Object Detection for Analysis of Piglets Feeding Behaviour in an Incubator
Author
Simonsen, Jonatan Emil
Term
4. term
Publication year
2022
Submitted on
2022-06-02
Pages
65
Abstract
Mange pattegrise har det svært i starten af livet. SEGES Innovation arbejder på en inkubator med kunstige patter for at reducere dødelighed kort efter fødslen. Dette speciale undersøger, om objektdetektion kan hjælpe ved automatisk at finde pattegrise ved patterne og afgøre, om de drikker. Vi afprøver flere variationer af Faster R-CNN, en udbredt dybdelæringsmodel til at finde objekter i billeder. Modellerne blev trænet og testet på et speciallavet datasæt indsamlet i samarbejde med SEGES Innovation. Målet var at opdage pattegrise, der bruger de kunstige patter i inkubatoren, og at vurdere, om en pattegris drikker. Ved en sikkerhedstærskel på 0,7 (hvor sikker modellen skal være, før den melder en observation) kunne systemet registrere drikkehandlinger med en gennemsnitlig præcision (AP, et standardmål for nøjagtighed) på 39,5 %, mens det at skelne mellem individuelle pattegrise havde en AP på 27,4 %. Samlet set opnåede modellerne ikke høj præcision i at opdage fodring ved de kunstige patter.
Many piglets face difficulties early in life. SEGES Innovation is developing an incubator with artificial teats to reduce deaths shortly after birth. This thesis examines whether object detection can help by automatically finding piglets at the teats and determining when they are drinking. We tested several variations of Faster R-CNN, a widely used deep-learning model for finding objects in images. The models were trained and evaluated on a custom dataset collected in collaboration with SEGES Innovation. The goal was to detect piglets using the artificial teats in the incubator and to judge whether a piglet was drinking. At a confidence threshold of 0.7 (how certain the model must be before reporting a detection), the system identified drinking events with an average precision (AP, a standard accuracy measure) of 39.5%, and distinguishing between individual piglets reached an AP of 27.4%. Overall, the models did not achieve high precision in detecting feeding at the artificial teats.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Documents