Infrared and visible image object detection based on dual-attention multiscale fusion network

This paper proposes DAMFNet, a dual-attention multi-scale fusion network integrated with YOLOv8 for robust infrared-visible object detection under complex environments. The framework addresses three key challenges through: (1) a cross-modal feature alignment module (CFAM) with deformable convolutions and spatial transformers for pixel-semantic alignment; (2) a dual-attention multi-scale fusion module (DAFM) combining frequency-domain enhancement with channel-spatial attention to suppress redundancy; and (3) a contrastive guidance loss (LCGL) synergizing detection and alignment losses for discriminative feature learning. The design significantly improves cross-modal detection accuracy in diverse scenarios.