RotFuserNet: A Feature-Enhanced and Multi-Scale Fusion Network for Rotated Object Detection in Remote Sensing Images

Abstract Detecting objects in remote sensing images is a demanding task due to complex and variable backgrounds, significant scale variations, and uneven object distributions. These challenges become even more acute when addressing tiny and rotated objects. Existing methods often fail to achieve robust detection due to limitations in feature extraction, insufficient contextual modeling, and ineffective label assignment strategies, which collectively reduce accuracy and reliability. To address these challenges, this paper introduces A Feature-Enhanced and Multi-Scale Fusion Network for Rotated Object Detection in Remote Sensing Images (RotFuserNet). The network utilizes reimagined large-kernel convolutional neural networks to improve feature extraction efficiency. Additionally, it proposes an Enhanced Feature Pyramid Network (En-FPN) to capture multi-scale object features while suppressing background noise. It employs a Context Anchor Attention mechanism to model both local and global contextual information effectively. RotFuserNet also introduces a Similarity Distance (SimD) label assignment approach, which adaptively evaluates bounding box similarity to achieve an optimal balance between positive sample selection and false positive reduction. Experiments on the benchmark DOTA-v1.0 dataset demonstrate the superior performance of RotFuserNet, achieving a 1.8% improvement in mean Average Precision (mAP) for small object detection and a precision score of 76.59%. Furthermore, as a generalization experiment, RotFuserNet was evaluated on the HRSC2016 dataset, achieving the mAP of 90.5%, showcasing its remarkable adaptability and robustness.