Color constancy based on texture pyramid matching and regularized local regression.

Considering that no single algorithm available is universal in color constancy, we propose an effective combination approach using a texture-based matching strategy and a local regression with prior-knowledge regularization. To represent the images, we construct a texture pyramid using an integrated Weibull distribution. Then we define an image similarity measure to search for the K most similar images of the test image. To combine the single algorithms, we integrate prior knowledge into a regularized local regression in a decorrelated color space. Regression weights are obtained on these similar images, and the regularization is implemented by the frequency ratio of the best single algorithm. Experiments on two real world datasets show our approach outperforms the state-of-the-art single algorithms and popular combination approaches with a performance increase of at least 29% compared to the best-performing single algorithm w.r.t median angular error.