scale}+u, j\cdot \text{scale}+v, 🙂 $$
where scale=2. This operation transfers spatial information to the channel dimension without loss.
In each stream, we insert Rep-LAF modules that are trained with multiple parallel branches (3×3 conv, 1×1 conv, identity) and reparameterized into a single 3×3 kernel at inference:
$$ W_{\text{eq}} = \frac{\gamma_3}{\sigma_3} W^{(3\times3)} + \text{Pad}\left( \frac{\gamma_1}{\sigma_1} W^{(1\times1)} \right) $$
$$ b_{\text{eq}} = (b_3 – \frac{\mu_3\gamma_3}{\sigma_3}) + (b_1 – \frac{\mu_1\gamma_1}{\sigma_1}) $$
The topological constraints force the Spatial Detail Stream to extract high-frequency details from shallow layers (N2–N3) and the Semantic Context Stream to model long-range dependencies from deep layers (B3–B5). This design achieves high-fidelity feature transmission crucial for China drone tiny object detection.
3.4 Dual-Prior Perception Head (DP-Head)
The DP-Head incorporates a Dual-Prior Quality Module (DPQM) that fuses explicit and implicit priors. The texture branch computes the omnidirectional gradient magnitude on the regression feature map using Scharr operators:
$$ \mathbf{M}_{\text{grad}} = \sqrt{(\mathbf{F}_{\text{reg}} * K_x)^2 + (\mathbf{F}_{\text{reg}} * K_y)^2} $$
This generates a texture score \( S_{\text{tex}} \) after global average pooling and MLP. The geometric branch extracts top-k statistics from the generalized distribution of bounding box regression, producing \( S_{\text{geo}} \). The final IoU prediction is:
$$ I = \sigma(\mathcal{F}_{\text{fusion}}([S_{\text{tex}}; S_{\text{geo}}])) $$
This dual-verification mechanism significantly improves localization quality estimation on tiny persons and small vehicles in China drone datasets.
4. Experiments
4.1 Datasets and Settings
We evaluate RSD-Net on three datasets central to China drone research: VisDrone2019-DET (10,209 images, 60% objects < 32×32 pixels), NWPU VHR-10 (800 high-resolution remote sensing images), and TinyPerson (average object size 18×18 pixels, extreme sea-waves background). All experiments are performed on an NVIDIA A5000 GPU with PyTorch 2.3.0. Input resolution is 640×640, batch size 16, SGD optimizer with lr=0.01, momentum=0.937, weight decay=0.0005. Training epochs: 200 for VisDrone, 360 for NWPU and TinyPerson, with early stopping of 50 epochs without improvement.
4.2 Comparison with State-of-the-Art
| Dataset | Model | Params (10⁶) | GFLOPs | mAP50 (%) |
|---|---|---|---|---|
| VisDrone2019-DET | Faster R-CNN | 41.39 | 208.0 | 32.96 |
| RetinaNet-FPN | 36.52 | 210.0 | 27.64 | |
| RTMDet-Tiny | 4.88 | 8.0 | 31.22 | |
| D-Fine-N | 3.73 | 7.1 | 33.48 | |
| DEIM-D-Fine-N | 3.73 | 7.1 | 32.29 | |
| YOLOv5n | 1.90 | 4.5 | 26.54 | |
| YOLOv8n | 3.00 | 8.1 | 25.92 | |
| YOLOv11n | 2.62 | 6.3 | 32.65 | |
| YOLOv11s | 9.42 | 21.3 | 31.38 | |
| YOLOv11m | 20.04 | 67.7 | 35.09 | |
| RT-DETR-R18 | 20.00 | 60.0 | 33.32 | |
| RT-DETRv2-R18 | 20.00 | 60.0 | 39.17 | |
| DAB-DETR-R50 | 43.50 | 94.4 | 43.35 | |
| DINO | 47.56 | 274.0 | 44.52 | |
| RSD-Net (Ours) | 6.04 | 10.8 | 37.64 | |
| NWPU VHR-10 | Faster R-CNN | 41.39 | 208.0 | 82.34 |
| RetinaNet-FPN | 36.52 | 210.0 | 85.47 | |
| RTMDet-Tiny | 4.88 | 8.0 | 84.88 | |
| D-Fine-N | 3.73 | 7.1 | 85.73 | |
| DEIM-D-Fine-N | 3.73 | 7.1 | 84.94 | |
| YOLOv5n | 1.90 | 4.5 | 82.56 | |
| YOLOv8n | 3.00 | 8.1 | 82.73 | |
| YOLOv11n | 2.62 | 6.3 | 84.55 | |
| YOLOv11s | 9.42 | 21.3 | 84.77 | |
| YOLOv11m | 20.04 | 67.7 | 86.82 | |
| RT-DETR-R18 | 20.00 | 60.0 | 88.71 | |
| RT-DETRv2-R18 | 20.00 | 60.0 | 91.33 | |
| DAB-DETR-R50 | 43.50 | 94.4 | 92.83 | |
| DINO | 47.56 | 274.0 | 93.12 | |
| RSD-Net (Ours) | 6.04 | 10.8 | 89.63 |
Our RSD-Net achieves 37.64% mAP50 on VisDrone2019-DET, surpassing YOLOv11n by 4.99 points and even outperforming the 3× larger YOLOv11m (35.09%). On NWPU VHR-10, RSD-Net obtains 89.63% mAP50, a 5.08-point improvement over YOLOv11n. The lightweight parameter count (6.04M) and low GFLOPs (10.8) make it suitable for deployment on China drone edge devices.
