A statistical learning approach for underwater color restoration with adaptive training based on visual attention

In most artificial vision systems the quality of acquired images is directly related with the amount of information that can be obtained from them, and, particularly in underwater robotics applications involving monitoring and inspection tasks this is crucial. Statistical learning methods like Markov Random Fields with Belief Propagation (MRF-BP) provide a solution by using existing essential correlations in training sets. However, as in any restoration/correction method for real applications, it is not possible to have color ground truth available on-line. In this paper, we present a MRF-BP model formulated in the chromatic domain of underwater scenes such that we synthesize the ground truth color to train the model and maximize the capabilities of our method. The generated ground truth introduces some improvements to existing color correction methods and visual attention considerations which also helps to choose a small size training set for the MRF-BP model. Feasibility of our approach is shown from the results in which a good color discrimination is observed even in poor visibility conditions.