In the domain of maritime search and rescue (SAR), the integration of drone technology has revolutionized emergency response by offering rapid deployment, wide-area coverage, and cost-effectiveness. However, the detection of human survivors in vast, dynamic ocean environments remains a formidable challenge due to the prevalence of small targets, complex sea conditions, and limited computational resources on board. Traditional single-stage object detectors, such as YOLO series, struggle with small object localization and adaptivity to varying illumination, wave interference, and fog. To address these limitations, this paper presents a novel model termed DDW-YOLO (Dynamic-Convolution, Dynamic-Upsampling, and Wasserstein-loss YOLO) built upon YOLO11, tailored specifically for unmanned aerial vehicle (UAV) maritime SAR missions. The model incorporates three key innovations: (1) a dynamic convolution module that adapts kernel weights based on input features, (2) a dynamic upsampling module that preserves high-frequency details via learnable offsets, and (3) a Wasserstein Wise Box (WWBox) loss function that mitigates gradient vanishing for non-overlapping boxes and suppresses noisy low-quality samples. Extensive ablation studies on the SeaDronesSee v2 dataset demonstrate that DDW-YOLO achieves a precision of 94.7% and recall of 92.7%, outperforming the baseline YOLO11 by 2.4% and 5.3%, respectively, while maintaining real-time inference at 232 FPS with only 4.67M parameters. This work provides an efficient and reliable visual detection solution for drone technology-based maritime search and rescue operations.
1. Introduction
Maritime accidents demand immediate lifesaving actions. Traditional rescue methods—helicopters and ships—are constrained by slow response times, high operational costs, and limited coverage. Drone technology, with its agility, low cost, and wide field of view, has emerged as a promising alternative for aerial maritime search. However, deploying object detection models on UAVs imposes stringent requirements: lightweight architecture, low latency, and robustness against environmental nuisances such as sea glint, waves, and small target sizes. Survivors in the water often appear as tiny blobs (e.g., a floating head or a small life jacket), making detection extremely difficult.
Current mainstream object detectors can be categorized into single-stage models (e.g., YOLO, SSD) and two-stage models (e.g., Faster R-CNN, Sparse R-CNN). While two-stage models offer higher accuracy, they are computationally heavy and unsuitable for onboard UAV processing. Single-stage models like YOLO are faster and lighter, yet they inherently lack sensitivity to small targets. Numerous improvements have been attempted—attention mechanisms, context modules, feature pyramid enhancements, and loss function refinements—but none fully satisfy the unique challenges of maritime SAR. This paper proposes a comprehensive solution built upon YOLO11, introducing dynamic convolution, dynamic upsampling, and a novel bounding-box loss function. The proposed DDW-YOLO significantly boosts detection performance while preserving real-time efficiency, making it an ideal candidate for drone technology-based emergency response.
2. Model Architecture and Improvements
2.1 Overview of DDW-YOLO
DDW-YOLO is constructed by augmenting YOLO11 with four major enhancements: (1) a small-target detection head, (2) dynamic convolution modules replacing standard convolutions, (3) dynamic upsampling modules (DySample) replacing fixed interpolation, and (4) a WWBox loss function combining Wasserstein distance and Wise IoU. The overall network architecture is illustrated conceptually in the earlier figure. The backbone and neck are revised to incorporate these modules, with the addition of a 160×160 detection layer dedicated to small objects. All modifications are designed to maintain low computational cost while boosting adaptability to maritime environments.
2.2 Small-Target Detection Head
In maritime SAR, a survivor’s visible part often is just the head—a minuscule object of roughly 10×10 pixels in a 4K image. The original YOLO11 uses three detection heads for large, medium, and small objects at strides 32, 16, and 8, respectively. To capture even smaller targets, we add a fourth head at stride 4. This is achieved by upsampling the 80×80 feature map from the neck by a factor of 2 (using the proposed dynamic upsampler) to obtain a 160×160 feature map, which is then concatenated with the low-level 160×160 feature map from the backbone (stage 2). The concatenated tensor passes through a C3k2 module to refine the features. Finally, a detection head (Detect) outputs bounding boxes and class probabilities at this high-resolution scale. The added head quadruples the detection resolution, enabling the model to recognize tiny objects that were previously missed.
