AI Boosts Ship‑Sea Target Detection: Lessons from the First Innovation Competition

Competition Overview

The first "Ship‑Sea Data Intelligent Application Innovation Competition" was jointly organized by the Taihu Laboratory (a deep‑sea technology research institute), Huawei, and the Wuxi municipal talent and science bureaus to promote the integration of modern information technology, next‑generation AI, and ship technology.

Challenge and Motivation

With growing demands for intelligent waterborne traffic, marine environment monitoring, and underwater exploration, improving detection and recognition of typical targets—such as ships, fishing nets, buoys, floating debris, reefs, and marine organisms—has become a critical AI problem in the maritime field.

Top Teams and Technical Solutions

In the first track "Surface/Underwater Typical Target Recognition", the top three teams were:

Zhang Junjie team, Huazhong University of Science and Technology

Shen Fei team, Nanjing University of Science and Technology

Xu Ke team, Zhejiang University

Common challenges included complex backgrounds, lighting and angle variations, data imbalance, and annotation errors, as well as a strict 30 FPS speed requirement. Teams addressed these issues with:

Extensive data augmentation, median filtering, mixup, and graph convolution to enhance feature extraction.

Multi‑scale training and test‑time augmentation (TTA) to boost mAP.

Multi‑label handling for ambiguous ship images.

Merge‑NMS to weight low‑confidence boxes instead of discarding them.

Stochastic Weight Averaging (SWA) for better generalization.

Knowledge distillation, model pruning, quantization, and TensorRT for speed‑accuracy balance.

Impact, Rewards, and Future Outlook

The competition featured real‑time, transparent ranking and offered substantial research funding—up to one million RMB for projects meeting national natural science fund standards—providing the biggest incentive for participants.

Beyond the prize money, the event aimed to bridge traditional computer‑vision methods with deep‑learning approaches, foster an innovative ecosystem, and stimulate downstream industry development. Organizers emphasized the goal of creating a digital‑maritime innovation environment, encouraging more talent to engage, and accelerating the transformation of research results into practical ship‑sea applications.

Note: This article was also published in Xinhua Daily.