DMEFF-Net: A Dynamic Multiscale Enhanced Feature Fusion Model for Small Object Detection in Remote Sensing Images

作者：Guoqing Zhou, Linbo Yu, Ertao Gao, Yanlin Lu · 发表于：IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 年份：2025 · DOI：10.1109/jstars.2025.3634977 · 被引用次数：7 · 研究领域：Advanced Neural Network Applications、Remote-Sensing Image Classification、Advanced Image Fusion Techniques

The performance and efficiency of small object detection are still very unsatisfactory due to the complex background interference for remote sensing images (RSIs) and the scale diversity among small objects. To solve this problem, this paper proposes a new method, named “dynamic multi-scale enhanced feature fusion network (briefly, DMEFF-Net)” for small object detection in RSIs. The proposed framework integrates four key components: a multi-scale convolutional decoupling module (MCDM), which enhances feature representation by integrating multiple depth separable convolution layers with different dilation rates; a novel neck structure, the dynamic multi-scale feature fusion network (DMFFN), which is built by integrating the fine-grained detail aggregation module (FDAM) to effectively represent feature information across different scales; a residual spatial context awareness module (RSCAM), which captures the relationship between local and global features to preserve details and understand full image context; and an adaptive threshold focal loss (ATFL), which enhances small object detection by decoupling them from the dominant background. The model is evaluated on four public datasets, USOD, AI-TOD, NWPU-VHR 10, and RS-STOD, and compared with several state-of-the-art methods including RefineDet, TPH-YOLOv5, SP-YOLOv8s and FFCA-YOLO. The experimental results demonstrate that the proposed method achieves superior detection accuracy compared with existing approaches across all eva...