Comparison of Artificial Neural Network and Polynomial Approximation Models for Reflectance Spectra Reconstruction

Area of Science:

• Computer Vision
• Spectroscopy
• Material Science

Background:

• Surface reflection knowledge is crucial for graphics and cultural heritage.
• Commercial RGB cameras offer high resolution and fast acquisition for reflectance spectrum (RS) mapping.
• Comparing artificial neural networks (ANN) and multivariate polynomial approximation (PA) for RS reconstruction is essential.

Purpose of the Study:

• To compare ANN and PA models for mapping RGB data to reflectance spectra.
• To investigate the impact of various parameters on ANN and PA model performance.
• To propose a profiling approach for optimizing ANN model parameters.

Main Methods:

• Utilized a training set of RGB-reflectance pairs for model development.
• Implemented and compared standard backpropagation (BP) and Levenberg-Marquardt (LM) learning algorithms for ANNs.
• Investigated parameters including hidden layers, neurons, polynomial degrees, inputs, and training set size.

Main Results:

• Two-layer ANNs with specific neuron configurations showed improved Mean Squared Error (MSE).
• ANN performance varied with learning algorithms (BP vs. LM) for one and two hidden layers.
• ANN models demonstrated superior fine-tuning capabilities compared to PA models.

Conclusions:

• ANN and PA methods for RS reconstruction are comparable, with ANNs offering advantages.
• ANN models can outperform PA models under realistic constraints due to better fine-tuning.
• A profiling approach aids in determining optimal ANN neuron configurations for varying training set sizes.