22 June 2026 in Neural Networks
23 January 2026 in IEEE Geoscience and Remote Sensing Letters
In recent years, object detection in visible (RGB) and infrared (IR) images has garnered extensive attention as a promising solution for achieving robust detection in all-weather scenarios. However, existing methods still face two major challenges: insufficient coordination between multimodal fusion and detection and inadequate feature extraction. To address these issues, this paper proposes a novel end-to-end network architecture for multimodal object detection, termed OSFGNet. Specifically, we introduce an OSM(object saliency-guided mechanism), which realizes adaptive fusion of ROI(region of interest) features from the detection head and fused features. This mechanism achieves collaborative optimization of multimodal features and detection while simplifying the training process, enabling integrated "fusion-detection" functionality. We design a tunable feature correlation factor that introduces learnable focusing parameters into attention mechanisms, dynamically adjusting attention distribution to balance local and global feature capture. Additionally, we embed an SFI(Spatial-Frequency Interaction) module to enhance fine-grained texture extraction through spatial-frequency interaction. Comprehensive experimental results demonstrate that our method achieves outstanding detection performance.