TL;DR
SUD-YOLO (Stable Underwater Detection) is a specialized enhancement of the YOLOv11 architecture designed to tackle the unique challenges of underwater imagery, such as blur, low contrast, and object deformation. By introducing novel feature sampling modules (SRFD and DRFD) and a decoupled EfficientHead, the researchers achieved significantly higher detection accuracy and training stability across multiple marine datasets compared to standard YOLO models.
What Problem Does This Paper Address?
Underwater target detection (UTD) is notoriously difficult for standard computer vision models. The "medium" itself—water—introduces light scattering, absorption, and turbidity, resulting in images that are often blurry, low-contrast, and color-distorted. Furthermore, camera jitter and the fluid movement of marine life cause significant object deformation.
While the latest YOLOv11 model is a powerhouse for real-time detection in terrestrial environments, it struggles underwater. The authors identify a "feature loss" problem: standard downsampling and upsampling operations in YOLOv11 discard critical texture information that is already sparse in murky underwater shots. This leads to low detection accuracy and unstable training.
Methods
The authors propose SUD-YOLO, which introduces three primary architectural changes to the YOLOv11 baseline:
- SRFD (Shallow Robust Feature Downing-sampling): Replaces the initial convolution layers in the backbone. It uses a "Cut-Slice" method (similar to space-to-depth operations) and Depthwise Convolutions (DwConv) to downsample features without losing the fine-grained texture details essential for identifying camouflaged or blurry marine organisms.
- DRFD (Deep Robust Feature Downing-sampling): Replaces the standard upsampling modules in the "neck" of the network. It incorporates Grouped Convolutions (GConv) and GeLU activation functions to prevent gradient disappearance in deeper layers, ensuring that high-level semantic information is preserved and fused effectively.
- EfficientHead: Traditional YOLO heads often use a "mixed" approach where classification (what is it?) and regression (where is it?) share the same features. SUD-YOLO adopts a decoupled head (inspired by YOLOv6) that separates these tasks. This prevents the two objectives from interfering with each other, which is particularly helpful when object boundaries are poorly defined due to underwater haze.
Datasets
The model was rigorously evaluated on three distinct underwater benchmarks:
- URPC2020: A dataset from the Underwater Robot Picking Contest featuring four categories: echinus (sea urchin), holothurian (sea cucumber), scallop, and starfish.
- Luderick: Focused on high-quality annotations of fish in river basin environments with dense seagrass backgrounds.
- Deepfish: A large-scale benchmark containing over 40,000 high-definition images from various marine habitats, used here specifically for fish detection.
Key Insights and Ideas
- Sampling is the Bottleneck: In underwater imaging, the "information density" is lower than in clear-air photography. The authors realized that standard pooling or strided convolutions act as a destructive filter. Their SRFD/DRFD modules treat sampling as a feature-preservation task rather than just a dimensionality reduction task.
- Stability via Architecture: One of the most striking findings wasn't just the final accuracy, but the stability of the training process. The inclusion of Batch Normalization within the sampling modules led to much smoother convergence curves, suggesting the model is less sensitive to the "noise" inherent in underwater data.
- Decoupling for Fuzzy Boundaries: By using the EfficientHead, the model can optimize for bounding box precision without compromising its ability to categorize species, a common trade-off in single-stage detectors working with low-quality input.
Key Results
SUD-YOLO consistently outperformed YOLOv3, v5, and the baseline v11 across all metrics:
- Accuracy Boost: On the Luderick dataset, SUD-YOLO achieved an [email protected]:0.95 of 0.627, a significant jump from the baseline YOLOv11's 0.586.
- Reliability: On the URPC2020 dataset, the model showed a 3% improvement in Recall over YOLOv11, meaning it missed fewer targets in complex, cluttered environments.
- Convergence: Training logs showed that SUD-YOLO reached near-optimal performance within the first 20 epochs, whereas baseline models exhibited much higher fluctuations in precision throughout the 100-epoch cycle.
Limitations and Future Work
While SUD-YOLO shows excellent generalization across the three tested datasets, the authors note that extreme conditions—such as heavy fog-like turbidity or near-zero light—still pose a challenge.
Future research directions include:
- Testing the architecture on other "noisy" domains, such as aerial photography in inclement weather (fog/rain).
- Further optimizing the computational efficiency of the SRFD/DRFD modules to ensure the model remains viable for deployment on low-power underwater autonomous vehicles (AUVs).
Summary updated