Multi-Type Ship Detection in Complex Marine Backgrounds Using an Enhanced YOLO-Based Network

Area of Science:

• Computer Vision
• Maritime Surveillance
• Deep Learning

Background:

• Accurate ship detection is crucial for maritime security and rights protection.
• Increasing vessel diversity and complex environments pose challenges for current marine surveillance systems.
• Existing methods struggle with small targets, feature discrimination, and precise localization.

Purpose of the Study:

• To develop a robust multi-scale detection algorithm for complex marine scenarios.
• To address high false-negative rates for small ship targets.
• To improve feature discrimination and localization accuracy in cluttered backgrounds.

Main Methods:

• Proposed AK-DSAM-YOLOv13 algorithm built on YOLOv13n architecture.
• Incorporated a lightweight cross-scale feature extraction module (AKC3k2) with Alterable Kernel Convolutions (AKConv).
• Designed a Dynamic Up-Sampling Dual-Stream Attention Merging (DyDSAM) structure and an Accuracy-Intersection-over-Union (AIoU) loss function.

Main Results:

• AK-DSAM-YOLOv13 demonstrated superior performance on CM-Ships, SeaShips, and McShips datasets.
• Significant improvements in detection accuracy, recall, and generalization capability were observed.
• The algorithm maintained low computational overhead while enhancing multi-scale target representation and feature fusion.

Conclusions:

• AK-DSAM-YOLOv13 offers an efficient and reliable solution for intelligent maritime visual monitoring.
• The proposed optimizations effectively tackle challenges in complex marine environments.
• This research advances ship target detection for enhanced maritime security.