Meistaravörn við verkfræðideild Orri Steinn Guðfinnsson
MSc í vélaverkfræði
Föstudaginn 13. janúar kl. 14:00 mun Orri Steinn Guðfinnsson verja 60 ECTS verkefni sitt til meistaragráðu í vélaverkfræði „Automatic behavior classification of aquatic animals in underwater videos“. Fyrirlesturinn fer fram í M102 og eru allir velkomnir.
Nemandi: Orri Steinn Guðfinnsson
Leiðbeinandi: Torfi Þórhallsson og Erik Martin Eineborg
Prófdómari: Guðmundur Einarsson
Útdráttur
This thesis investigates the feasibility of automatic classification of fish behavior in the cluttered low-visibility environment that typically surrounds towed fishing gear. A deep learning dataset was created from a limited amount of raw underwater video footage that depicts a scene in front of a fishing trawl where the conventional herding mechanism has been replaced by directed laser light. The goal is to train a Two-Stream network that can detect the presence of a fish in the videos and classify whether or not the fish reacts to the lasers. To achieve this, multiple versions of the dataset were created on which the two streams of the network were trained end-to-end. The results of the experiments were used to improve the classification accuracy of the network on the dataset by optimizing the model and the dataset. The results revealed that the Two-Stream network's spatial stream achieved an average accuracy of 90% when classifying between videos that contained a fish, and videos that did not. However, when the Two-Stream network was trained to classify the behavior of the fish depicted in the videos in a 3-class task, the accuracy achieved by the model was only 63.4%. With further inspection of the model's performance, the reason for the relatively low accuracy is most likely linked to the dataset, leaving room for improvement.
