Color Constancy Algorithms for Multi-source Non-uniform Scenes Based on Local Estimation

Abstract

This paper presents a color constancy solution for multi-source non-uniform scenes. Firstly, the image partitioning strategy based on Tk-means is used to realize the conversion of multi-light source non-uniform scene to single source uniform scene. A single source color estimate is then performed on the local scene using the BP neural network algorithm. Aiming at the characteristics of multi-light source non-uniform scene, it is proposed to use FF attention mechanism to estimate the weight of the light source color estimated by each local scene, and obtain a composite light source as an approximate estimation of the scene light source. Finally, the paper reconstructs the image through the diagonal model. The experimental results show that the proposed strategy can solve the color constancy problem of multi-source non-uniform scenes.