HyperColorization: propagating spatially sparse noisy spectral clues for reconstructing hyperspectral images

Area of Science:

• Computational Imaging
• Image Processing
• Spectroscopy

Background:

• Hyperspectral cameras present significant spatial-spectral resolution trade-offs.
• These cameras are highly susceptible to shot noise, impacting image quality.
• Existing methods struggle to balance resolution and noise reduction.

Purpose of the Study:

• To introduce a colorization algorithm for reconstructing hyperspectral images from limited data.
• To demonstrate the algorithm's generalizability across different spectral dimensions.
• To improve the efficiency and robustness of hyperspectral image reconstruction.

Main Methods:

• Reconstruction of hyperspectral images using a grayscale guide image and sparse spectral clues.
• Application of a low-rank space for colorization to reduce computational load and noise.
• Incorporation of guided sampling, edge-aware filtering, and dimensionality estimation for enhanced robustness.

Main Results:

• The algorithm successfully reconstructs dense hyperspectral images from sparse samples.
• Colorization in a low-rank space significantly reduces compute time and shot noise impact.
• The proposed method outperforms previous algorithms in key performance metrics (SSIM, PSNR, GFC, EMD).

Conclusions:

• The developed algorithm offers a viable solution to the spatial-spectral-temporal resolution trade-off in hyperspectral imaging.
• This approach enables the reconstruction of high-quality hyperspectral images from various scanning systems.
• The findings pave the way for more efficient and robust hyperspectral data acquisition and processing.