Depth simulation of underwater images for data augmentation

Abstract

A method of simulating underwater images at a depth displaced by some desired factor is proposed with an application for data augmentation in machine learning. An at-depth PSF is generated and convolved with a recovered undegraded image to account for decreased resolution effects, and a transmission map is generated via single-image underwater dark channel prior (UDCP) and displaced by a distance d to reconstruct a final at-depth image. The accuracy of the depth simulation is compared between rescale only and PSF with rescale methods, using actual coral at-depth as the ground truth. The results of this validation show a high structural similarity for the simulated images compared with just a single rescale.