DCAF-Net: Dual-Branch Collaborative Attention Fusion Network For Multispectral Object Detection

To address the challenge of cross-modal feature collaboration and fusion for infrared and visible object detection in complex environments, this paper proposes a Dual-branch Collaborative Attention Fusion Network. The network constructs a frequency-domain and spatial-domain coupled framework. Through its dual-domain attention module, it leverages frequency-domain self-attention to model global dependencies and combines channel-spatial attention to achieve spatial calibration, thereby enhancing single-modal features. A multi-head cross-attention module is designed to decouple cross-modal interactions into semantic subspaces, enabling deep fusion through independent learning within attention heads followed by gated integration. A lightweight dual-branch backbone network is adopted to preserve modal independence while reducing model complexity. Experiments demonstrate that this method outperforms mainstream algorithms in both detection accuracy and speed on the DroneVehicle dataset, with significant improvements in small object detection performance.