When maritime accidents occur, the rapid deployment of China drone systems plays a pivotal role in locating survivors amidst vast and treacherous sea conditions. Compared to traditional rescue vessels and aircraft, China drone technology offers distinct advantages: rapid response, extensive coverage, and cost efficiency. However, the complex sea surface — characterized by intense sunlight reflections, wave interference, fog, and the extreme small size of human heads or life vests — poses severe challenges to visual detection models. Existing single-stage object detectors, including the widely adopted YOLO family, often suffer from insufficient detection of small targets and poor adaptability to dynamic environments. In this work, I address these pressing issues by developing a novel model named DDW-YOLO (Dynamic convolution, Dynamic upsampling, and Wasserstein-loss YOLO). This model builds upon the YOLO11 baseline and integrates three key innovations: a dedicated small-target detection layer, dynamic convolution and dynamic upsampling modules for environmental adaptation, and a novel Wasserstein-wise box loss (WWbox) that combines Wasserstein distance with wise IoU. Extensive experiments on the SeaDronesSee v2 dataset demonstrate that DDW-YOLO achieves a precision of 94.7% and a recall of 92.7%, significantly outperforming its predecessors, making it an ideal solution for China drone maritime search and rescue missions.
1. Introduction
The deployment of China drone in maritime search and rescue operations has become increasingly critical. However, the inherent limitations of lightweight detection models often lead to missed detections of drowning individuals, whose heads or bright safety vests may occupy only a few dozen pixels in a 4K image. The YOLO11 model, while efficient, struggles with such small targets and cannot adapt to rapidly changing illumination and wave patterns. To overcome these shortcomings, I propose DDW-YOLO, which introduces three major improvements:
- A small-target detection layer that fuses high-resolution shallow features with deep semantic features, enabling reliable detection of objects as small as 10×10 pixels.
- Dynamic convolution that adaptively adjusts convolution kernels based on input features, allowing the model to handle diverse sea states (calm, choppy, foggy, etc.) without retraining.
- Dynamic upsampling which replaces fixed-interpolation upsampling with a learning-based point sampling approach, preserving high-frequency details crucial for distinguishing a swimmer from a wave crest.
- A WWbox loss that combines Wasserstein distance and wise IoU, solving the gradient vanishing problem when bounding boxes do not overlap, and adaptively weighting difficult samples.
The rest of the paper is organized as follows. Section 2 details the proposed architecture. Section 3 presents the experimental setup, ablation studies, and comparison results. Section 4 draws conclusions.
2. Proposed Method
2.1 Overall Architecture
DDW-YOLO inherits the backbone and neck structure of YOLO11 but replaces standard convolutions with dynamic convolutions in key stages, substitutes nearest-neighbor upsampling with dynamic upsampling, and adds an additional detection head for small targets. The network outputs predictions at four scales: 160×160, 80×80, 40×40, and 20×20.
The following table summarizes the modules integrated into the baseline YOLO11:
| Component | Function | Impact on Performance |
|---|---|---|
| Small-target detection layer | Upsamples 80×80 feature maps to 160×160 and concatenates with shallow features | +2.1% recall, +2.8% mAP50 |
| Dynamic convolution | Routes adaptive kernel weights via a lightweight network | +0.5% precision, +0.8% mAP50-95 |
| Dynamic upsampling | Learns sample offsets to recover high-frequency details | +1.0% recall, +0.6% mAP50-95 |
| WWbox loss | Combines Wasserstein distance and wise IoU | +1.4% precision, +1.2% mAP50 |
2.2 Small-Target Detection Layer
In maritime rescue, a drowning person’s head often appears as a tiny dot in the image. The original YOLO11 only has detection heads for feature maps of sizes 80×80, 40×40, and 20×20. To capture minute features, I add a new detection head on a 160×160 feature map. This is achieved by taking the 80×80 feature map from the neck, applying 2× dynamic upsampling, and concatenating it with the corresponding shallow feature map from the backbone (also 160×160). The concatenated features are then refined by a C3k2 module before being fed into the detection head. This design not only preserves fine spatial details but also integrates high-level semantic context, enabling accurate identification of small targets even under heavy sea clutter.
