Dataset artificial augmentation with a small number of training samples for reflectance estimation

Area of Science:

• Computer Vision
• Color Science
• Machine Learning

Background:

• Accurate spectral reflectance estimation is crucial for many applications.
• Current methods are limited by the size and diversity of training datasets.
• Existing datasets often lack sufficient coverage and representation of color samples.

Purpose of the Study:

• To introduce a novel dataset augmentation technique for spectral reflectance estimation.
• To enhance the accuracy of reflectance estimation by increasing training data artificially.
• To evaluate the effectiveness of the proposed augmentation method on standard and real-world datasets.

Main Methods:

• Artificial dataset augmentation using light source spectra tuning.
• Generating a large number of augmented color samples from a small initial set.
• Performing reflectance estimation using both original and augmented datasets.
• Investigating the impact of augmented sample size on performance.

Main Results:

• Successfully augmented 140 Color Constancy Standard Group (CCSG) samples to over 13,791.
• Significantly improved reflectance estimation performance compared to benchmark CCSG datasets.
• Demonstrated superior performance across multiple standard datasets (IES, Munsell, Macbeth, Leeds) and a hyperspectral database.

Conclusions:

• The proposed dataset augmentation approach is practical and effective.
• Artificial augmentation of training data substantially enhances spectral reflectance estimation accuracy.
• This method offers a viable solution for improving color-related machine learning models.