A two-stage approach for ship detection in restricted visibility based on dehazing and SE-YOLO algorithms

ABSTRACT

The rapid and automatic detection of ships in restricted visibility is crucial for protecting the safety of maritime navigation. The frequent presence of fog in maritime environments causes restricted visibility, contributing to a higher occurrence of maritime accidents. Aiming to address ship target detection's low accuracy in restricted visibility weather, this paper proposes an improved dehazing model based on a two-stage ship detection algorithm. First, during the dehazing stage, a specialized model for maritime environments addresses the problem of unclear ship features in foggy images. The model is trained by a combination of synthetic images training and real images fine-tuning. Then, during the detection stage, a modified YOLO network for small ship detection called SE-YOLO is introduced to solve the issue of small ship targets' low detection accuracy. On the one hand, the modified Spatial Pyramid Pooling – Fast(SPPF) module is designed to reduce information loss during feature extraction; on the other hand, the attention mechanisms are integrated to enhance the network's sensitivity to the details of small ship targets and improve the model's overall detection performance for ships. Moreover, to simulate the real sea scene and test the effectiveness of our method, a maritime-haze dataset containing different concentrations of fog and various brightness is made for this research. Finally, The experimental results indicate that, compared to the traditional YOLOv5 method, our method performs better on detecting ships in restricted visibility environments, with the mean average precision (mAP) increased from 62.64% to 77.23%.

KEYWORDS:

• Deep learning
• image dehazing
• object detection
• neural networks
• marine vehicles

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The datasets generated during and analysed during the current study are availablefrom the corresponding author on reasonable request.

Additional information

Funding

This work was supported by the National Natural Science Foundation of China under Grants 51939001, 61976033, 61803064, and in part by the Science and Technology Innovation Funds of Dalian under Grant 2018J11CY022, and in part by the Liaoning Revitalization Talents Program under Grant XLYC1908018, XLYC1807046, and in part by the Natural Foundation Guidance Plan Project of Liaoning under Grant 2019-ZD-0151, and in part by the Fundamental Research Funds for the Central Universities under Grant 3132022143.