2.3 Dynamic Convolution Module
Standard convolution applies the same kernel to all input samples, which is suboptimal for variable sea conditions. Dynamic convolution (DyConv) adapts its kernel weights per input using a routing network. As shown in Table 1, the routing network computes expert weights via global average pooling and a fully connected layer followed by Sigmoid activation. Unlike Softmax (which enforces competition), Sigmoid allows multiple experts to be activated simultaneously, enabling the model to blend receptive fields and capture diverse patterns. The forward pass performs grouped convolution with batch-specific kernels, maintaining low FLOPs despite increased parameter count. This module is inserted at six locations in the backbone and neck, replacing standard 3×3 convolutions.
Table 1: Dynamic convolution operation steps
| Step | Description | Equation |
|---|---|---|
| 1. Global Average Pooling | Compress spatial dimensions to a global descriptor | $$ \mathbf{p} = \frac{1}{H W} \sum_{i=1}^{H} \sum_{j=1}^{W} \mathbf{x}_{:,:,i,j} $$ |
| 2. Linear Projection | Map pooled vector to expert weights | $$ \mathbf{\alpha} = \sigma(\mathbf{p} \mathbf{W}_r) $$ |
| 3. Weight Mixing | Combine static experts weighted by α | $$ \mathbf{W}_{\text{mix}} = \mathbf{\alpha} \cdot \mathbf{W} $$ |
| 4. Grouped Convolution | Apply divided kernels per sample batch | $$ \mathbf{y} = \text{GroupConv}(\mathbf{x}_{\text{reshaped}}, \mathbf{W}_{\text{mix}}) $$ |
The router uses Sigmoid activation, ensuring that each expert can contribute independently. This flexibility allows the model to adjust its receptive field and filter characteristics based on the input, e.g., suppressing wave noise in clear water or enhancing contrast in foggy conditions. Compared with standard convolution, DyConv increases parameter count but reduces redundant computation because expert weights are computed on the fly.
2.4 Dynamic Upsampling Module (DySample)
Traditional nearest-neighbor or bilinear interpolation upsampling loses high-frequency details—edges, textures—that are critical for distinguishing survivors from debris or waves. DySample overcomes this by learning sample positions via point-based offsets. As illustrated in Table 2, the module first uses two parallel 1×1 convolutions: one predicts the base offsets, and the other predicts a scaling factor (through Sigmoid). The final offset vector is computed as:
$$ \Delta = f_{\text{offset}}(\mathbf{x}) \cdot \sigma(f_{\text{scope}}(\mathbf{x})) \cdot 0.5 + \mathbf{p}_{\text{init}} $$
where $$ \mathbf{p}_{\text{init}} $$ is the initial grid position. The offset is then applied to generate a deformed coordinate grid, and bilinear sampling is performed using the original feature map. This content-aware process retains fine details that interpolation methods discard, making DySample particularly effective for small object detection in dynamic maritime scenery.
Table 2: Dynamic upsampling module operations
| Stage | Operation | Purpose |
|---|---|---|
| 1. Offset Prediction | 1×1 conv → (2×G×S²) channels | Predict initial sample offsets |
| 2. Scope Prediction | 1×1 conv → Sigmoid → scale | Compute adaptive offset magnitude |
| 3. Offset Computation | Multiply offsets by scaled scope + init | Final per‑sample offset |
| 4. Grid Generation | coords = 2·C + Δ − 1 | Create deformed sampling grid |
| 5. Bilinear Sampling | grid_sample(input, coords) | Generate upsampled feature map |
DySample is used in two places: (1) in the small-target detection head to upsample the 80×80 map to 160×160, and (2) in the neck as part of the feature pyramid to fuse high-level semantics with low-level details. Its computational overhead is minimal—only two extra 1×1 convolutions—making it suitable for real-time drone technology applications.
2.5 WWBox Loss Function
Standard IoU loss suffers from two major problems in maritime SAR: (1) gradient vanishing when predicted and ground-truth boxes do not overlap (IoU = 0), which happens frequently for small floating objects; (2) sensitivity to low-quality samples (noisy annotations, partial occlusion) that dominate the gradient. To address these, we propose WWBox loss, a weighted combination of Wasserstein Distance loss and Wise IoU (WIoU) loss.
