Structure-preserving spectral reflectance estimation using guided filtering.

Area of Science:

• Computational imaging
• Spectral imaging
• Image processing

Background:

• Multispectral cameras provide valuable spectral information for material classification.
• Estimating light spectra from multispectral images is crucial but challenging.
• Noise significantly impacts spectral reconstruction, especially in video capture due to short exposure times.

Purpose of the Study:

• To develop a novel algorithm for robust spectral reconstruction from noisy multispectral images.
• To reduce the influence of noise on spectral estimation without requiring camera calibration or training data.

Main Methods:

• A novel spectral reconstruction algorithm based on guided filtering is proposed.
• The guided filtering technique leverages spatial information to preserve image structures and reduce noise.
• The method was evaluated using spectra from natural images and a real-world nine-channel multispectral camera.

Main Results:

• The proposed algorithm significantly outperforms state-of-the-art spatial reconstruction methods in noisy scenarios.
• Mean squared error and spectral angle were reduced by up to 46% and 35%, respectively, in noisy conditions.
• The technique demonstrated effective 'out-of-the-box' performance without calibration or training.

Conclusions:

• The guided filtering-based spectral reconstruction method offers a robust solution for noisy multispectral imaging.
• This approach improves the accuracy and reliability of spectral estimation for material classification tasks.
• The algorithm's independence from calibration and training makes it broadly applicable to various multispectral imaging systems.