Analiza modelov za preurejanje obraznih slik v lepotni industriji

JAKIMOVSKA, MARIJA (author), Štruc, Vitomir (mentor)

Ličenje obraza je postopek, ki spremeni videz obraza s posebnimi izdelki za osebno nego, kot so puder, senčila za oči, šminke itd. V zadnjem času je vse več pozornosti deležna tehnologija prenosa ličil, ki s pomočjo algoritmov računalniškega vida omogoča prenos ličil iz izbrane referenčne slike na želeno vhodno sliko obraza. Takšno tehnologijo je mogoče uporabiti za različne tipe ženskih obrazov, različne scene, različne starosti, različne kože itd. Idealna metoda prenosa ličil mora zagotoviti videz obraza vhodne slike obraza, pri čemer se prenese le slog ličenja referenčne slike, končna izhodna slika pa samodejno predstavlja kombinacijo vhodne slike obraza in izbranega izgleda referenčnega ličila. Ker so mimika obraza ljudi, oblike obraza, oblike ustnic, oblike obrvi, oblike oči ter razdalja med obrvmi in očmi različni, je lahko izhodno ustvarjena slika popačena in vsebuje vidne artefakte. Poleg tega je stil ličenja spremenljiv, nanj pa vplivata tudi starost in rasa. Poleg prenosa ličil sodobni modeli računalniškega vida omogočajo tudi preurejanje drugih lastnosti obraznih slik, kot je barva ali oblika las, odtenka kože, starosti itd. Takšni modeli so prav tako izrednega pomena za lepotno industrijo, saj omogočajo uporabnikom vnaprejšnji vpogled v morebitne spremembe v izgledu zaradi obiska pri frizerju, kozmetičarki ipd. V tem delu predstavimo model prenosa ličil BeautyGAN, s katerim na dano vhodno sliko, ki je brez ličila, prenesemo ličila, ki so privzeta iz določene referenčne slike. Prav tako opišemo model StarGAN, namenjen preurejanju izgleda obraznih slik. Delovanje izbranih modelov v zaključnem delu analiziramo v eksperimentalni evalvaciji na javno dostopnih podatkovnih zbirkah. Pokažemo, da je z modelom BeautyGAN mogoče izvesti vizualno preprečljiv prenos ličil za širok spekter vhodnih slik z različnimi lastnostmi ter da model StarGAN zagotavlja smiselno preurejanje obraznih slik, a z nekoliko nižjo kakovostjo rezultatov, kot jih generira BeautyGAN.

• Language: Slovenian
• Keywords: prenos ličila na obrazu, referenčna slika, izhodna generirana slika
• Work type: Bachelor thesis/paper
• Organization: FE - Faculty of Electrical Engineering
• Year: 2022
• PID: 20.500.12556/RUL-140231
• COBISS.SI-ID: 121434115
• Publication date in RUL: 13.09.2022

Secondary language

• Language: English
• Title: Analysis of face image editing models for the beauty industry

Abstract

Facial make-up is a procedure that transforms the appearance of the face with specific personal care products such as powder, eye shadow, lipstick, etc. Recently, make-up transfer technology, which uses computer vision algorithms to transfer make-up from a selected reference image to a desired input image of the face, has received more and more attention. Such technology can be applied to different types of female faces, different scenes, different ages, different skins etc. An ideal make-up transfer method should provide the facial appearance of the input facial image, where only the make-up style of the reference image is transferred, and the final output image automatically represents a combination of the input facial image and the selected appearance of the reference make-up. Since human facial expressions, face shapes, lip shapes, eyebrow shapes, eye shapes and the distance between the eyebrows and the eyes are different, the output generated image may be distorted and contain visible artefacts. In addition, make-up style is variable and influenced by age and race. In addition to make-up transfer, modern computer vision models also allow other features of facial images to be rearranged, such as hair colour or shape, skin tone, age etc. Such models are also of great importance for the beauty industry, as they allow users to see in advance possible changes in appearance due to a visit to a hairdresser, beautician etc. In this thesis, we present a makeup transfer model BeautyGAN, which transfers makeup defaulted from a given reference image to a given input image that is non-makeup. We also describe a model StarGAN, designed to rearrange the appearance of facial images. In this thesis, we analyse the performance of the selected models in an experimental evaluation on publicly available datasets. We show that the BeautyGAN model can be used to perform visually avoidable makeup transfer for a wide range of input images with different features, and that the StarGAN model provides meaningful reordering of facial images, but with slightly lower quality results than those generated by BeautyGAN.

• Keywords: facial make-up transfer, reference image, output generated image