Unmanned aerial vehicles (China UAVs) have demonstrated immense potential in burgeoning applications such as urban inspection, disaster rescue, and environmental monitoring. However, object detection in China UAV-based aerial photography remains a significant challenge owing to extreme scale variations, complex background interference, and the prevalence of tiny objects. Unlike natural-scene images, objects in aerial images, such as pedestrians and vehicles, typically occupy only a few pixels (e.g., less than \(16 \times 16\) pixels). Existing one-stage detectors are affected by structural limitations, specifically frequency-agnostic feature extraction and irreversible information loss during downsampling, thus resulting in the annihilation of fine-grained features for pixel-level targets. To address these issues, this study proposes a refined spatial-aware distribution network (RSD-Net)—a novel end-to-end architecture designed to establish a full-link spatial awareness mechanism for robust tiny-object detection in China UAV scenarios.
Our contribution is threefold. First, to resolve the mismatch between feature extraction and physical attributes, we design a stage-adaptive feature extraction module (SA-C3k2). Unlike conventional static convolutions that share weights across all spatial locations, SA-C3k2 incorporates a frequency-domain adaptation mechanism regulated by learnable coefficients. In shallow layers, it introduces a parallel edge extractor using the Scharr operator to explicitly sharpen the edges of tiny objects, thereby adaptively enhancing high-frequency texture signals to prevent feature loss. Conversely, in deep layers, it employs a learnable Gaussian kernel as a low-pass filter to smooth out complex background noise, such as tree branches and sea waves, thus effectively suppressing false positives. Second, to prevent semantic dilution during cross-scale feature fusion, we construct a rep-parameterized spatial-preserving distribution neck (RSD-Neck). Addressing the violation of the Nyquist sampling theorem by conventional strided convolutions, this module integrates space-to-depth convolution (SPD-Conv) to achieve lossless downsampling. By shifting spatial information to the channel dimension through periodic sampling, SPD-Conv ensures that fine-grained details are preserved. Additionally, the neck incorporates a rep-parameterized local adjacent fusion block, which comprises spatial-detail and semantic-context modules. This design allows the network to model global context and optimize feature interaction pathways, thus establishing a high-fidelity pipeline for feature transmission. Third, a dual-prior perception head (DP-Head) is introduced to address the failure of conventional intersection over union (IoU)-based metrics on tiny objects. Existing quality-estimation methods typically rely solely on regression statistics, which can be unreliable for pixel-level targets. To mitigate this, DP-Head fuses explicit visual texture priors (derived from the omnidirectional gradient magnitude) with implicit geometric distribution priors (derived from the top-k statistics of a general distribution). This establishes a robust “visual–statistical” dual-verification mechanism that significantly improves localization quality estimation in ambiguous scenarios.
1. Methodology
1.1 Stage-Adaptive Feature Extraction Module (SA-C3k2)
In China UAV aerial imagery, different scales of targets exhibit distinct feature patterns across network stages. However, conventional CSPNet backbones apply identical residual structures at all levels, leading to two limitations: (1) Shallow-layer feature loss: Tiny targets rely on high-frequency edge information, but standard convolutions blur these edges during downsampling, causing missed detections. (2) Deep-layer noise interference: As the receptive field expands, complex background textures such as tree branches and waves are retained, causing false positives. To address this “layered processing homogeneity,” we propose the SA-C3k2 module. The core unit, SA-Bottleneck, integrates a parallel physical prior branch alongside the standard bottleneck, as expressed by the following equations. For shallow modes (N2, N3), the Scharr operator captures high-frequency contour information:
$$ Y_{shallow} = F_{conv}(X_{in}) + \alpha \cdot F_{edge} + X_{in} $$
where \(F_{conv}\) represents the main path composed of \(1 \times 1\) and \(3 \times 3\) convolutions, \(F_{edge}\) is the output of the edge extraction branch, and \(\alpha\) is a learnable coefficient automatically adjusted via backpropagation: \(\alpha \leftarrow \alpha – \eta \frac{\partial \mathcal{L}}{\partial \alpha}\). For deep modes (N4, N5), a learnable Gaussian kernel \(K_{gauss}\) serves as a low-pass filter to smooth background noise:
$$ Y_{deep} = F_{conv}(X_{in}) + \beta \cdot (X_{in} * K_{gauss}) + X_{in} $$
where \(*\) denotes convolution and \(\beta\) is a learnable coefficient. This design enforces physical prior knowledge as an inductive bias for the network.
