DAF-Net: A Dual Attention Fusion Network with Window-Guided Attention and Invertible Modules for Infrared-Visible Image Fusion

Infrared and visible image fusion plays a pivotal role in mission-critical vision systems by combining complementary thermal radiation and textural information. While deep learning approaches have advanced beyond traditional multi-scale decomposition methods, existing solutions face fundamental limitations: convolutional neural networks (CNNs) inadequately model long-range dependencies, while vision transformers suffer from computational complexity and local feature extraction deficiencies. To address these challengesWe propose a Dual Attention Fusion Network (DAF-Net) featuring three key components: (1) A hybrid encoder combining convolutional and window-based attention operations to process both global thermal patterns and local textures, (2) An invertible fusion module that preserves critical features during modality integration, and (3) A multi-scale architecture maintaining spatial consistency through efficient depth-wise convolutions and attention-guided connections. Experimental results show that our method achieves competitive performance in standard metrics, outperforming both classical fusion baselines and recent advanced approaches, while maintaining computational efficiency comparable to mainstream CNN-based methods.