Answers are grounded in this paper and cite their sources.
Small-target detection based on uncrewed aerial vehicle aerial imagery has been widely adopted in remote sensing. However, under complex environments and varying illumination conditions, single-modal processing of small targets often suffers from weak feature representation, making effective feature extraction a formidable challenge. To address these challenges, we propose a novel multimodal remote sensing detector, namely dual-modal hierarchical fusion network (DMHF-Net). This network offers four scalable model sizes to cater to deployment requirements across diverse scenarios. The DMHF-Net leverages frequency-split receptive field dilated convolution to expand the receptive field, enabling lightweight frequency-aware feature extraction. The extracted features are fed into the modality-channel joint attention fusion module, which fuses the dual-modal features and dynamically adjusts the feature weights of each modality to output feature maps with different scales and channel dimensions. Via the global–spatial channel joint weighting module and multimodal scale influence detection head, the importance of spatial regions and individual channels is adaptively assigned, further enhancing the feature expression of small targets. After detection, critical samples are weighted in the intersection-over-union-aware weighted loss and participate in backpropagation to improve the network. Extensive experiments were conducted on improved VEDAI, RGBT-tiny, LLVIP, DroneVehicle, and M3FD datasets. Our experimental results demonstrate that the DMHF-Net achieves remarkable performance levels, with an excellent tradeoff between detection accuracy and computational cost maintained across all model variants. The results also show the superiority over other state-of-the-art object detection methods.
No takes yet. Share an insight, caveat, or question.
Cao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: