Model Pruning for Infrared-Visible Image Fusion

Infrared-visible image fusion combines complementary information from both modalities, enhancing scene perception in applications such as surveillance and autonomous driving. However, existing deep learning-based methods are often computationally expensive. Our approach involves training an over-parameterized fusion network, applying structured pruning to reduce model complexity, and fine-tuning the pruned model to maintain performance. The pruning process leverages the L2-norm of the Restormer blocks, ensuring that less critical components are removed while preserving essential fusion quality. Experiments on the benchmark datasets demonstrate that our approach achieves high fusion quality with significantly reduced computational costs. Ablation studies further validate the effectiveness of our pruning strategy.