条件扩散模型驱动的两阶段近红外图像着色方法OA

Existing near-infrared(NIR)image colorization methods often suffer from luminance-chrominance confu-sion,texture blurring,and structural artifacts in the generated results.To address these issues,this paper proposes a dual-stage learning framework.In the first stage,a conditional diffusion model is employed to map NIR images into an intermediate grayscale domain that is aligned with the grayscale distribution of target visible-light(VIS)color im-ages while preserving high-frequency details.In the second stage,a mature pre-trained grayscale image colorization network is reused to colorize the predictions from the first stage,enabling high-fidelity color image reconstruction.This strategy effectively decouples the luminance and texture mapping process from the color reconstruction process,significantly reducing the dependence on large-scale paired NIR and VIS color image samples.Experimental results on the public dataset demonstrate that the proposed dual-stage learning framework achieves significant advantages in perceptual quality-related metrics,with the Fréchet distance and perceptual similarity metrics reaching 36.03 and 0.369 6,respectively,outperforming the compared methods,while maintaining comparable performance to state-of-the-art approaches in structure fidelity-related metrics.subjective visual comparisons indicate that the proposed method produces better color consistency and texture detail preservation.