Classification and Localization of Sea Animals in Brackish Waters with Diverse Visibility Using Deep Learning

Stranovsky, Lukás

4. term

Vision, Graphics and Interactive Systems, Master

2020

2020-05-20

92

Vandforurening er stigende, og Den Europæiske Union kræver, at medlemslande overvåger deres havmiljøer. En måde at vurdere forurening på er at se på ændringer i dyrs adfærd. Det skaber behov for computersyn, der automatisk kan overvåge dyr under vandet. Ydelsen afhænger dog af sigtbarheden, som påvirkes af turbiditet—vandets uklarhed forårsaget af partikler i vandet. Dette speciale bygger på tidligere arbejde med et særligt Brackish-datasæt (fra områder, hvor fersk- og saltvand mødes), der er stærkt påvirket af turbiditet. Målet er at undersøge, hvordan turbiditet påvirker en undervandsobjektdetektor—et program, der finder og klassificerer dyr i billeder. Ved at bruge YOLOv3, en hurtig detektor i realtid, forbedrer projektet de oprindelige basisresultater med 9,2% i mAP (mean Average Precision, et udbredt mål for detektionsnøjagtighed). Forbedringen bekræftes på et nyt, manuelt annoteret datasæt kaldet Brackish X, der bruges til at vurdere, hvor godt en trænet model generaliserer til nye data. Specialet undersøger også, hvilke visuelle kendetegn der bedst kan bruges til at estimere turbiditetsniveauet. Til sidst testes den bedste model på det oprindelige testsæt, opdelt i tre undergrupper efter den estimerede turbiditet, for at se, hvordan sigtbarhed påvirker detektion. Resultaterne peger på næste skridt for forsøg i kontrollerede omgivelser.

Water pollution is increasing, and the European Union asks member states to monitor marine environments. One way to gauge contamination is by observing changes in animal behavior. This creates a need for computer vision systems that can automatically monitor underwater animals. However, performance depends on underwater visibility, which is influenced by turbidity—the cloudiness of water caused by suspended particles. This thesis builds on earlier work using a unique Brackish dataset (from waters where fresh and salt water mix) that is strongly affected by turbidity. The goal is to examine how turbidity influences an underwater object detector—a program that finds and classifies animals in images. Using YOLOv3, a fast real-time detector, the project improves the baseline accuracy reported in the prior study by 9.2% in mAP (mean Average Precision, a common measure of detection accuracy). The improvement is confirmed on a newly created, manually annotated dataset called Brackish X, which helps assess how well a trained model generalizes to new data. The thesis also investigates which visual features are most useful for estimating turbidity levels. Finally, the best-performing model is tested on the original test set after splitting it into three groups based on estimated turbidity, to see how visibility affects detection. The results provide direction for future experiments in controlled environments.

[This abstract was generated with the help of AI]

Underwater vision ; deep learning ; computer vision ; turbidity

Download
View record in AAU Student Projects