Wasserstein Distance Loss: Each bounding box is modeled as a 2D Gaussian distribution with mean μ (center) and covariance Σ (size). The Wasserstein–2 distance between two Gaussians is:
$$ W_2^2 = \| \mu_1 – \mu_2 \|_2^2 + \text{Tr}\!\left( \Sigma_1 + \Sigma_2 – 2(\Sigma_1^{1/2} \Sigma_2 \Sigma_1^{1/2})^{1/2} \right) $$
For axis-aligned boxes B₁ = (x₁,y₁,x₂,y₂), the Gaussian has μ = ((x₁+x₂)/2, (y₁+y₂)/2) and Σ = diag((x₂−x₁)/4, (y₂−y₁)/4). This distance is differentiable and provides a non-zero gradient even when boxes are far apart, thus solving the non-overlap problem.
Wise IoU Loss: WIoU dynamically scales the IoU contribution based on the quality of the sample. The core mechanism involves a non-monotonic focusing factor:
$$ \text{scaled\_loss} = \alpha / \beta, \quad \beta = \text{IoU} / \overline{\text{IoU}}, \quad \alpha = \delta \cdot \gamma^{\beta – \delta} $$
where $$ \overline{\text{IoU}} $$ is the exponentially moving average of IoU (updated each batch). Parameters δ and γ control the shape; we set δ=2, γ=0.5 following typical recommendations. This scaling reduces the influence of low-quality samples (with IoU far from the mean) while focusing on moderate-quality ones.
Combined Loss: WWBox = α·W₂ + (1-α)·WIoU, with α=0.5 (determined via grid search). The loss is applied to all detection heads. Table 3 summarizes the comparison.
Table 3: Comparison of bounding-box regression losses
| Loss | Gradient when IoU=0 | Sensitivity to low-quality | Small-target adaptability |
|---|---|---|---|
| IoU & variants (GIoU, DIoU) | Zero or near‑zero | High (equal weighting) | Poor when incomplete overlap |
| Wasserstein | Non‑zero (distance‑based) | Moderate (no quality weighting) | Good for non‑overlap |
| WIoU | Zero for non‑overlap | Low (adaptive scaling) | Good for quality balancing |
| WWBox (Ours) | Non‑zero | Low | Excellent |
3. Experiments and Analysis
3.1 Dataset and Setup
We used the SeaDronesSee v2 dataset, a large-scale UAV maritime dataset containing over 54,000 4K images with six object categories (swimmer, boat, jet ski, life-saving equipment, buoy, etc.) and more than 400,000 instances. To fit our GPU memory (Tesla V100 32GB), we randomly sampled 5,000 images and cropped each to 640×640 using area-based clustering. The training parameters are fixed across all experiments: learning rate = 0.01, SGD optimizer with momentum 0.937, batch size = 64, 400 epochs, and fixed random seed 1. Evaluation metrics include Precision, Recall, mAP50, mAP50-95, FPS, parameters (M), and GFLOPs.
3.2 Ablation Study
We conducted systematic ablation to assess each component’s contribution. The baseline is YOLO11n. Results are shown in Table 4.
Table 4: Ablation study results on SeaDronesSee v2
| Small-Head | DyConv | DySample | WWBox | 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 |
Observations: Adding the small-target detection head (row 2) boosts recall by 3.7% and mAP50 by 2.8%, albeit with a decrease in FPS from 285 to 204. Integrating DyConv (row 3) recovers speed (FPS 250) while further improving recall (91.7%) and mAP50-95 (64.5%). This synergy arises because DyConv adapts to the richer feature space introduced by the new head. DySample (row 4) lifts recall to 92.5% and mAP50-95 to 65.1%, demonstrating the value of detail preservation. Finally, incorporating WWBox loss (row 5) brings all metrics to their peak: precision 94.7% (+2.4% vs baseline), recall 92.7% (+5.3%), mAP50 95.3% (+3.5%), and mAP50-95 66.2% (+5.1%). The total parameter growth is modest (2.58M to 4.67M), and the final model operates at 232 FPS—well above real-time threshold.
