Frequency-aware wavelet fusion network for visible and near-infrared images

The fusion of visible and near-infrared (VNIR) images aims to synthesize enhanced images that integrate the unique features from two different modalities. However, prevalent deep learning methods usually struggle to achieve an optimal balance between fusing common and differential features. Moreover, conventional downsampling operations may lose critical spatial detail. To address these, we propose a frequency-aware wavelet fusion (FAWFuse) network, which enriches multi-scale features based on spatial-frequency relationships while performing feature fusion in the wavelet domain. The proposed FAWFuse comprises three core steps. First, a WavePool block is designed as the downsampler to enrich feature representations and preserve details adaptively by coordinating spatial and frequency information via a learnable weight adaptation mechanism. Second, we propose the dynamic frequency modulation (DFM) block that utilizes a dynamic attention mechanism for cross-modal frequency domain modulation to enhance the disadvantage features of VNIR images, and the enhanced features are fed into the wavelet adaptive fusion (WAF) block that merges each frequency component individually from the perspectives of global and local modulation. Third, we introduce a perception-inspired loss function, grounded in the frequency characteristics of the input images, to further improve fusion quality without ground-truth images. Extensive experiments demonstrate that the proposed method produces high-quality fused images with rich details and natural color, outperforming other representative methods both visually and quantitatively.