CDSANet: A CNN-ViT-Attention Network for Ship Instance Segmentation.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Ship instance segmentation is crucial for maritime applications like surveillance and port management.
• Challenges include dense ship distribution, scale/shape variations, and limited datasets.
• Existing methods like YOLOv8 struggle with global-local interactions and precise mask boundaries.

Purpose of the Study:

• To develop a novel one-stage network, CDSANet, for accurate ship instance segmentation.
• To improve the modeling of global-local interactions and feature representation for ship detection.
• To enhance segmentation accuracy and generalization capabilities in remote sensing imagery.

Main Methods:

• Proposed CDSANet integrates convolutional operations, Vision Transformers, and attention mechanisms.
• Utilized a Convolutional Vision Transformer Attention (CVTA) backbone for enhanced feature extraction.
• Employed dynamic-weighted DOWConv in the neck for multi-scale instance handling and SIoU loss for improved localization.

Main Results:

• CDSANet achieved a mask Average Precision (AP) of 75.9% on the VLRSSD dataset.
• Outperformed the YOLOv8 baseline by 1.8% in mask AP (50-95).
• Demonstrated superior performance in handling dense ship distributions and scale variations.

Conclusions:

• CDSANet offers a significant advancement in ship instance segmentation for remote sensing.
• The integrated architecture effectively addresses limitations of previous methods.
• The proposed network shows strong potential for real-world maritime surveillance and management applications.