This is the authors' abstract. We don't add key points for this paper.
Remote sensing object detection plays a crucial role in land resource monitoring, urban planning, and environmental protection. However, the detection process is often hindered by significant scale variations, complex background interference, and a surge in parameters resulting from high resolution inputs. These factors pose challenges for existing algorithms to balance accuracy and efficiency on lightweight edge devices. To effectively address these issues, the DyRSNet-Lite, an end-to-end lightweight dynamic multi-scale remote sensing object detection network, is proposed. First, a symmetric sampling flow structure is constructed to transition from static compression to dynamic recovery. Within this structure, the ADown hierarchical compression downsampling preserves subtle features, and offset adaptive upsampling restores details, enabling the network to adapt to extreme scale differences. Second, the DyRSFusion dynamic multi-scale feature fusion module is designed to achieve adaptive feature selection via large kernel convolutional sequences and a dynamic spatial kernel selection mechanism. This module significantly enhances context capture and detail retention in complex environments. Furthermore, the SPPF-Lite lightweight feature pyramid module is developed to maintain large-scale sensitivity while significantly reducing computational complexity through depthwise separable convolutions and multi-level max pooling with decreasing kernel sizes. Extensive experiments conducted on the DOTAv1.0, DIOR, and VisDrone datasets demonstrate that DyRSNet-Lite accurately detects multi-scale objects in complex backgrounds. With 1.90M parameters and 5.7 GFLOPs, the proposed network significantly outperforms current state-of-the-art methods and achieves an efficient balance between high accuracy and lightweight design.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: