WDHN-Det: Wavelet-Coordinated Dynamic Hyperbolic Normalization for Multisource Remote Sensing Object Detection

Multi-source remote sensing image object detection is essential for disaster monitoring and national defense. However, challenges remain due to heterogeneous imaging mechanisms, limited multi-scale feature representation, and complex background interference. To address these issues, we propose WDHN-Det(Wavelet-Coordinated Dynamic Hyperbolic Normalization for Multi-Source Remote Sensing Object Detection), a unified object detection framework designed for robust and accurate detection in multi-source remote sensing scenarios. The framework incorporates three core modules: (1) a Cross-Scale Wavelet Feature Fusion (CSWFF) module that utilizes Haar wavelet decomposition in place of conventional downsampling operations, enabling cross-scale feature integration while preserving edge details and reducing high-frequency information loss; (2) a Dynamic Hyperbolic Pyramid Squeeze Attention (DyHPSA) module that combines dynamic attention with hyperbolic normalization, enabling adaptive feature modulation through learnable curvature parameters to improve robustness against outliers; and (3) a Cross-Modal Adaptive Fine-Grained Pyramid Fusion (CM-AFGPF) module that embeds dual-dimension attention into the SPPF structure, refining feature granularity and boosting detection accuracy. Extensive experiments conducted on the RSOD (optical) and SSDD (SAR) datasets show that WDHN-Det achieves state-of-the-art performance, with mAP/Precision of 95.86%/95.08% on RSOD and 98.7%/98.7% on SSDD. Compared with baseline methods, WDHN-Det improves mAP by 2.68% and 0.41%, and Precision by 6.68% and 0.66% on RSOD and SSDD, respectively, while maintaining high detection efficiency. These results demonstrate the effectiveness and generalization capability of WDHN-Det for multi-source remote sensing object detection. The project is publicly available on https://github.com/cicadachan/WDHN-Det.