EMSF-Net: An Efficient Multiscale Fusion Network With Lightweight Attention for Object Detection in Remote Sensing Images

Remote-sensing object detection is an important technique for addressing global challenges, including climate change and environmental pollution. However, existing methods still suffer from limited model adaptability, high miss-detection rates, and considerable deployment costs, particularly in UAV-based and edge-computing scenarios. To tackle these issues, an Efficient Multi-Scale Fusion network, termed EMSF-Net, is proposed by integrating the strengths of (Convolutional Neural Networks) CNNs and the Swin Transformer. The network incorporates an efficient multi-scale fusion module (MSF-Block) to achieve complementary integration of global semantic information and local fine-grained details, a lightweight detection head (LDetect) to improve classification and localization performance, two feature-enhancement modules-an enhanced feed-forward module (C2PSA-GDFN) and a depth-enhanced fusion module (C2D)-that jointly strengthen feature representation and fusion. In addition, a multi-scale optimization strategy based on the P2 prediction layer and depthwise convolution (DWConv) is constructed to reinforce fine-grained feature extraction and spatial localization. A customized SIoU loss function combined with a loss-weight optimization strategy is further designed to enhance bounding-box regression accuracy and training stability. Experimental results demonstrate that EMSF-Net achieves improvements of 8.4%, 3.3%, 8.8%, and 5.4% in mAP over the baseline model YOLOv11s across four mainstream datasets, while maintaining a favorable balance between detection accuracy and computational efficiency. Moreover, the reliability and generalization performance of the proposed model are further validated on the SIMD and VEDAI dataset, offering an effective and reliable approach for remote-sensing object detection in resource-limited settings.