This would be a great application for deep learning. Use the authors' method to generate a lot of uncorrected-corrected pairs. Or, use a graphics engine to rend

This would be a great application for deep learning. Use the authors' method to generate a lot of uncorrected-corrected pairs. Or, use a graphics engine to render realistic underwater scenes with and without water color. Then use a convolutional neural network to learn to mimic the transformation. Then any photographer can apply the learned filter without a color card or depth information. Edit: already been done: https://arxiv.org/abs/1702.07392 https://arxiv.org/abs/1702.07392