A Modality Alignment and Fusion-Based Method for Around-the-Clock Remote Sensing Object Detection.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Cross-modal remote sensing object detection faces challenges due to modality differences and feature degradation in adverse weather.
• Existing methods struggle with reliable around-the-clock detection, limiting applications in surveillance and reconnaissance.

Purpose of the Study:

• To develop a novel framework for robust cross-modal remote sensing object detection.
• To address challenges of modality disparity and feature degradation in visible light, infrared, and synthetic aperture radar (SAR) data.
• To enhance performance for 24/7 surveillance, military reconnaissance, and emergency response operations.

Main Methods:

• A multi-scale feature extraction module using hierarchical convolutional architecture to capture details and context.
• An innovative feature interaction module with cross-attention for long-range dependencies and adaptive noise suppression.
• A feature correction fusion module for spatial alignment and global feature consistency across modalities.

Main Results:

• The framework achieved state-of-the-art mean average precision (mAP) scores on challenging datasets: 66.3% (LLVIP), 58.6% (OGSOD), and 71.7% (Drone Vehicle).
• Demonstrated significant improvements over existing methods, particularly in scenarios with modality differences or extreme weather conditions.
• Validated effectiveness across visible light, infrared, and SAR modalities.

Conclusions:

• The proposed framework offers a robust solution for cross-modal object detection under challenging conditions.
• Advances the technical frontier of remote sensing object detection for critical 24/7 applications.
• Provides practical value for mission-critical operations requiring reliable around-the-clock surveillance and reconnaissance.