Automatska anotacija skupa podataka semantičke segmentacije metodama polu nadziranog učenja

master's thesis

• Title (english): Automatic annotation of data set for semantic segmentation using semi-supervised learning methods
• Author: Petar Slaviček
• Mentor: Zdenko Kovačić (mentor)
• Committee member: Zdenko Kovačić (član povjerenstva)
• Committee member: Goran Vasiljević (član povjerenstva)
• Committee member: Dean Martinović (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2023-06-29, Croatia

Abstract

U ovom radu istražuje se problem semantičke segmentacije trsa u vinogradima s ciljem generiranja skupa podataka uz minimalno ljudsko anotiranje. Korištenjem VineSet skupa podataka, istrenirana je YOLOv5 detekcija trsa, koja se zatim iskoristila za generiranje skupa podataka rezanjem trsova iz slika. Izrezani trsovi su potom semantički segmentirani pomoću metode polunadziranog učenja SSL_ELN. Na generiranom skupu podataka provedena je evaluacija s nizom modela semantičke segmentacije i enkodera kako bi se odredio najprikladniji model za primjenu na mobilnom robotu u vinogradima. Rezultati pokazuju da je PSP model sa mobilenet_v2 enkoderom najbolji kompromis između točnosti i brzine predikcije te se čini najpogodniji za uporabu u ovoj specifičnoj primjeni.

Abstract (english)

In this work, we investigate the problem of semantic segmentation of grapevines in vineyards with the aim of generating a dataset with minimal human annotation. Using the VineSet dataset, YOLOv5 vine detection was trained, which was then used to generate a dataset by cropping vines from images. The cropped vines were then semantically segmented using the SSL_ELN semi-supervised learning method. Evaluation was conducted on the generated dataset with a series of semantic segmentation models and encoders to determine the most suitable model for application on a mobile robot in vineyards. The results show that the PSP model with the mobilenet_v2 encoder is the best compromise between accuracy and prediction speed, making it the most suitable for use in this specific application.

• Keywords: Semantička segmentacija; detekcija objekata; polunadzirano učenje
• Keywords (english): Semantic segmentation; object detection; semi-supervised learning
• Language: croatian
• URN:NBN: urn:nbn:hr:168:433798
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-10-11 10:27:33