多尺度區域強化之姿態遷移用於自動人像生成;Multi-Scale Region Reinforcement on Pose Transfer for Automatic Person Image Generation

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/83943

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/83943

Title:	多尺度區域強化之姿態遷移用於自動人像生成;Multi-Scale Region Reinforcement on Pose Transfer for Automatic Person Image Generation
Authors:	陳思頴;Chen, Sih-Ying
Contributors:	資訊工程學系
Keywords:	生成對抗網路;姿態轉換;OpenPose;Generative Adversarial Network;Pose Transfer;OpenPose
Date:	2020-07-20
Issue Date:	2020-09-02 17:43:56 (UTC+8)
Publisher:	國立中央大學
Abstract:	随著人工智慧與深度學習領域的蓬勃發展，已被廣泛應用於不同領域中，不論是在語義分析、影像識別等，都有相當顯著的貢獻。如今人工智慧的目標不在是讓電腦擁有智慧，而是希望讓電腦也具有創造力，如寫詩、作曲、或者是影像生成等，透過人工智慧，無中生有，創造出無限潛能。本篇論文提供一個姿態遷移系統，藉由人物圖像與目標姿態，讓電腦自動生成出符合目標姿態的人物圖像。本論文使用了漸進式的姿態遷移生成模型架構，透過漸近式的方式將人物圖像的姿態轉換至目標姿態。在轉換的過程中，我們提出了多尺度區域提取器 (Multi-Scale Region Extractor)，透過擷取人物影像中特定的區域位置的特徵圖，來改善自動編碼器遺失資料訊息的問題，同時也降低了姿態遷移中斷肢的可能性。並針對於多尺度區域特徵提取器，設計了區域風格損失函數 (Region Style Loss)，來優化訓練生成模型的過程。最後，基於本系統的架構下，只要使用一張人物圖像，便可以針對喜好生成出不同舞蹈風格的影片。;With the vigorous development of artificial intelligence and deep learning, they have been widely used in different fields. Whether in semantic analysis, image recognition, etc., there are quite significant contributions. The goals of artificial intelligence are to make computer creative, such as writing poems, composing, or making images, making out of noting, rather than to have intelligence. This thesis proposes a DanceGAN, which can make computer generate character images that matches the target posture automatically. In this thesis, we use a progressive pose transfer to generate a model architecture, which transforms the pose of the character images to the target pose in an asymptotic manner. In the transform process, we propose the Multi-Scale Region Extractor to capture specific area of the character image to improve the missing data message problems of auto encoder. We also design the Region Style Loss for Multi-Scale Region Extractor to improve the training process of generating model. Finally, based on the architecture of this system, we can generate different dancing style according to your favorite using only one character image.
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	69	View/Open

社群 sharing

Loading...