Sammendrag
In this work, we propose a more accurate and robust white balance architecture for single and multi-illuminant scenes with low and high-frequency changes in illumination. In order to do that, we start by reviewing the state of the art of most important methods for white balance and illuminant estimation. Then, we introduce and analyze the large-scale multi-illuminant (LSMI) dataset that will be used for training. Next, we introduce the MEM-CNN architecture for multi-illuminant estimation, and we re-train and test it using the new dataset in order to analyze its performance. After that, we propose, implement and compare different improvements and optimizations to the original MEM-CNN architecture in order to enhance the accuracy and robustness of the predictions. Finally, we propose and evaluate an additional novel architecture that uses a more robust approach based on multiple estimations in order to achieve a combined solution and increase further the general white balance performance. In the experimental results, based on both visual and angular error comparisons, we conclude that our proposed model, named Advanced MEM-CNN, reaches good accuracy either in single or multi-illuminant scenes with low and high-frequency changes in illumination, thus being a good alternative to current state-of-the-art models.