关键词: 无人机航拍图像, 小目标检测, 可见光-红外, 注意力机制, 轻量化设计

Abstract: In unmanned aerial vehicle (UAV) aerial photography tasks, the core challenge for achieving precision detection in complex environments lies in effectively suppressing inter-modal interference while efficiently fusing cross-modal complementary features. To address this challenge, this paper proposes PCMA-Net, a dual-modal fusion algorithm for small object detection in UAV imagery. First, a parallel attention and gated recalibration module is designed. By capturing detailed features through a parallel dual-path architecture and employing dynamic gated masking for secondary feature refinement, this module suppresses complex background artifacts and enhances the model's discriminative power for small objects in complex backgrounds. Second, a cross-modal attention interaction module is constructed. By pre-modeling the response distribution to focus on salient regions, it utilizes bidirectional cross-attention to achieve pixel-level dynamic alignment of heterogeneous features, thereby resolving feature mismatch and submergence issues. Finally, we propose dual-branch attention guided convolution, which utilizes channel splitting and task-aware weight generation to achieve global context modeling with minimal overhead, thereby improving inference efficiency on embedded platforms. Experimental results on the DroneVehicle dataset demonstrate that the proposed method outperforms the baseline and state-of-the-art fusion algorithms by 1.9% and 1.7% in mAP@0.5, respectively. Furthermore, on the VEDAI dataset, PCMA-Net exceeds the baseline by 5.6% in mAP@0.5 and achieves state-of-the-art performance across most specific categories. Notably, PCMA-Net achieves a detection speed of 45.1 FPS, fully meeting the real-time requirements for UAV-based deployment.

Key words: aerial imagery, small object detection, visible-infrared, attention mechanism, lightweight design