Barvanje črnobelih slik z globokimi modeli

Godec, Primož (author), Zupan, Blaž (mentor)

Barvna fotografija je prišla v vsakdanjo uporabo šele v zadnjih 50 letih, zato so razni arhivi polni črno-belih fotografij, katere bi njihovi lastniki radi obarvali. V ta namen so bili razviti različni algoritmični pristopi. V disertaciji predstavljamo nekaj novih avtomatskih pristopov za barvanje črno-belih slik in videov, ki so osnovani na strojnem učenju in konvolucijskih nevronskih mrežah. Pristope primerjamo s pristopi iz sorodnih del in jih preizkusimo na starih črno-belih slikah. Iz rezultatov je razvidno, da naši pristopi dosegajo kvaliteto barvanja pristopov iz sorodnih del. Naš nov pristop, ki obarva slike po delih, pa izboljša barvanje slik velikosti, ki so različne od tistih, na katerih je bila mreža naučena. Ta pristop je tudi naučen hitreje kot obstoječi pristopi, ki za barvanje uporabljajo celotne slike.

• Language: Slovenian
• Keywords: umetna inteligenca, strojno učenje, globoke nevronske mreže, barvanje črno-belih slik
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2017
• PID: 20.500.12556/RUL-94553
• Publication date in RUL: 04.09.2017

GODEC, Primož, 2017, Barvanje črnobelih slik z globokimi modeli [online]. Master’s thesis. [Accessed 18 April 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=94553

Secondary language

• Language: English
• Title: Deep models for image coloring

Abstract

Since the color photography came into everyday use in the last fifty years our grandparents are still owning many black and white photographs which we would like to colorize. Researchers are therefore encouraged to develop algorithmic approaches for black and white photographs and video colorization. We have developed a set of automatic approaches based on the machine learning and neural networks, which are using regression and classification. We compared them with approaches from related work. Our approaches reach the quality of colorization comparable to those from related works. Our new approach on image parts improves colorization of images which size is different from those from the training set. This approach is also faster in training than existing approaches that uses full images for learing.

• Keywords: umetna inteligenca, strojno učenje, globoke nevronske mreže, barvanje črno-belih slik