Computational color constancy can be done in two steps, illumination estimation and color correction (also referred to as white balancing). The illumination estimation problem is ill-posed, since it is impossible to determine which combination of illumination and reflectance created the color that is seen by the camera sensor. The number of illuminants present in the scene defines the type of problem, i.e., single-illumination and multi-illumination problems. The focus of the thesis is on multi-illumination methods. These can be separated into multi-illumination estimation methods and image segmentation methods based on illumination. The goal of segmentation methods is to create a mask indicating the areas where only one illuminant is present. They are designed to work in concert with more well-researched single-illumination estimation methods. Multi-illumination estimation methods produce an estimation for each pixel in the scene. The focus is on the implementation and analysis of methods for both multi-illuminant estimation and segmentation. They are implemented using deep-learning-based techniques and achieve state-of-the-art performance on their respective problems. Deep-learning models based on convolutional layers are selected because they show great performance and generalization capabilities. For deep-learning methods, the question of the training dataset is important. For the proposed task, a real-world dataset with 2500 images was used for scenes with two sources of illumination. For scenes with three or more sources, an artificial dataset was created.