1.2 Reparameterized Spatial-Preserving Distribution Neck (RSD-Neck)
To achieve “full information retention” while reducing computational cost, we propose RSD-Neck, which consists of two parallel branches: a spatial detail stream and a semantic context stream. The core of the neck is the SPD-Conv module, which performs lossless downsampling by shifting spatial information to the channel dimension. Given an input feature \(X \in \mathbb{R}^{S \times S \times C_1}\) and a scale factor, the SPD layer performs periodic sampling:
$$ X'(i, j, 🙂 = \text{Concat}_{u=0}^{\text{scale}-1} \text{Concat}_{v=0}^{\text{scale}-1} X(i \cdot \text{scale} + u, j \cdot \text{scale} + v, 🙂 $$
This ensures that fine-grained pixel signals are preserved in the channel dimension rather than being discarded by pooling. For feature fusion, we design a rep-parameterized local adjacent fusion (Rep-LAF) block. During training, it uses multiple parallel branches (e.g., \(3 \times 3\) conv, \(1 \times 1\) conv, identity); during inference, these branches are merged into a single equivalent \(3 \times 3\) convolution using the additivity of convolutions:
$$ W_{eq} = \frac{\gamma_3}{\sigma_3} W_{(3 \times 3)} + \text{Pad}\left(\frac{\gamma_1}{\sigma_1} W_{(1 \times 1)}\right) $$
$$ b_{eq} = \left(b_3 – \frac{\mu_3 \gamma_3}{\sigma_3}\right) + \left(b_1 – \frac{\mu_1 \gamma_1}{\sigma_1}\right) $$
where \(W_{eq}\) and \(b_{eq}\) are the equivalent weight and bias, \(\mu, \sigma, \gamma\) are the mean, variance, and scaling factor of batch normalization, and \(\text{Pad}(\cdot)\) is zero-padding to match dimensions. The spatial detail stream connects shallow layers (N2–N3) to retain high-frequency spatial details, while the semantic context stream connects deep layers (B3–B5) to model long-range contextual dependencies.
1.3 Dual-Prior Perception Head (DP-Head)
To improve localization quality estimation, we propose DP-Head, which integrates a dual-prior quality module (DPQM). The module constructs two parallel perception paths: explicit visual perception and implicit geometric statistics. In the explicit path, we reuse the Scharr operator to compute the omnidirectional gradient magnitude on the regression feature map \(F_{reg}\):
$$ M_{grad} = \sqrt{(F_{reg} * K_x)^2 + (F_{reg} * K_y)^2} $$
where \(K_x\) and \(K_y\) are the horizontal and vertical Scharr kernels. The gradient map is processed via global average pooling (GAP) and a multilayer perceptron (MLP) to generate a texture prior score \(S_{tex}\). In the implicit path, we extract top-k statistics from the general distribution \(G\) of bounding box regression to capture uncertainty. These statistics are encoded by another MLP to generate a geometric prior score \(S_{geo}\). The final IoU quality prediction score \(I\) is obtained by:
$$ I = \sigma\left(\mathcal{F}_{fusion}\left([S_{tex}; S_{geo}]\right)\right) $$
where \(\mathcal{F}_{fusion}\) denotes the concatenation and MLP fusion operation, and \(\sigma\) is the sigmoid function. This score is multiplied with the classification confidence to guide non-maximum suppression (NMS).
