漸進式人臉角度轉換;Progressive Face Transfer With Multiple Discriminators For Face Rotation

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/89780

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/89780

Title:	漸進式人臉角度轉換;Progressive Face Transfer With Multiple Discriminators For Face Rotation
Authors:	王心妙;Wang, Hsin-Miao
Contributors:	資訊工程學系
Keywords:	生成對抗網路;多角度人臉生成;Generative Adversarial Network;Multi-view Face Generation
Date:	2022-07-20
Issue Date:	2022-10-04 11:59:27 (UTC+8)
Publisher:	國立中央大學
Abstract:	本篇論文主要探討如何將影像中的人臉隨心所欲的轉至任一角度，人臉角度轉換在電腦視覺領域裡一直是具有挑戰性的任務，過去的研究大多基於Metric Learning來實現，但隨著人工智慧快速的崛起，漸漸有許多學者發現基於深度學習的方法，在人臉角度轉換上有更出色的表現。本論文提出一個漸進式的生成器，並配合多個鑑別器來實現漸進式的將人臉轉換至各種目標角度。在架構上，生成器我們使用了Pose-Attentional Transfer Network實現漸進式的人臉角度轉換，並配合三個Discriminator，其中一個是用來學習旋轉角度的區別，一個是用來提升區分臉部結構的多樣性和獲取局部感知訊息的能力，而最後一個主要的功能是加強人臉重點區域的生成品質。基於本模型的架構下，使用者只要輸入一張原始人臉影像和一張目標人臉影像，即可將原始人臉旋轉至目標角度。最後，根據實驗結果，我們提出的方法在多項指標上，均有較好的表現。 ;The purpose of this thesis is to convert faces to various angles. Face angle transfer has always been a challenging task in computer vision. In the past, most studies were based on metric learning. However, with the rise of artificial intelligence, many scholars have found that methods based on deep learning have better performance in face angle transfer. In this paper, we proposed a progressive generator that cooperated with multiple discriminators to gradually transform faces to various target angles. In terms of architecture, the generator uses the Pose-Attentional Transfer Network for progressive face angle transfer, and cooperates with three discriminators. One of the discriminators is used to learn the difference of rotation angles. One is used to improve the diversity of different facial structures and the ability to obtain local perceptual information. And the last discriminator is to enhance the generation quality of key areas of the face. Based on the framework of this model, the user only needs to input an original face image and a target face image, then the original face can be transferred to the target angle. Finally, according to the experimental results, our method has good performance in various indicators.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	39	View/Open

社群 sharing

Loading...