2.3 Dynamic Convolution
Standard convolutions apply fixed kernels to all inputs, which is suboptimal for non-stationary scenes like the ocean. I adopt a dynamic convolution module that consists of a routing network and a conditional convolution branch. For an input feature map \(x \in \mathbb{R}^{B \times C \times H \times W}\), the routing network first computes global descriptors via average pooling:
$$ \text{pooled\_inputs} = \frac{1}{H \times W} \sum_{i=1}^{H} \sum_{j=1}^{W} x_{:,:,i,j} \in \mathbb{R}^{B \times C} $$
Then, a lightweight fully connected layer with Sigmoid activation generates \(E\) expert weights:
$$ \text{routing\_weights} = \sigma(\text{pooled\_inputs} \cdot W_r) \in \mathbb{R}^{B \times E} $$
The conditional convolution branch maintains a set of \(E\) static convolutional kernels. The final kernel for each sample is a linear combination weighted by the routing weights:
$$ W_{\text{combined}} = \alpha \cdot W, \quad \alpha \in \mathbb{R}^{B \times E}, \quad \sum_{j=1}^{E} \alpha_{i,j} = 1 $$
The output is then computed using grouped convolution where the group size equals the batch size. This approach adds minimal computational cost (<1% increase in GFLOPs) while significantly improving adaptability to varying sea conditions.
2.4 Dynamic Upsampling
Conventional upsampling (e.g., bilinear or nearest) cannot adapt to local content and often blurs fine edges. I replace them with a learned upsampling method based on point sampling and dynamic offsets. Given an input feature map, two parallel 1×1 convolutions predict:
- Main offset \(f_{\text{offset}}(x)\) with output channels = \(2 \times G \times S^2\) (where \(G\) is the number of groups and \(S\) is the upsampling factor)
- Dynamic scope coefficient \(f_{\text{scope}}(x)\) passed through Sigmoid to limit the offset range
The final offset is computed as:
$$ \Delta = f_{\text{offset}}(x) \cdot \sigma(f_{\text{scope}}(x)) \cdot 0.5 + P_{\text{init}} $$
where \(P_{\text{init}}\) is the initial grid shift. The offset-adjusted coordinates are then used for bilinear sampling. This mechanism allows the model to learn when to stretch, compress, or shift sampling points, preserving high-frequency details such as the outline of a swimmer’s head against the sky horizon.
2.5 WWbox Loss
In maritime imagery, it is common for predicted boxes to have zero Intersection over Union (IoU) with ground truth, causing gradient saturation in standard IoU-based losses. To address this, I introduce a loss function that models bounding boxes as 2D Gaussian distributions and computes the Wasserstein distance. For two boxes \(B_1 = (x_1, y_1, x_2, y_2)\) and \(B_2 = (x’_1, y’_1, x’_2, y’_2)\), the Gaussian parameters are:
$$ \mu_1 = \left( \frac{x_1 + x_2}{2}, \frac{y_1 + y_2}{2} \right), \quad \Sigma_1 = \begin{bmatrix} \frac{x_2 – x_1}{4} & 0 \\ 0 & \frac{y_2 – y_1}{4} \end{bmatrix} $$
The squared Wasserstein distance is:
$$ W_2^2(N_1, N_2) = \|\mu_1 – \mu_2\|_2^2 + \text{Tr}\left( \Sigma_1 + \Sigma_2 – 2(\Sigma_1^{1/2} \Sigma_2 \Sigma_1^{1/2})^{1/2} \right) $$
This distance is non-zero even when bounding boxes do not overlap, providing continuous gradients. Furthermore, I combine it with wise IoU (WIoU) which adaptively scales the loss contribution based on the outlier factor \(\beta\):
$$ \beta = \frac{\text{IoU}}{\text{iou\_mean}}, \quad \text{scaled\_loss} = \frac{\alpha}{\beta} $$
where \(\alpha = \delta \cdot \gamma^{\beta – \delta}\) controls the monotonicity. The final WWbox loss is:
$$ \mathcal{L}_{\text{WWbox}} = 0.5 \cdot W_2^2 + 0.5 \cdot \text{WIoU} $$
This hybrid loss ensures robust optimization under severe occlusion, reflection, and wave interference typical in China drone maritime rescue scenarios.
3. Experiments and Analysis
3.1 Dataset and Settings
I evaluate DDW-YOLO on the SeaDronesSee v2 dataset, which contains 54,000+ aerial images taken by UAVs over the sea. The dataset includes various objects: swimmers, boats, buoys, life-saving equipment, etc. To balance computational cost, I randomly sampled 5,000 images and cropped them to 640×640. All experiments are conducted on a Tesla V100-SXM2-32GB GPU using PyTorch 2.4. The training hyperparameters are fixed as shown in Table 1.
