Machine Learning-Based Estimation of Correlated Color Temperature from Raw Image Data

Correlated Color Temperature (CCT) is vital in lighting applications, influencing human perception, mood, and productivity. Traditional estimation methods using spectrophotometers are accurate but costly, highlighting the need for accessible alternatives. This study presents an image-based CCT estimation method using machine learning models applied to RAW images of a three-step grayscale ColorChecker chart. Ground-truth CCT was recorded using a Konica Minolta CL-500A Illuminance Spectrophotometer in a controlled setup. A DSLR camera with a fixed focal-length lens captured 1200 images under varied lighting. MATLAB was used to extract average pixel values from grayscale patches, forming inputs for regression models. Two models were implemented: a Bayesian Neural Network (BNN) trained with Bayesian Regularization and a Matern Gaussian Process Regression (GPR) model. Both were validated using a diverse dataset. Results showed that the Matern GPR model consistently achieved absolute estimation errors below 8%, outperforming the BNN in accuracy and generalization. This demonstrates the potential of cost-effective, camera-based CCT estimation for applications in photography, cinematography, architectural lighting, and human-centric lighting systems, offering a practical alternative to traditional spectrophotometry.