Resilient Machine Learning Techniques for Improving Underwater Images

Color correction of underwater images for species identification, marine conservation, and other challenging tasks requires accurate visual information even in complex conditions and environment’s diversity. The recent Sea-Thru model, which comes from the atmospheric formation model, outlined how the attenuation coefficient of the signal is not uniform across the scene but depends on the object distance, light reflection, whereas also different from the coefficient governing distance backscattering. For the aforementioned reason, Sea-Thru, which uses a physical model of light attenuation model to improve underwater imaging, leverage a Structure From Motion (SFM) range map built using multiple images from different directions and the information of dark pixels for depth estimation. To improve Sea-Thru, a resilient methodological framework is here outlined, leveraging the advancements of self-supervised and monocular depth estimation (e.g., monodepth2) which propose a minimum reprojection loss to handle occlusion through a full-resolution multi-scale sampling method, ignoring loss in training pixels that violate camera motion assumptions with auto-masking procedure.