红外线的
材料科学
图像融合
融合
光学
计算机视觉
人工智能
图像处理
计算机科学
热成像
遥感
作者
Chenyu Zhang,Junyue Sun
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2026-06-01
卷期号:: 42-50
标识
DOI:10.1109/icbase70763.2026.11619411
摘要
To address the challenges of high inference latency and difficulty in edge deployment faced by current diffusion models in infrared and visible image fusion, this paper proposes a lightweight fusion method based on offline knowledge distillation and structural re-parameterization. A pre-trained diffusion model serves as the teacher network, and a two-stage offline distillation framework is constructed to transfer the generated cross-modal diffusion priors to a lightweight student network. The student network is designed upon the principle of structural reparameterization: during training, a multi-branch Edge-Enhanced Re-parameterization Block (EERB) is employed to strengthen feature representation, which is equivalently folded into a single-path convolution during inference for ultra-fast deployment. Furthermore, a multi-level synergistic optimization engine incorporating color-difference-aware and edge-aware losses is proposed to guide the student network in overcoming the performance bottleneck imposed by model compression. Extensive experiments on multiple datasets demonstrate that the proposed method achieves substantial improvements over the teacher model in core metrics such as mutual information fidelity and color difference control, while maintaining extreme inference speed, successfully reconciling high-fidelity fusion quality with real-time processing performance.
科研通智能强力驱动
Strongly Powered by AbleSci AI