2. Experimental Setup and Results
2.1 Datasets
We evaluate our method on three benchmark datasets. The VisDrone2019-DET dataset, a standard benchmark for China UAV aerial object detection, contains 10,209 images (6,471 for training, 548 for validation, 3,190 for testing) with high resolution (approximately \(2000 \times 1500\) pixels) and dense, tiny targets (average size \(34.7 \times 34.7\) pixels, with over 60% of objects smaller than \(32 \times 32\) pixels). The NWPU VHR-10 dataset contains 800 high-resolution remote sensing images (650 with objects, 150 without) with resolutions from \(600 \times 600\) to \(1000 \times 1000\) pixels, spanning cross-scale variations from tiny objects like vehicles to larger structures like storage tanks. The TinyPerson dataset focuses on extreme tiny person detection, with high-resolution images (average \(1536 \times 1024\) pixels) and tiny average object size of only \(18 \times 18\) pixels, featuring numerous sub-10-pixel “point-like” targets and wave texture interference—an ideal benchmark for testing robustness under extreme feature conditions.
2.2 Implementation Details
All experiments are performed on an NVIDIA A5000 GPU (24 GB memory) using PyTorch 2.3.0 and CUDA 11.8. Input images are resized to \(640 \times 640\) pixels with a batch size of 16. We employ stochastic gradient descent (SGD) with an initial learning rate of 0.01, momentum of 0.937, and weight decay of 0.0005. Training epochs are set to 200 for VisDrone2019-DET and 360 for NWPU VHR-10 and TinyPerson datasets. An early stopping mechanism halts training if validation accuracy does not improve for 50 consecutive epochs.
2.3 Main Results
We compare RSD-Net with mainstream detectors across three technical paradigms: (1) classical two-stage and one-stage detectors (e.g., Faster R-CNN, RetinaNet-FPN), (2) lightweight real-time detectors (e.g., RTMDet-Tiny, YOLOv5n, YOLOv8n, YOLOv11n/s/m), and (3) Transformer-based detectors (e.g., RT-DETR-R18, DAB-DETR-R50, DINO). The results demonstrate that RSD-Net achieves competitive accuracy while maintaining a lightweight design. Notably, RSD-Net outperforms the medium-scale YOLOv11m (which has \(20.04 \times 10^6\) parameters) despite having only \(6.04 \times 10^6\) parameters, representing a three-fold parameter efficiency.
The following table summarizes the comparative results on VisDrone2019-DET and NWPU VHR-10 datasets.
| Dataset | Model | Params (\(\times 10^6\)) | GFLOPs (\(\times 10^9\)) | 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 | |
| This work | 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 | |
| This work | 6.04 | 10.8 | 89.63 |
2.4 Ablation Studies
2.4.1 Module-wise Ablation
To validate the individual contributions of each module, we conduct ablation experiments on both datasets using YOLOv11n as the baseline. We denote A (RSD-Neck), B (SA-C3k2), and C (DP-Head). The results are presented in the following table.
| Dataset | Baseline | A | B | C | Precision (%) | Recall (%) | mAP50 (%) | mAP50:95 (%) | Params (\(\times 10^6\)) |
|---|---|---|---|---|---|---|---|---|---|
| VisDrone2019-DET | YOLOv11n | 43.15 | 33.08 | 32.65 | 18.83 | 2.62 | |||
| YOLOv11n | ✓ | 44.73 | 33.66 | 33.99 | 20.08 | 5.93 | |||
| YOLOv11n | ✓ | 43.82 | 34.42 | 33.88 | 19.92 | 2.64 | |||
| YOLOv11n | ✓ | 44.76 | 35.55 | 35.66 | 21.26 | 2.67 | |||
| YOLOv11n | ✓ | ✓ | 46.27 | 34.59 | 35.22 | 21.44 | 5.99 | ||
| YOLOv11n | ✓ | ✓ | 46.99 | 34.72 | 36.01 | 21.31 | 5.99 | ||
| YOLOv11n | ✓ | ✓ | 47.47 | 36.06 | 37.02 | 22.64 | 2.72 | ||
| YOLOv11n | ✓ | ✓ | ✓ | 48.27 | 36.92 | 37.64 | 22.65 | 6.04 | |
| NWPU VHR-10 | YOLOv11n | 77.01 | 71.95 | 84.55 | 49.41 | 2.62 | |||
