Performance Comparison of Classical Methods and Neural Networks for Colour Correction.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Color correction converts RAW camera RGB to standard color spaces like CIE XYZ.
• Traditional methods include linear, polynomial, and root-polynomial least-squares regression.
• Neural networks (NNs) are emerging as alternatives for color correction.

Purpose of the Study:

• Compare neural network (NN) color correction with regression methods.
• Evaluate NN performance against advanced regression techniques.
• Investigate and improve NN exposure invariance for robust color correction.

Main Methods:

• Comparative analysis of NN and regression (linear, polynomial, root-polynomial least-squares) models.
• Adaptation of regression methods to minimize perceptual color error.
• Development of exposure-invariant NN solutions using data augmentation and novel architectures.
• Cross-dataset training and testing to assess model generalization.

Main Results:

• NNs show improvement over simple least-squares but are surpassed by root-polynomial regression.
• Perceptual error minimization reduces the NN advantage over linear least-squares.
• NNs are sensitive to exposure changes; proposed solutions improve invariance.
• Regression methods consistently outperform NNs in color correction accuracy, even across different datasets.

Conclusions:

• Advanced regression techniques, particularly root-polynomial, offer superior color correction compared to current NN approaches.
• Exposure invariance is a critical challenge for NNs in color correction, addressed through data augmentation and architectural design.
• Regression methods demonstrate greater robustness and higher accuracy in color correction, especially when models are trained and tested on diverse datasets.