Physically Plausible Spectral Reconstruction.

Area of Science:

• Computational imaging
• Color science
• Computer vision

Background:

• Spectral reconstruction algorithms aim to recover spectral reflectance from RGB sensor data.
• Current deep learning methods achieve good spectral accuracy but lack physical correctness and robustness to exposure changes.
• Physically incorrect spectra do not accurately represent the original scene's spectral properties.

Purpose of the Study:

• To develop a physically accurate spectral reconstruction method that ensures recovered spectra induce the original RGB values.
• To improve the robustness of spectral recovery algorithms to variations in image exposure.
• To enhance color fidelity under different viewing conditions, including varying illuminations and cameras.

Main Methods:

• Developed a physically accurate recovery method based on the decomposition of spectra into a fundamental metamer and metameric blacks.
• Incorporated exposure variations into the training process to ensure robustness.
• Evaluated spectral recovery accuracy and RGB prediction under various conditions.

Main Results:

• The proposed method achieves physically accurate spectral recovery, provably inducing the same RGBs.
• Demonstrated improved and stabilized performance across varying exposure levels compared to state-of-the-art methods.
• Achieved zero colorimetric error and significantly improved color fidelity (up to 60% reduction in error).

Conclusions:

• The developed method offers a significant advancement in physically accurate spectral reconstruction.
• The approach provides robust and accurate spectral recovery, even with changes in exposure.
• This work enhances color fidelity and reliability in computational imaging applications.