Določanje berljivosti napisov na poljubnih ozadjih s pomočjo algoritmov strojnega učenja

diplomsko delo

Elvisa Alibašić (Author), Luka Fürst (Mentor)

Abstract

Pri izdelavi vizualnega gradiva je pomembno, da so napisi na slikah in videu berljivi ne glede na ozadje. S tem problemom se srečujemo v številnih dejavnostih, še zlasti v oglaševalski in filmski industriji. Cilj diplomske naloge je razviti in preizkusiti metodo določanja berljivosti napisov na poljubnih ozadjih. Učno množico smo oblikovali s pomočjo anketiranja, in sicer tako, da smo za večjo množico slik anketirance povprašali, ali so berljive. Nato smo zajeli ključne podatke (kontrast, svetloba ipd.). Pri tem smo uporabili barvna modela RGB in HSL. Na osnovi zajetih podatkov smo zgradili linearni model. Odgovor na vprašanje berljivosti smo v okviru naloge obravnavali kot dvojiški, zato smo model zgradili s pomočjo logistične regresije. Zgrajeni model smo ovrednotili z metodami, kot sta AUC in prečno preverjanje. Končni klasifikacijski model je bil pri napovedovanju berljivosti napisov natančen v 68 odstotkih. Na podlagi rezultatov bo lahko oglaševalec med samodejno generiranimi oglasi izbral najbolj berljive.

Keywords

strojno učenje;oblikovanje;oglasi;računalništvo in informatika;univerzitetni študij;diplomske naloge;

Data

Language:	Slovenian
Year of publishing:	2020
Typology:	2.11 - Undergraduate Thesis
Organization:	UL FRI - Faculty of Computer and Information Science
Publisher:	[E. Alibašić]
UDC:	004.85(043.2)
COBISS:	27817219
Views:	743
Downloads:	176
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	Determining the legibility of text on arbitrary backgrounds using machine learning algorithms
Secondary abstract:	In visual media production (e.g., in marketing and film industry), it is important that the text on images and video is legible regardless of the background. The goal of the thesis is to develop and evaluate a method to determine the legibility of text on arbitrary backgrounds. The dataset was created using surveys. For a large dataset of photos, we asked the participants whether they are legible or not. Subsequently, we gathered key features (contrast, lightness etc.) by using the RGB and HSL color models. The gathered data were employed to build a linear model. Because we perceive legibility as binary, we used logistic regression. The model was evaluated using such methods as AUC and cross validation. The final classification model is 68% accurate at predicting legibility. Based on these results, advertisers can, from a set of generated ads, select the most legible.
Secondary keywords:	machine learning;design;ads;computer and information science;diploma;
Type (COBISS):	Bachelor thesis/paper
Study programme:	1000468
Embargo end date (OpenAIRE):	1970-01-01
Thesis comment:	Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages:	34 str.
ID:	12025905