Color Constancy Based on SAAF-SVR

Variations in illumination conditions significantly impact the color values in images, thereby affecting the accuracy of image processing tasks. This paper presents a self-attention autoencoding feature Support Vector Regression (SAAF-SVR) algorithm, which utilizes the probability distribution of L-r-g color space as original features. By incorporating a self-attention mechanism into an autoencoder, the features have been reconstructed. Then the SVR is used to estimate the illumination color. Experimental results demonstrate that, compared to other feature-based methods, the proposed approach effectively handles noise and diverse illumination variations in images, achieving notable improvements in both the accuracy and stability of color constancy.