Learning geometric Jensen-Shannon divergence for tiny object detection in remote sensing images.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Tiny objects in remote sensing images present unique detection challenges due to their limited pixel representation.
• Standard object detectors struggle with feature extraction and are sensitive to Intersection-over-Union (IoU) thresholds for small objects.

Purpose of the Study:

• To develop a specialized detector, JSDNet, to overcome the limitations of general object detectors in tiny object detection.
• To improve the accuracy and robustness of tiny object detection in remote sensing data.

Main Methods:

• Integration of Swin Transformer as a backbone for enhanced feature extraction of tiny objects.
• Modeling anchor boxes and ground-truth as 2D Gaussian distributions for a statistical representation of tiny objects.
• Introduction of the JSDM module, utilizing geometric JS divergence for robust anchor box regression, mitigating IoU sensitivity.

Main Results:

• JSDNet demonstrated superior performance in detecting tiny objects.
• The proposed method outperformed state-of-the-art general object detectors on benchmark datasets (AI-TOD and DOTA).

Conclusions:

• JSDNet effectively addresses the challenges of tiny object detection by employing a novel statistical distribution-based approach.
• The network's design, incorporating Swin Transformer and JS divergence, offers a significant advancement in remote sensing image analysis for small object identification.