Haifeng Liang
Haifeng Liang has a total of 9 co-authors. Here are the top co-authors by number of shared publications:
- Ang Zhang
Technology Center of Qinhuangdao Customs, Qinhuangdao 066004, China
1 shared publication - Bin Fan
Institute of Optics and Electronics, Chinese Academy of Sciences, Chengdu 610209, China
1 shared publication - Bing Su
Biomedical Research Institute, Shenzhen PKU-HKUST Medical Center, Shenzhen, Guangdong 518036, China
1 shared publication - Dun Liu1 shared publication
- Jiang Bian1 shared publication
- Jianying Chen1 shared publication
Latest Publications
- Open Access
- 23 June 2026
- Published by Elsevier in SSRN Electronic Journal
GFE: Gated Flow Enhancement with Vision-Language Model Priors for Diffractive Snapshot Multispectral Imaging
Diffractive optical element (DOE)-based snapshot multispectral imaging provides a compact and lightweight solution for spectral acquisition. However, diffraction-induced degradations typically cause low contrast, structural blurring, and color cast, limiting both spatial and spectral imaging quality. Existing reconstruction methods still struggle to achieve satisfactory restoration quality under such degradations. Vision-language models (VLMs) can provide sharp and visually rich restoration priors, but their generated results may contain hallucinated textures or spurious details. Consequently, these priors cannot be directly injected into physical spectral reconstruction. To address this problem, we propose the Gated Flow Enhancer (GFE), a two-stage backend enhancement framework that conservatively uses VLM priors to improve reconstruction quality without modifying the DOE hardware or pretrained solvers. GFE first constructs an admitted guide image through a discrepancy-aware admission mechanism. It then enhances the initial multispectral estimate using gated flow updates within a frozen spectral latent space. This design selectively transfers beneficial structural information while suppressing content inconsistent with the real image. On multiple datasets and with different pretrained solvers, GFE consistently improves reconstruction quality, with average gains of 3.02 dB in PSNR and 0.098 in SSIM and a 2.66 reduction in SAM. To validate practical measurement performance, we further build a real DOE-based prototype and conduct real-world experiments using ColorChecker measurements and a 9-channel reference spectral camera. The results show that GFE improves both visual and spectral quality in real captures and yields more consistent spectral responses across independent measurement devices.
- 23 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessGFE: Gated Flow Enhancement with Vision-Language Model Priors for Diffractive Snapshot Multispectral Imaging
Diffractive optical element (DOE)-based snapshot multispectral imaging provides a compact and lightweight solution for spectral acquisition. However, diffraction-induced degradations typically cause low contrast, structural blurring, and color cast, limiting both spatial and spectral imaging quality. Existing reconstruction methods still struggle to achieve satisfactory restoration quality under such degradations. Vision-language models (VLMs) can provide sharp and visually rich restoration priors, but their generated results may contain hallucinated textures or spurious details. Consequently, these priors cannot be directly injected into physical spectral reconstruction. To address this problem, we propose the Gated Flow Enhancer (GFE), a two-stage backend enhancement framework that conservatively uses VLM priors to improve reconstruction quality without modifying the DOE hardware or pretrained solvers. GFE first constructs an admitted guide image through a discrepancy-aware admission mechanism. It then enhances the initial multispectral estimate using gated flow updates within a frozen spectral latent space. This design selectively transfers beneficial structural information while suppressing content inconsistent with the real image. On multiple datasets and with different pretrained solvers, GFE consistently improves reconstruction quality, with average gains of 3.02 dB in PSNR and 0.098 in SSIM and a 2.66 reduction in SAM. To validate practical measurement performance, we further build a real DOE-based prototype and conduct real-world experiments using ColorChecker measurements and a 9-channel reference spectral camera. The results show that GFE improves both visual and spectral quality in real captures and yields more consistent spectral responses across independent measurement devices.