Segmentacija in klasifikacija ultrazvočnih vokalizacij podgan

Absec, Marija (author), Marolt, Matija (mentor)

V diplomskem delu smo raziskali avtomatizirano detekcijo in klasifikacijo ultrazvočnih vokalizacij (USV) podgan. Analizirali smo uveljavljene metode, kot so MUPET, DeepSqueak in USVSEG, ter jih preizkusili na podatkovni zbirki Inštituta za patološko fiziologijo UL MF. Razvili smo lastno metodo, ki temelji na doučenju modela YOLOv11, ki hkrati izvaja detekcijo in klasifikacijo. Rezultati kažejo, da je naša metoda kljub enostavni zasnovi primerljiva z obstoječimi rešitvami in pri klasifikaciji dosega mero F1 nad 80 \%. Delo prispeva k napredku avtomatske analize USV in odpira možnosti za nadaljnje raziskave na področju.

• Language: Slovenian
• Keywords: ultrazvočne vokalizacije, segmentacija, klasifikacija
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2025
• PID: 20.500.12556/RUL-167933
• COBISS.SI-ID: 232069379
• Publication date in RUL: 20.03.2025

Secondary language

• Language: English
• Title: Segmentation and Classification of Rat Ultrasonic Vocalizations

Abstract

In this thesis, we explored the automated detection and classification of ultrasonic vocalizations (USVs) in rats. We analyzed established methods such as MUPET, DeepSqueak, and USVSEG and tested them on the dataset from the Institute of Pathophysiology at the Faculty of Medicine, University of Ljubljana. We developed our own method based on fine-tuning the YOLOv11 model, which performs both detection and classification simultaneously. The results show that despite its simple design, our method is comparable to existing solutions and achieves an F1 score above 80 \% in classification. This work contributes to the advancement of automatic USV analysis and opens up opportunities for further research in the field.

• Keywords: ultrasonic vocalizations, segmentation, classification