ESS-H2RBox: A Remote Sensing Model for Small Object Detection

Remote sensing image analysis plays a critical role in fields such as urban planning and traffic management. To address the issues of missed detection and low accuracy in small target detection within the DOTA v1 dataset, this paper proposes an object detection network, ESS-H2RBox, which utilizes deep neural networks to extract and integrate multi-scale features and is trained using supervised learning with labeled data. ESS-H2RBox incorporates the EMA attention mechanism module into the backbone network to enhance the weighting capability of multi-scale features, significantly improving the detection accuracy for small and rotated targets. Additionally, this paper introduces the SSPAFPN module based on mmdet.FPN, which enhances feature fusion accuracy by modeling the similarity relationships between regions and applying an adaptive channel weighting mechanism. Furthermore, ESS-H2RBox integrates CIoULoss to optimize rotated bounding box detection, thereby significantly improving the model’s localization accuracy and robustness. Experimental results show that ESS-H2RBox achieves a 2.16% improvement in average precision compared to H2RBox on the large-scale remote sensing dataset DOTA v1, advancing the development of remote sensing object detection technology.