Physics-Driven Computational Multispectral Imaging for Accurate Color Measurement.

Area of Science:

• Computer Vision
• Materials Science
• Biomedical Imaging

Background:

• Accurate color measurement is vital for vision-based systems but is hindered by variable lighting, complexity, and subjectivity.
• Dental color measurement presents unique challenges due to strict perceptual and spectral fidelity requirements.

Purpose of the Study:

• To develop a deep-learned framework for accurate snapshot spectral reflectance prediction from RGB images.
• To address limitations of existing color measurement techniques in complex lighting scenarios, using dental applications as a validation case.

Main Methods:

• A deep-learned, end-to-end spectral reflectance prediction framework utilizing a physically interpretable network for feature fusion.
• Development of a dual-attention modular-information fusion neural network to recover spectral reflectance directly from RGB images.
• Creation of a dataset with 4000 RGB-hyperspectral image pairs under complex illumination using a custom optical system.

Main Results:

• The proposed framework achieved high accuracy in predicting teeth spectral reflectance, with a Mean Squared Error (MSE) of 0.0024 and a Structural Similarity Index Measure (SSIM) of 0.8724.
• Demonstrated effective performance across multiple scenarios for both natural teeth and ceramic materials.
• Successfully recovered spectral reflectance from RGB images under complex lighting conditions.

Conclusions:

• The deep learning framework provides a robust solution for accurate color measurement, overcoming metamerism-induced color mismatches.
• This advancement enables enhanced optical property characterization, 3D surface reconstruction, and computer-aided restorative design.
• The physically interpretable network design contributes to physically informed feature fusion for improved spectral prediction.