Synthetic subsea imagery for inspection under natural lighting with marine-growth

Gathering real-world high-quality data from underwater environments is cost-intensive, as is labeling this data for machine learning. Given this, synthetic data represents a possible solution that delivers ground-truth training data. Nevertheless, rendering and modeling of underwater environments are challenging due to several factors, including attenuation, scattering, and turbidity. The focus of this study is on the creation of a simulated underwater environment constructed for the purposes of simulating marine growth on offshore structures. The main requirement is the creation of renderings of sufficient quality and quantity with respect to the representation of marine-species distribution and intra-class variation, and sufficiently accurate recreation of lighting and turbidity (Jerlov water type) conditions underwater. Underwater rendering has been implemented using Blender, with a CAD model of an actual offshore installation and combined with marine growth from 2D/3D scanned and hand-modeled entities. The proposed approach provides for the generation of synthetic images usable for training computer vision models in marine-growth inspection applications as well as other related underwater applications. This has been demonstrated in a case study, wherein the utility of the rendered dataset has been briefly demonstrated in a neural network marine-growth segmentation task, targeting native fouling species for North-sea installations with a pixel-level fouling coverage. The produced renderings are available as a dataset of 1038 scene renders, using varying poses and randomized representative marine growth; each render includes RGB images, ground-truth segmentation masks, water-free RGB images, and depth information. In future work, the expansion with additional species and objects in other oceanic and coastal environments is envisioned. • Synthetic images rendered using Blender can represent subsea environments. • Effects of Jerlov water types and depth-dependent attenuation under natural light are captured. • Machine-learning segmentation of fouling is possible using synthetic data.