logo SBA

ETD

Digital archive of theses discussed at the University of Pisa

 

Thesis etd-01262023-213301


Thesis type
Tesi di laurea magistrale
Author
DEL TURCO, FRANCESCO
URN
etd-01262023-213301
Thesis title
Deep Learning methods for Visual Fish re-identification
Department
INGEGNERIA DELL'INFORMAZIONE
Course of study
ARTIFICIAL INTELLIGENCE AND DATA ENGINEERING
Supervisors
relatore Prof. Cimino, Mario Giovanni Cosimo Antonio
relatore Prof. Falchi, Fabrizio
relatore Prof. Carrara, Fabio
relatore Prof. Bibbiani, Carlo
relatore Dott.ssa Sangiacomo, Chiara
Keywords
  • artificial intelligence
  • deep learning
  • metric learning
  • neural network
  • re-identification
  • recall
  • triplet loss
  • zebrafish
Graduation session start date
17/02/2023
Availability
Withheld
Release date
17/02/2026
Summary
Fishes are widely used in scientific research due to some particular characteristics, like the strong similarity of their genome with the humans’ one or the very short life cycles; moreover, many of them provide simpler systems than other animals like mice and pigs for the study of complex processes. A typical task to be performed when studying different fish individuals during their life is re-identification, in which individuals of the same species under study must be recognized at different growth stages in order to understand the progress of the breeding, matching different weights and lengths obtained at different times. For this task, in some cases, it is enough to use GPS tags or microchips, but they’re not a feasible solution for tiny fishes like the Danio rerio (zebrafish), which is one of the most used animals in scientific research, in particular considering their juvenile stages. The purpose of this study is, therefore, to exploit a metric learning approach applied through a Triplet Loss Network to build software able to perform, starting from an input image, the re-identification task in order to ease the whole study process by providing the most similar identities to the researcher and decrease the selection pool for re-identification. With this approach, we were able to obtain a 96% recall@10 over a group of 30 identities and 44.57% over 180 identities taken from our own zebrafish dataset.
File