YOFOR : You only focus on object regions for tiny object detection in aerial images

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Deep learning has advanced object detection, but challenges remain in high-resolution remote sensing images.
• Complex backgrounds, uneven object distribution, and class imbalance hinder existing detector performance.

Purpose of the Study:

• To propose YOFOR (You Only Focus on Object Regions), an adaptive local sensing enhancement network.
• To address challenges in object detection for remote sensing images, including dense objects and class imbalance.

Main Methods:

• Developed an adaptive local sensing module to localize and crop dense object regions.
• Implemented a fuzzy enhancement module to reduce background interference and improve object visibility.
• Introduced a class balance module to mitigate the long-tailed class problem by analyzing class distribution and object proximity.

Main Results:

• The adaptive local sensing module effectively handles uneven object distribution.
• The fuzzy enhancement module improves detection by weakening background interference.
• The class balance module alleviates the long-tailed class problem, enhancing overall detection performance.
• All components are unsupervised and easily integrated into existing object detection networks.

Conclusions:

• YOFOR demonstrates significant effectiveness and adaptability across VisDrone, DOTA, and AI-TOD datasets.
• The proposed method offers a robust solution for object detection in challenging remote sensing scenarios.