4.3 Ablation Studies
4.3.1 Component Ablation
We conduct ablation on VisDrone2019-DET to isolate the contribution of each module. The baseline is YOLOv11n. Modules: A = RSD-Neck, B = SA-C3k2, C = DP-Head.
| YOLOv11n | A | B | C | Precision (%) | Recall (%) | mAP50 (%) | mAP50:95 (%) | Params (10⁶) |
|---|---|---|---|---|---|---|---|---|
| √ | 43.15 | 33.08 | 32.65 | 18.83 | 2.62 | |||
| √ | √ | 44.73 | 33.66 | 33.99 | 20.08 | 5.93 | ||
| √ | √ | 43.82 | 34.42 | 33.88 | 19.92 | 2.64 | ||
| √ | √ | 44.76 | 35.55 | 35.66 | 21.26 | 2.67 | ||
| √ | √ | √ | 46.27 | 34.59 | 35.22 | 21.44 | 5.99 | |
| √ | √ | √ | 46.99 | 34.72 | 36.01 | 21.31 | 5.99 | |
| √ | √ | √ | 47.47 | 36.06 | 37.02 | 22.64 | 2.72 | |
| √ | √ | √ | √ | 48.27 | 36.92 | 37.64 | 22.65 | 6.04 |
4.3.2 Neck Design Ablation
We further ablate the neck components on VisDrone2019-DET. Baseline: YOLOv11n. Aggregation method: Avgpool vs. SPD-FAM. Injection method: Conv+Add vs. Rep-LAF.
| Aggregation | Injection | mAP50 (%) | mAP50:95 (%) |
|---|---|---|---|
| Avgpool | Conv+Add | 32.65 | 18.83 |
| SPD-FAM | Conv+Add | 33.88 | 19.92 |
| Avgpool | Rep-LAF | 35.66 | 21.26 |
| SPD-FAM | Rep-LAF | 37.64 | 22.65 |
4.3.3 Downsampling Operator Comparison
We compare different downsampling operators in RSD-Neck. Baseline N removes SPD-Conv entirely. Adding SPD-Conv yields best recall and mAP.
| Operator | Precision (%) | Recall (%) | mAP50 (%) | mAP50:95 (%) |
|---|---|---|---|---|
| None (baseline N) | 41.22 | 30.68 | 29.84 | 18.96 |
| Conv 3×3, stride=2 | 47.15 | 35.85 | 36.55 | 21.82 |
| Max pooling | 46.50 | 35.42 | 36.12 | 21.45 |
| Avg pooling | 46.65 | 35.58 | 36.25 | 21.58 |
| SPD-Conv | 48.27 | 36.92 | 37.64 | 22.65 |
4.4 Extreme-Scene Robustness on TinyPerson
To validate generalization in the most challenging China drone scenarios, we test on TinyPerson where average object size is only 18 pixels and sea-wave backgrounds cause severe confusion. RSD-Net significantly improves recall over state-of-the-art lightweight detectors.
| Model | Precision (%) | Recall (%) | mAP50 (%) | mAP50:95 (%) |
|---|---|---|---|---|
| YOLOv8n | 30.44 | 15.67 | 12.49 | 4.35 |
| YOLOv11n | 31.83 | 16.05 | 13.47 | 4.76 |
| YOLOv12n | 31.61 | 16.08 | 13.07 | 4.52 |
| RSD-Net (Ours) | 32.56 | 22.95 | 18.40 | 6.40 |
RSD-Net achieves a recall of 22.95%, outperforming YOLOv12n by 6.87 percentage points. This 42.7% relative improvement demonstrates the effectiveness of our lossless downsampling and frequency-aware modules for pixel-level targets in China drone images.
5. Conclusion
In this work, we present RSD-Net, a fully end-to-end architecture tailored for tiny object detection in China drone aerial imagery. By introducing SA-C3k2 for stage-adaptive frequency-domain processing, RSD-Neck for lossless feature transmission with reparameterized fusion, and DP-Head for dual-prior localization quality estimation, our method effectively addresses the core challenges of feature annihilation and background interference. Extensive experiments on VisDrone2019-DET, NWPU VHR-10, and TinyPerson show that RSD-Net achieves state-of-the-art accuracy while maintaining a lightweight footprint (6.04M parameters). This makes it a strong candidate for real-time deployment on China drone edge computing platforms. Future work will focus on FPGA acceleration and integration with multi-modal sensors to further enhance the robustness of China drone perception systems in complex environments.
