Cyclic Angle Encoding and Angle-Embedded Attention for Robust Oriented Object Detection in Remote Sensing Images

Rotated object detection in remote sensing images is crucial for various applications, including surveillance, urban planning, and environmental monitoring. However, this task is challenging due to the diverse orientations and scales of objects, as well as the complex background in aerial images. Traditional regression-based methods often suffer from boundary issues, leading to significant loss values and reduced detection accuracy. To address these challenges, we introduces a novel method for oriented object detection. We propose a cyclic angle encoding approach that converts angle prediction from a regression problem to a classification problem, effectively mitigating boundary issues and improving angle prediction accuracy. Additionally, we introduce an embedded angle information deformable attention mechanism to better align features with sampling points, enhancing the model's performance in rotated object detection tasks. Experiments on the DOTA-v1.0 dataset, a large-scale aerial image dataset, demonstrate the effectiveness of our method. Our approach achieves a mean average precision (mAP) of 80.98%, outperforming several state-of-the-art methods. Notably, our method achieves a precision of 90.70% in the aircraft category and shows robust performance across various categories and challenging scenarios. The proposed innovations in angle encoding and attention mechanism significantly improve detection accuracy and model robustness.