TL2Fusion: real-time lightweight image fusion technique on embedded platforms

Infrared and visible image fusion is crucial for multimodal perception, but high-performance models incur excessive computational costs for resource-constrained edge devices. This work introduces TL2Fusion, a unified compression-to-deployment framework that transforms the SeAFusion network into a hardware-efficient implementation. First, a dual-stream-adapted Tucker decomposition achieves substantial model compression. Second, a structured pruning strategy leverages hierarchical dependency analysis and L2-norm-based filter sensitivity to intelligently remove redundant convolutional filters. Finally, a targeted retraining phase restores model accuracy. Evaluations on TNO, FMB, and MSRS datasets demonstrate competitive fusion quality and superior performance in downstream segmentation tasks. Furthermore, by optimizing the GRDB module for FPGA deployment, TL2Fusion reduces memory footprint by 71.3–73.2% and improves power efficiency by 64.2% with minimal computational overhead. This research provides a validated pipeline for deploying real-time, energy-efficient multimodal perception systems at the edge.