Abstract
Accurate vessel draft reading (VDR) is essential for intelligent maritime surveillance, primarily to monitor ship loading and prevent overloading. While computer vision offers a cost-effective solution for estimating draft depth, traditional methods often struggle with low-quality imagery and high computational demands. This paper presents MTL-VDR, a multi-task learning framework designed for robust and real-time VDR. The system integrates draft mark detection, scale recognition, vessel/water segmentation, and adaptive depth estimation. By employing a shared backbone for simultaneous recognition and segmentation, and incorporating a rule-based correction module, MTL-VDR achieves high accuracy even under challenging conditions like stained hulls. Experimental results demonstrate superior performance, with a computational speed exceeding 40 FPS, meeting the requirements for real-time maritime safety monitoring.
Introduction
Vessel draft reading (VDR) measures the vertical distance between the waterline and the bottom of the hull, serving as the primary indicator of a ship's displacement and total load. In maritime surveillance, automated VDR is critical for identifying overloaded vessels that pose significant safety risks. However, automating this task is difficult due to environmental noise, such as hull corrosion, stains, varying lighting, and the dynamic nature of the waterline. Traditional automated approaches often process detection and segmentation as isolated, computationally expensive steps, which limits their utility for real-time applications.
The MTL-VDR Framework
The proposed MTL-VDR system streamlines the draft reading process into four distinct stages: initial detection, multi-task analysis, rule-based correction, and final depth estimation.
- Draft Mark Detection: The system first identifies the region of interest containing the draft marks using a customized YOLOv8 architecture.
- Multi-Task Learning (MTL) Subnet: This core component uses a shared backbone to perform two tasks in parallel: recognizing individual draft scales (digits and units) and segmenting the image into vessel and water regions.
- Spatial Distribution Correction: To handle obscured or damaged scales, a correction module applies physical rules—such as the standard 0.2m interval between scales—to validate and correct recognition results.
- Adaptive Depth Estimation: Finally, the system calculates the draft depth by geometrically relating the recognized scales to the segmented waterline.
Architecture and Methodology
The heart of MTL-VDR is its multi-task architecture, which balances efficiency and accuracy. By sharing a backbone (DarkNet-53), the model avoids redundant feature extraction for recognition and segmentation.
To optimize this joint learning process, the authors utilize a dynamic loss weighting strategy based on homoscedastic uncertainty. This ensures that neither the detection task nor the segmentation task dominates the training. Furthermore, a specialized Mean Absolute Vertical Distance (MAVD) loss is introduced for the segmentation subnet. Unlike standard pixel-wise losses, MAVD focuses specifically on the accuracy of the water-vessel boundary, which is the most critical factor for precise depth estimation.
Robustness and Real-World Performance
One of the primary innovations of MTL-VDR is its ability to handle "dirty" data. In real-world maritime environments, draft scales are frequently stained by mud or rust. The system’s spatial correction module acts as a fail-safe; if the neural network misidentifies a digit, the module uses the vertical positions of surrounding, correctly identified scales to interpolate the true value. This physics-based prior prevents the "silly" mistakes common in pure deep learning models.
The performance was evaluated on a newly constructed VDR benchmark dataset containing over 2,200 images. MTL-VDR achieved a Mean Absolute Draft Depth Error (MADDE) of 0.074 meters, significantly outperforming traditional decoupled methods. Crucially, the system maintains an inference speed of approximately 60 FPS on mid-range hardware, making it highly suitable for deployment in live surveillance networks.
Conclusion
MTL-VDR represents a significant advancement in automated maritime surveillance. By integrating multi-task learning with physics-based priors, the system overcomes the limitations of traditional computer vision methods in challenging environments. Future work will focus on enhancing performance in low-visibility conditions like fog or night and adapting the system for the diverse unit systems and hull shapes found in international oceanic shipping.
Summary updated