Multi-ship detection and classification with feature enhancement and lightweight fusion

Area of Science:

• Computer Vision
• Marine Technology
• Artificial Intelligence

Background:

• Classical ship detection methods face challenges like background interference and multi-scale object recognition.
• Inadequate training data for small sample recognition hinders traditional techniques.
• Ensuring maritime safety and efficient ship traffic management necessitates robust detection systems.

Purpose of the Study:

• To develop an enhanced ship multi-target detection model for improved accuracy and reliability.
• To address limitations of classical methods in complex marine environments.
• To enhance the extraction of semantic features for ship identification against intricate backdrops.

Main Methods:

• Modified YOLOv8 algorithm as the baseline model (YOLOv8n).
• Integration of ESSE module and GSConvns technology into the YOLOv8 backbone.
• Incorporation of Wise-IoU technology for improved detection.

Main Results:

• Achieved average detection accuracies of 82.1% (Dockship), 99.1% (Seaships), and 91.7% (Infrared Offshore Ship dataset).
• Demonstrated significant improvement over the baseline model.
• Validated enhanced feature extraction, lightweight fusion, and detection capabilities through IoU and ablation studies.

Conclusions:

• The proposed enhanced YOLOv8 model significantly improves ship target detection performance.
• The modifications enhance multi-scale feature extraction and semantic characteristic extraction, especially in challenging conditions.
• The approach offers potential for increased automation, dependability, and quality in ship target detection systems.