M-ReDet: A mamba-based method for remote sensing ship object detection and fine-grained recognition.

Area of Science:

• Remote Sensing and Geospatial Analysis
• Computer Vision and Image Processing
• Artificial Intelligence and Machine Learning

Background:

• Ship object detection and fine-grained recognition are crucial for applications like maritime surveillance and navigation.
• Existing methods struggle with accurate detection of diverse ship types in complex remote sensing imagery.
• Improving the precision of ship classification and localization is an ongoing challenge.

Purpose of the Study:

• To enhance the accuracy of ship object detection and fine-grained recognition in remote sensing images.
• To develop a novel method, Mamba-ReDet (M-ReDet), that improves upon existing ReDet architectures.
• To address the limitations in extracting and refining fine-grained features for different ship categories.

Main Methods:

• Proposed Mamba-ReDet (M-ReDet) incorporating a Mamba-ReResNet backbone for fine-grained feature extraction using Mamba's selective memory.
• Introduced Ship Object Perception Module (SOPM) and Ship Feature Refinement Module (SFRM) to extract and fuse spatial-positional information for feature refinement.
• Utilized KFIoU and Focal Loss for improved regression and classification accuracy during training.

Main Results:

• Mamba-ReDet achieved a mean Average Precision (mAP0.5) of 43.29% on the FAIR1M(ship) dataset.
• Mamba-ReDet achieved a mAP0.5 of 82.09% on the DOTAv1.0 visible light (RGB) dataset.
• Demonstrated performance improvements of 2.78% and 3.34% over the baseline ReDet on the respective datasets.

Conclusions:

• The proposed Mamba-ReDet (M-ReDet) effectively enhances ship object detection and fine-grained recognition in remote sensing images.
• The integration of Mamba's selective memory and specialized modules significantly improves feature extraction and refinement capabilities.
• The method shows superior performance compared to existing approaches, offering a valuable advancement for maritime applications.