The Effect of Physics-Based Corrections and Data Augmentation on Transfer Learning for Segmentation of Benthic Imagery

Ocean observation has been greatly improved by the use of Autonomous Underwater Vehicles and Remotely Operated Vehicles, and the high quality and high quantity of imagery they produce. This quantity of images collected on research cruises has, however, become intractable using traditional manual image analysis methods. There is a growing need for automation of this process, using methods such as deep learning to analyse and summarise the information present, yet research into how to improve their performance for underwater images is currently limited. This paper presents a study into the effect of using physics-based corrections and data augmentation to aid the performance of DeepLabV3, a state of the art image segmentation system. Using metadata about the physical environment the images were taken in, particularly the altitude of the vehicle for each image, and the known wavelength dependent light attenuation over distance through water, reduces the generalisation error of the DeepLabV3 system.