Zebrafish Occlusion Detection
Author
Stjernholm, Niclas Hjorth
Term
4. term
Publication year
2019
Submitted on
2019-10-11
Abstract
Zebrafisk (Danio rerio) bruges som modelorganisme i biomedicinsk forskning, fx til at teste adfærdsmæssige reaktioner på forskellige stoffer. For at analysere adfærd spores hver fisks bevægelse i videooptagelser. Når flere fisk svømmer sammen i et akvarium, kan de dække for hinanden i billedet (okklusioner), så det bliver svært at følge hver enkelt fisk korrekt. Formålet med afhandlingen er at opdage sådanne okklusioner for potentielt at forbedre sporing og gøre det tydeligere, hvornår der kræves brugerindgriben. Afhandlingen foreslår to løsninger med forskellige grader af automatisering. Den første er en billedklassifikationsmodel, der afgør, om et billede indeholder en okklusion. Den anden er en objektdetektor baseret på Faster Region-based Convolutional Neural Network (Faster R-CNN), en dyb læringsmetode, som lokaliserer okkluderede områder og kategoriserer dem i én af seks foruddefinerede typer. Arbejdet præsenterer en ny kategorisering af zebrafisk-okklusioner og implementerer et proof-of-concept til multiklasse-objektdetektion, der kan finde flere okklusioner i samme billede.
Zebrafish (Danio rerio) are widely used as a model organism in biomedical research, for example to test behavioral responses to different substances. To analyze behavior, researchers track each fish’s movement from video. When several fish swim together in an aquarium, they can overlap in the image (occlusions), making it difficult to follow each individual accurately. This thesis aims to detect such occlusions to potentially improve tracking and to indicate when user intervention may be needed. Two solutions with different levels of automation are proposed. The first is an image classifier that decides whether a frame contains an occlusion. The second is an object detector based on Faster Region-based Convolutional Neural Network (Faster R-CNN), a deep learning method that localizes occluded regions and assigns them to one of six predefined occlusion types. The work introduces a new categorization of zebrafish occlusions and implements a proof-of-concept multi-class object detector capable of finding multiple occlusions in a single image.
[This abstract was generated with the help of AI]
Keywords
Documents