| Parameter | Value | Parameter | Value |
|---|---|---|---|
| Learning rate | 0.01 | Momentum | 0.937 |
| Epochs | 400 | Optimizer | SGD |
| Batch size | 64 | Random seed | 1 |
Metrics include Precision (P), Recall (R), mAP50, mAP50-95, FPS, and model parameters. The formulas are:
$$ \text{Precision} = \frac{TP}{TP + FP} $$
$$ \text{Recall} = \frac{TP}{TP + FN} $$
$$ \text{mAP50} = \int_0^1 P(r) \, dr $$
$$ \text{mAP50-95} = \frac{1}{10} \sum_{k=0}^{9} \text{mAP}_{\text{IoU}=0.5+0.05k} $$
3.2 Ablation Study
To verify the contribution of each component, ablation experiments are performed by incrementally adding modules to the YOLO11 baseline. Results are summarized in Table 2.
| Small Target Layer | Dynamic Conv | Dynamic Up | WWbox Loss | Precision (%) | Recall (%) | mAP50 (%) | mAP50-95 (%) | FPS | Params (M) |
|---|---|---|---|---|---|---|---|---|---|
| ✗ | ✗ | ✗ | ✗ | 92.3 | 87.4 | 91.8 | 61.1 | 285 | 2.58 |
| ✓ | ✗ | ✗ | ✗ | 94.4 | 91.1 | 94.6 | 64.3 | 204 | 2.63 |
| ✓ | ✓ | ✗ | ✗ | 92.3 | 91.7 | 93.9 | 64.5 | 250 | 4.64 |
| ✓ | ✓ | ✓ | ✗ | 93.3 | 92.5 | 94.1 | 65.1 | 238 | 4.67 |
| ✓ | ✓ | ✓ | ✓ | 94.7 | 92.7 | 95.3 | 66.2 | 232 | 4.67 |
Key observations: Adding only the small‑target detection layer boosts recall by 3.7% but reduces FPS due to the extra head. Dynamic convolution recovers speed (250 vs 204 FPS) while improving mAP50‑95. Dynamic upsampling further enhances recall and mAP50‑95. Finally, the WWbox loss pushes precision to 94.7% and recall to 92.7%, a 2.4% and 5.3% improvement over the baseline, respectively.
3.3 Comparison with State-of-the-Art
I compare DDW-YOLO against YOLO11n, YOLOv10n, and two RT-DETR variants. All models are trained under identical conditions. Table 3 reports the results.
| Model | Precision (%) | Recall (%) | mAP50 (%) | mAP50‑95 (%) | FPS | Params (M) | GFLOPs |
|---|---|---|---|---|---|---|---|
| DDW-YOLO (Ours) | 94.7 | 92.7 | 95.3 | 66.2 | 232 | 4.67 | 8.4 |
| YOLO11n | 92.3 | 87.4 | 91.8 | 61.1 | 285 | 2.58 | 6.3 |
| YOLOv10n | 91.0 | 89.1 | 92.0 | 62.6 | 350 | 2.71 | 8.4 |
| RT-DETR-l | 73.2 | 73.0 | 75.8 | 40.0 | 147 | 32.0 | 103.5 |
| RT-DETR-ResNet50 | 79.9 | 82.9 | 83.2 | 47.9 | 127 | 41.9 | 125.6 |
DDW-YOLO significantly outperforms all competitors in both accuracy and speed‑efficiency trade‑offs. Despite having 4.67M parameters (slightly more than YOLO11n’s 2.58M), it achieves 8.4 GFLOPs, comparable to YOLOv10n, while delivering the highest precision and recall. The RT-DETR models, although powerful, are too heavy for real‑time deployment on edge devices commonly used in China drone applications.
3.4 Qualitative Analysis
In practical maritime search scenarios, DDW-YOLO exhibits remarkable robustness. While the baseline YOLO11 often misses swimmers whose bodies are partially submerged (only the orange life jacket visible), DDW-YOLO successfully detects them. In images with strong sea surface glare, the dynamic upsampling preserves the boundary of the target, allowing accurate localization. The dynamic convolution adapts to the varying texture of waves and ship wakes, preventing false positives.
4. Conclusion
I have presented DDW-YOLO, a robust detection model specifically designed for China drone maritime search and rescue operations. By integrating a small‑target detection layer, dynamic convolution, dynamic upsampling, and a novel WWbox loss, the model achieves a precision of 94.7% and a recall of 92.7% on the challenging SeaDronesSee v2 dataset. The improvements in recall (5.3% over baseline) directly translate to more lives saved by reducing missed detections. The model retains a high frame rate of 232 FPS with only 8.4 GFLOPs, making it suitable for real‑time deployment on lightweight UAV platforms. Future work will focus on further compression for edge AI hardware and adaptation to night‑vision and infrared imaging. DDW-YOLO represents a significant step forward in integrating advanced deep learning with practical China drone emergency response systems.