Detekcija kapljica kiše na zaštitnom staklu između scene i kamere

undergraduate thesis

• Title (english): Detection of raindrops on the protective glass between the scene and the camera
• Author: Josip Srzić
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Franjo Matković (član povjerenstva)
• Committee member: Zorana Doždor (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2022-07-08, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Mogućnost detekcije kapljica kiše važna je u mnogim modernim automatiziranim
sustavima. Ovaj rad doskače tom problemu korištenjem suvremenih tehnika iz područja računalnog vida. Cilj je bio napraviti klasifikator koji će za danu sliku utvrditi
nalaze li se na njoj kapljice ili ne. U radu su objašnjeni osnovni principi strojnog učenja
i konvolucijskih neuronskih mreža. Opisan je korišteni skup podataka te programska
implementacija u sklopu razvojnog okvira PyTorch. Odabran je model ResNet-18
prednaučen na skupu ImageNet koji se već pokazao pouzdan na mnogim klasifikacijskim
zadacima. Isprobane su razne kombinacije hiperparametara i pronađen je najbolji
model metodom unakrsne validacije. Ostvarena je točnost od 97.71% na skupu za validaciju
i 92.26% na skupu za testiranje. Na kraju je napravljena diskusija postignutih
rezultata i potencijalne primjene modela u stvarnim sustavima.

Abstract (english)

The ability to detect raindrops is important in many modern automated systems.
This paper addresses this problem by utilizing modern techniques in the field of computer
vision. The goal was to create a classifier that can determine for a given image
whether it contains raindrops or not. The paper explains the basic principles behind
machine learning and convolutional neural networks. The dataset used and the software
implementation written using PyTorch are described. A ResNet-18 model initialized
with weights from the ImageNet dataset was selected, which has already proven
reliable on many classification tasks. Various combinations of hyperparameters were
tested and the best model was found using the cross-validation method. An accuracy
of 97.71 % on the validation set and 92.26 % on the test set was achieved. Finally,
a discussion of the results achieved and the potential application of the model in real
systems was made.

• Keywords: duboko učenje; računalni vid; klasifikacija slika; detekcija kapljica; konvolucijske neuronske mreže; duboki modeli s rezidualnim vezama; ResNet-18
• Keywords (english): deep learning; computer vision; image classification; raindrop detection; convolutional neural networks; deep residual networks; ResNet-18
• Language: croatian
• URN:NBN: urn:nbn:hr:168:719136
• Study programme: Title: Computing; Study programme type: university; Study level: undergraduate; Academic / professional title: sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva (sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-15 11:15:02