Infrared and visible image fusion using NSCT-HRCNN: a training-free feature extraction approach

Infrared and visible images capture distinct environmental features that image fusion can integrate into a single, information-rich representation. Traditional fusion algorithms often struggle with low precision, color distortion, and detail loss, while deep learning methods require extensive computational resources. We propose an image fusion algorithm combining the Non-Subsampled Contourlet Transform with a Hierarchical Random-Coupled Neural Network (NSCT-HRCNN). This training-free approach eliminates costly training and hyperparameter tuning, and operates on CPU-only devices without requiring high-end GPUs. It preserves natural color from visible images while often providing clear infrared details with reduced artifacts. Comparative experiments demonstrate that NSCT-HRCNN significantly outperforms state-of-the-art methods on the VIFB dataset. The code is publicly available at https://github.com/HaoranLiu507/NSCT-HRCNN-VIFusion.