Performance evaluation of semantic segmentation in subsea conditions using AI trained on virtual images: Master’s Thesis Project in Sustainable Energy Engineering with Specialization in Offshore Energy Systems
Translated title
Performance evaluation of semantic segmentation in subsea conditions using AI trained on virtual images
Author
Ozalla Sanchez, Miguel
Term
4. semester
Education
Publication year
2025
Submitted on
2025-06-02
Pages
120
Abstract
Inspektion af undervandsinfrastruktur er afgørende, men uklart vand (turbiditet) gør det ofte svært at se vigtige dele. Dette speciale undersøger, om en computervisionsmodel stadig kan finde og afgrænse objekter under sådanne forhold. Vi skaber syntetiske træningsdata (computergenererede billeder), så vi præcist kan styre, hvor uklart vandet er. I Blender renderes 1.800 billedpar: et råbillede og en tilhørende segmenteringsmaske, som angiver for hvert pixel, om det tilhører et objekt eller baggrunden. Vi træner et semantisk segmenteringsnetværk (DeepLab v3+ med en ResNet-18 backbone), en dyb læringsmodel der mærker billeder på pixelniveau. Der laves fem datasæt: tre med lav, middel og høj turbiditet; ét uden turbiditet eller diffraktionseffekter (næsten som om der ikke er vand); og ét, der kombinerer alle turbiditetsindstillinger med forskellige vandfarver. Netværket trænes separat på hvert datasæt og evalueres på et turbiditets-testdatasæt for at se, hvor følsom ydeevnen er over for de forhold, modellen er trænet på. Vi sammenligner også modeller trænet kun på lav, middel eller høj turbiditet for at undersøge, hvordan mangfoldigheden i træningsdata påvirker prædiktionernes nøjagtighed. Studiet giver praktiske indsigter til at udvikle mere robuste visionsystemer til f.eks. rørledningsinspektion og vurdering af konstruktioners tilstand under vand, hvor pålidelig objektdetektion i uklart vand er afgørende for sikkerhed og effektiv drift.
Inspecting underwater infrastructure is vital, but murky water (turbidity) often makes it hard to see important parts. This thesis tests whether a computer vision model can still find and outline objects in such conditions. We generate synthetic training data (computer-made images) so we can precisely control how cloudy the water is. In Blender, we render 1,800 image pairs: a raw image and a matching segmentation mask that marks each pixel as belonging to an object or the background. We train a semantic segmentation network (DeepLab v3+ with a ResNet-18 backbone), a deep learning model that labels images at the pixel level. Five datasets are created: three with low, medium, and high turbidity; one with no turbidity or diffraction effects (similar to having no water); and one that combines all turbidity settings with different water colors. We train the network separately on each dataset and evaluate it on a turbidity test set to see how sensitive performance is to the conditions seen during training. We also compare models trained on only low, medium, or high turbidity to study how the diversity of training data affects prediction accuracy. The study provides practical guidance for building more robust vision systems for tasks such as pipeline inspection and assessing the structural health of underwater infrastructure, where reliable object detection in cloudy water is essential for safety and efficiency.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Documents