Smart Ship Draft Reading by Dual-Flow Deep Learning Architecture and Multispectral Information.

Area of Science:

• Maritime Logistics
• Computer Vision
• Remote Sensing

Background:

• Ship draft surveys are crucial for accurate bulk cargo weighing in maritime trade.
• Current visual draft reading methods suffer from safety, cost, and accuracy limitations.
• Existing image processing techniques struggle with water surface interference like reflections.

Purpose of the Study:

• To develop an accurate and reliable automatic ship draft reading system.
• To overcome the limitations of RGB-based image processing in maritime environments.
• To introduce a novel approach integrating Near-Infrared (NIR) and RGB data for draft reading.

Main Methods:

• A dataset of 524 annotated multispectral images (RGB and NIR) was created.
• A dual-branch backbone network (BIF) was proposed to extract and fuse spectral information.
• The BIF backbone was integrated with YOLOv8 for draft detection and UPerNet for waterline segmentation.

Main Results:

• Draft detection achieved a mean Average Precision (mAP) of 99.2% using YOLOv8 with the BIF backbone.
• Waterline segmentation mIoU improved from 98.9% to 99.3% when using UPerNet with the BIF backbone.
• The method demonstrated a draft reading inaccuracy of less than ±0.01 m.

Conclusions:

• Integrating NIR and RGB data significantly improves automatic draft reading accuracy.
• The proposed BIF dual-branch backbone effectively extracts and combines multispectral information.
• This method offers a robust solution for accurate and automated ship draft surveys in maritime transportation.