M²S-YOLOv8: Multi-Scale and Asymmetry-Aware Ship Detection for Marine Environments.

Area of Science:

• Computer Vision
• Marine Robotics
• Artificial Intelligence

Background:

• Ship detection is crucial for marine environmental perception but faces challenges like target occlusion, multi-scale targets, asymmetric geometries, and harsh imaging conditions.
• Existing vision-based methods struggle with degraded image quality and complex scenarios common in real marine environments.

Purpose of the Study:

• To develop an improved ship detection framework, M2S-YOLOv8, addressing limitations of current methods in complex marine settings.
• To enhance the perception of multi-scale and asymmetric vessel targets, improve feature fusion efficiency, and stabilize small-target regression.

Main Methods:

• Proposed M2S-YOLOv8 framework based on YOLOv8, incorporating a Multi-Scale Asymmetry-aware Parallelized Patch-wise Attention (MSA-PPA) module in the backbone.
• Introduced a Deformable Convolutional Upsampling (DCNUpsample) operator in the Neck for efficient adaptive feature fusion.
• Developed a Wasserstein-Distance-Based Weighted Normalized CIoU (WA-CIoU) loss function to improve small-target regression stability.

Main Results:

• M2S-YOLOv8 demonstrated improved performance on the Unmanned Vessel Zhoushan Perception Dataset (UZPD) and the Singapore Maritime Dataset (SMD).
• The framework achieved a balance between lightweight design and real-time inference capabilities.
• Enhanced perception of multi-scale and asymmetric targets, efficient feature fusion, and stable localization were observed.

Conclusions:

• M2S-YOLOv8 offers a robust solution for ship detection in challenging marine environments.
• The proposed enhancements significantly improve detection accuracy and localization stability.
• The framework shows strong potential for deployment on edge devices of unmanned marine platforms for reliable marine perception.