| YOLOv11n | ✓ | 89.15 | 74.63 | 85.25 | 50.24 | 5.93 | |||
| YOLOv11n | ✓ | 82.47 | 78.82 | 86.58 | 54.61 | 2.64 | |||
| YOLOv11n | ✓ | 87.27 | 76.51 | 85.13 | 52.73 | 2.67 | |||
| YOLOv11n | ✓ | ✓ | 86.90 | 83.01 | 87.84 | 55.14 | 5.99 | ||
| YOLOv11n | ✓ | ✓ | 89.14 | 78.66 | 87.68 | 55.37 | 5.99 | ||
| YOLOv11n | ✓ | ✓ | 91.25 | 82.55 | 88.93 | 55.44 | 2.72 | ||
| YOLOv11n | ✓ | ✓ | ✓ | 92.39 | 86.91 | 89.63 | 56.61 | 6.04 |
2.4.2 Neck Network Ablation
We analyze the contributions of the aggregation and injection modules within the RSD-Neck. The baseline uses YOLOv11n with Avgpool for aggregation and Conv+Add for injection. The following table shows the effect of replacing these with SPD-FAM and 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 |
2.4.3 Downsampling Operator Ablation
We compare different downsampling operators within RSD-Neck, using a baseline without SPD-Conv (denoted as N). The results demonstrate the superiority of SPD-Conv.
| Method | Precision (%) | Recall (%) | mAP50 (%) | mAP50:95 (%) |
|---|---|---|---|---|
| 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 |
| +SPDConv | 48.27 | 36.92 | 37.64 | 22.65 |
2.5 Robustness Evaluation on Extreme Tiny Object Scenarios
To further evaluate the robustness of RSD-Net in extreme tiny object scenarios, we extend experiments to the TinyPerson dataset, which features high-resolution images with average target size of only 18 pixels and severe background interference from wave textures. We compare RSD-Net with current mainstream detectors, including YOLOv8n, YOLOv11n, and the latest YOLOv12n.
| 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 |
| This work | 32.56 | 22.95 | 18.40 | 6.40 |
On the TinyPerson dataset, all models achieve low absolute mAP scores (mAP50 < 20%) due to extreme challenges: (1) annotation ambiguity for sub-18-pixel targets, (2) confusion between non-rigid targets and high-frequency wave textures, and (3) physical feature annihilation for sub-10-pixel objects. Despite these physical limits, RSD-Net demonstrates significant relative performance gains. While mainstream detectors (e.g., YOLOv12n) stagnate at a recall bottleneck of approximately 16%, RSD-Net achieves a breakthrough recall rate of 22.95% (a 6.87 percentage point improvement over YOLOv12n), and mAP50 increases by 5.33 percentage points. This confirms the superior cross-domain adaptability of RSD-Net in managing pixel-level tiny objects in diverse and complex China UAV aerial environments.

3. Conclusion
In this paper, we propose a refined spatial-aware distribution network (RSD-Net) to address the challenges of feature annihilation and background interference in China UAV tiny object detection. The SA-C3k2 module achieves frequency-domain adaptive feature extraction, effectively balancing shallow texture preservation with deep noise suppression. The RSD-Neck introduces SPD-Conv lossless downsampling and global context modeling, breaking through the information loss bottleneck during downsampling. The DP-Head constructs a “visual-statistical” dual-verification mechanism, improving the reliability of localization quality estimation. Experimental results show that RSD-Net improves mAP50 by 4.99 and 5.08 percentage points over the baseline on VisDrone2019-DET and NWPU VHR-10 datasets, respectively. On the TinyPerson extreme scenario test, the model successfully breaks through the recall bottleneck of general detectors, validating the robustness of the architecture. Future work will focus on further optimizing inference latency and exploring hardware acceleration based on FPGA or edge chips to meet the real-time engineering requirements of China UAV inspection tasks.
