Anotacija kolonoskopskih slik s temeljnimi modeli

diplomsko delo

Miha Lazić (Author), Žiga Emeršič (Mentor)

Abstract

Anotacija slik je ključen, a večkrat zamuden korak pri pripravi slikovnih podatkovnih zbirk. Poleg grafičnega vmesnika anotacijskega orodja, na hitrost anotacije vplivajo implementirani segmentacijski pristopi. Z razvojem globokega učenja na področju računalniškega vida se je pojavila možnost nadomestitve ročne anotacije in tradicionalnih segmentacijskih algoritmov s hitrejšimi in bolj natančnimi pristopi. Eden takšnih je temeljni model Segment Anything, ki smo ga analizirali v večih različicah (ViT-b, ViT-l, ViT-h, MobileSAM, SAM-Med2D, MedSAM) in testirali na podatkovni zbirki kolonoskopskih slik Kvasir-SEG in kolonoskopskih inštrumentov Kvasir-Instrument. Ovrednotili smo natančnost segmentacije in časovno zahtevnost modelov z resničnimi maskami objektov in na podlagi rezultatov, implementirali funkcionalnosti najboljšega modela v prototipni anotacijski program.

Keywords

anotacija;segmentacija;temeljni model; kolonoskopske slike ;visokošolski strokovni študij;diplomske naloge;

Data

Language:	Slovenian
Year of publishing:	2024
Typology:	2.11 - Undergraduate Thesis
Organization:	UL FRI - Faculty of Computer and Information Science
Publisher:	[M. Lazić]
UDC:	004.85:004.93:616.34 78.082.4(043.2)
COBISS:	208428803
Views:	151
Downloads:	39
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	Colonoscopy image annotation with foundation models
Secondary abstract:	Image annotation is a crucial but often time-consuming step in preparing image datasets. In addition to the graphical interface of the annotation tool, the speed of annotation is influenced by the implemented segmentation approaches. With the development of deep learning in the field of computer vision, the possibility has arisen to replace manual annotation and traditional segmentation algorithms with faster and more accurate approaches. One such model is the foundation model Segment Anything, which we analyzed in various versions (ViT-b, ViT-l, ViT-h, MobileSAM, SAM-Med2D, MedSAM) and tested on the Kvasir-SEG dataset of colonoscopic images and the Kvasir-Instrument dataset of colonoscopic instruments. We evaluated the segmentation accuracy and time complexity of the models with ground-truth object masks and, based on the results, implemented the functionalities of the best model into a prototype annotation program.
Secondary keywords:	annotation;segmentation;deep learning;vision transformer; foundation model;segment anything;colonoscopy;computer science;diploma;Računalniški vid;Strojno učenje;Kolonoskopija;Računalništvo;Univerzitetna in visokošolska dela;
Type (COBISS):	Bachelor thesis/paper
Study programme:	1000470
Embargo end date (OpenAIRE):	1970-01-01
Thesis comment:	Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages:	1 spletni vir (1 datoteka PDF (33 str.))
ID:	24862810