Refined One-Stage Oriented Object Detection Method for Remote Sensing Images.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Multi-class object detection in remote sensing is crucial but hindered by scale imbalance and arbitrary object orientations.
• Objects in remote sensing data often exhibit extreme aspect ratios, posing significant detection challenges.

Purpose of the Study:

• To enhance multi-class object detection in remote sensing images, particularly for objects with arbitrary orientations and extreme aspect ratios.
• To address scale imbalance and angle sensitivity issues in object detection models.

Main Methods:

• Proposed the Asymmetric Feature Pyramid Network (AFPN) with an asymmetric convolutional block to handle extreme aspect ratios.
• Introduced the Dynamic Feature Alignment (DFA) module to align mismatched features caused by anchor deviation.
• Developed a refined Area-IoU regression loss, combining area-guided and IoU-guided losses, to tackle scale imbalance and angle sensitivity.

Main Results:

• The proposed method demonstrated effectiveness on DOTA, HRSC2016, and ICDAR2015 datasets.
• The asymmetric convolutional block improved spatial representation with minimal computational overhead.
• The DFA module and Area-IoU loss successfully addressed feature misalignment and regression challenges.

Conclusions:

• The integrated approach of AFPN, DFA, and Area-IoU loss significantly improves multi-class object detection in remote sensing.
• The method offers a robust solution for detecting objects with challenging characteristics like extreme aspect ratios and arbitrary orientations.