Združevanje slik na podlagi nenatančnih mask

Črne, Ema (author), Čehovin Zajc, Luka (mentor)

Pri združevanju slik se večina pristopov osredotoča na izboljšanje mask, ki ločujejo ospredje od ozadja. Kot alternativo zahtevni in počasni metodi finega izboljševanja mask smo si v nalogi zastavili cilj doseči podobne rezultate samo z uporabo približnih mask in globokega učenja. Približne maske za učenje modela smo izpeljali iz podanih natančnih mask, ki smo jih sami deformirali. V okviru naloge smo preučili vplive različnih parametrov in za končni model izbrali tiste, ki so se izkazali za najuspešnejše. Končni model smo nato preizkusili tudi z raznovrstnimi maskami pridobljenimi z drugimi metodam za določanje mask. Model kljub svoji majhnosti in enostavnosti prikaže obetavne rezultate.

• Language: Slovenian
• Keywords: združevanje slik, konvolucijske nevronske mreže, globoko učenje, samokodirniki
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science ; FMF - Faculty of Mathematics and Physics
• Year: 2023
• PID: 20.500.12556/RUL-150185
• COBISS.SI-ID: 168447747
• Publication date in RUL: 14.09.2023

ČRNE, Ema, 2023, Združevanje slik na podlagi nenatančnih mask [online]. Bachelor’s thesis. [Accessed 18 April 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=150185

Secondary language

• Language: English
• Title: Image compositing with non-accurate masks

Abstract

In the field of image compositing most approaches focus on improving the precision of masks that distinguish between foreground and background. As an alternative to computationally intensive and time consuming image matting method, this research aims at achieving similar outcomes using only imprecise masks and the process of deep learning. These imprecise masks were created by deforming given exact masks. The work investigates the impact of different parameters and uses the combination of the most effective ones for the final model. The final model was then tested with a variety of masks obtained from other unrelated methods for foreground mask extraction. Despite its small size and simplicity, the model demonstrates promising results.

• Keywords: image compositing, convolutional neural networks, deep learning, autoencoders