3.3 Comparison with State-of-the-Art
We compared DDW-YOLO against YOLO11n, YOLOv10n, RTDETR-l, and RTDETR-ResNet50. All models were trained and evaluated under the same settings. Table 5 presents the results.
Table 5: Comparison of object detectors on SeaDronesSee v2
| Model | Precision (%) | Recall (%) | mAP50 (%) | mAP50-95 (%) | FPS | Params (M) | GFLOPs |
|---|---|---|---|---|---|---|---|
| YOLO11n (baseline) | 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 |
| RTDETR-l | 73.2 | 73.0 | 75.8 | 40.0 | 147 | 32.0 | 103.5 |
| RTDETR-ResNet50 | 79.9 | 82.9 | 83.2 | 47.9 | 127 | 41.9 | 125.6 |
| DDW-YOLO (Ours) | 94.7 | 92.7 | 95.3 | 66.2 | 232 | 4.67 | 8.4 |
YOLO-based models significantly outperform RTDETR series on this low-latency maritime task. The Transformer-based detectors, despite their larger capacity, fail to capture small objects effectively and have higher computational cost (103–125 GFLOPs vs 8.4 GFLOPs). Among YOLO variants, YOLO11n shows higher precision than YOLOv10n (92.3% vs 91.0%) but lower recall (87.4% vs 89.1%). DDW-YOLO surpasses both in every metric, achieving a 2.4% improvement in precision and 5.3% in recall over YOLO11n. The FPS of 232 is sufficient for real-time UAV deployment.
3.4 Qualitative Analysis
To visually inspect the improvements, we compared predictions of YOLO11n and DDW-YOLO on challenging test images. In typical marine scenes with strong reflections, YOLO11n often misses swimmers whose arms blend with wave crests. DDW-YOLO, with its dynamic upsampling preserving edge details, correctly locates the person. In another example, a survivor is partially submerged showing only an orange vest; the dynamic convolution in DDW-YOLO adaptively contextualizes the object, inferring the presence of a human. When a life jacket is fragmented or overlapping with debris, the WWBox loss enables better regression by not being misled by low-quality anchor boxes. These qualitative results confirm that the proposed improvements are robust against the typical failure modes of standard detectors.
4. Discussion
The success of DDW-YOLO lies in the synergistic combination of four targeted enhancements. The small-target detection head increases the spatial resolution of feature maps, directly addressing the scale challenge. Dynamic convolution and upsampling allow the model to adapt its processing to the specific visual characteristics of each frame—an essential property for the non-stationary maritime environment. Finally, the novel loss function effectively handles the twin problems of non-overlapping boxes and noisy annotations, which are particularly prevalent in small-target scenarios. The overall architecture maintains YOLO’s philosophy of speed and simplicity, with only a 1.8× increase in parameters but a 2.4%–5.3% gain in detection accuracy.
One limitation is that the new small-target head requires additional memory and computation, reducing FPS from 285 to 232. However, 232 FPS is still far above the typical UAV camera frame rate (30 fps), making the model suitable for real-time operation. Future work could explore pruning or quantization to further reduce the model size while preserving accuracy. Moreover, the dataset used (SeaDronesSee v2) is fixed-wing UAV captured; rotary-wing UAVs may have different motion blur characteristics. Adapting DDW-YOLO to such domain shifts via domain adaptation or continuous learning would be valuable.
5. Conclusion
In this paper, we have presented DDW-YOLO, a purpose-built object detection model for UAV-based maritime search and rescue. By augmenting YOLO11 with a small-target detection head, dynamic convolution, dynamic upsampling, and a Wasserstein-WIoU loss function, we have achieved significant improvements in both precision and recall while retaining real-time inference speed. The model demonstrates excellent adaptability to complex marine environments, effectively detecting tiny survivors that are often missed by conventional detectors. With only 4.67 million parameters and 8.4 GFLOPs, DDW-YOLO represents a lightweight yet powerful solution for drone technology-enabled emergency response. The proposed techniques can be extended to other challenging small-object detection tasks such as surveillance, wildlife monitoring, and disaster management. We believe this work will contribute to the advancement of drone technology in lifesaving applications, ultimately increasing the chances of successful maritime rescues.