PEFuse: Progressive Emphasis of Dual-Frequency Feature for Infrared and Visible Image Fusion

Infrared and visible image fusion (IVF) is a fine-grained cross-modal technique that integrates thermal cues from infrared images with detailed textures from visible images at the pixel level, enhancing visual quality and downstream task performance. However, existing methods for dual-frequency feature decoupling and fusion often lack effective high-frequency separation and sufficient interaction between high- and low-frequency features, limiting the preservation of fine textures and structural details. To address these challenges, we propose a framework for progressive emphasis of dual-frequency feature for IVF (PEFuse). Specifically, PEFuse employs a discrete cosine transform-based high-frequency extractor to disentangle texture and edge information, followed by a cross modulation collaborative fusion module that strengthens the complementarity between frequency components during early fusion. For frequency-specific refinement, we design a multikernel weighted convolution to enhance high-frequency details and a downsample top-k self-attention to capture low-frequency global context. Furthermore, an enhanced attention fusion module is integrated to progressively and adaptively guide the interaction and integration of frequency features throughout the fusion pipeline. Experimental results demonstrate that PEFuse achieves state-of-the-art fusion quality on standard infrared-visible benchmarks and significantly improves performance in downstream tasks, such as object detection.