請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/71416
標題: | 利用使用者引導之最佳化的二維內容設計 2D Visual Content Design Driven by Human-Guided Optimization |
作者: | I-Chao Shen 沈奕超 |
指導教授: | 陳炳宇(Bing-Yu Chen) |
關鍵字: | 電腦圖學,向量圖,數值最佳化,機器學習, computer graphics,vector graphics,machine learning,numerical optimization,human-in-the-loop, |
出版年 : | 2020 |
學位: | 博士 |
摘要: | 一大部分人類是影像的狂熱消費者。儘管如此,大部分的人只能使用並且觀看視覺資料,只有少部分的人有足夠多的專業和天份能夠有效率的利用影響資料來表示他們自己。即使是最普遍的二維視覺資料如影像和影片,大部分的人們都沒辦法有效率地從頭產生他們,或是改變這些資料來增加他們的美感。比如說,專業的美工人員可以有效率地利用向量圖軟體來產生二維標誌圖片。相對的,一般使用者經常需要花費很長時間但還是沒辦法產生有美感的圖片設計。
在這篇論文研究中,我們調查並探索了幾種資料驅動的方法來弭平這個不對等的分佈。我們主要透過結合人類的先備知識以及嶄新的最佳化演算法來達到這個目的。首先,我們探索了如何讓使用者可以直接去探索和尋找利用生成影像模型 (generative image modeling)達到他想要的圖片。我們的方法提供多個滑桿 (slider) 讓使用者更有效率去瀏覽可能生成的圖片,並且允許使用者透過影像編輯工具來指定想要的影像特徵。接著,我們探索了如何產生符合人類視覺期望的半結構化 (semi-structured) 美工圖片向量化 (vectorization) 演算法。這些半結構化的美工圖片往往具備了區塊顏色區別性很強,部分連續邊界的特性。我們利用以前對人類視覺對於形狀的反應的研究來產生符合人的視覺系統會預期的結果。同時,我們也探索了如何利用單一物品形態的標籤來自動產生這些二維的向量美工圖。最後,我們提出了一個演算法和系統來幫助使用者設計多視角的向量美工圖案。 在這些研究的過中,我們透過線上群眾外包平台的方式,來利用人類感知的比較作為衡量的標準。從結果中可以看到,我們提出的方法都能夠準確的捕捉人類的先備知識和喜好;也因次,我們的方法產生的結果設計都能夠獲得較多使用者的喜愛。未來,我們預想我們提出的這些方法和經驗,可以提供一個重要的基礎給之後嘗試要設計計算輔助系統的研究。 Humans consume visual content avidly for a very long time. The magnitude of the consumption grows exponentially in the past few years due to widespread online social networks and content sharing services, such as Facebook, Instagram, and Youtube. However, there is a huge asymmetry–while everybody avidly consumes visual data, only a few are talented enough to effectively express themselves visually. Even for the most common visual content such as 2D images and videos, most of us still cannot efficiently design them from scratch or manipulate them to enhance their aesthetics. For example, professional artists can generate a 2D icon quite efficiently using a vector graphics authoring tool. On the contrary, naïveusers often spend long hours but still fail to generate an aesthetic design. In this dissertation, we investigate several data-driven approaches for eliminating this asymmetry by combining human priors (including their preferences and knowledge) with novel optimization methods. First, we investigate how to enable the users to control the image generation process using a deep generative model. Second, we investigate methods for generating 2D clipart from existing low-resolution raster icon images and single category labels. Third, we investigate a method on how to generate2D clipart from unseen viewpoints given only a single viewpoint. Specifically, we propose the following three human-guided optimization methods to facilitate efficient 2D visual content design. 1.First, we present a human-in-the-optimization method that allows users to directly explore and search the latent vector space of generative image modeling. Our system provides multiple candidates, and the user selects the best blending result using multiple sliders and image editing tools. 2.Second, we propose approaches (i) to convert artist-drawn images stored as raster images to their vector image form and (ii) to generate the 2D vector clipart directly from a single category label. We first leverage previous studies about human perception of shapes to generate vector images consistent with viewer expectations. Furthermore, we design a generative model to synthesize clipart directly from a single category label. And we trained this generative model on a new clipart dataset of man-made objects called ClipNet. 3.Third, we design an assistive system for clipart design by providing visual scaffolds from the unseen viewpoints. We combined user-provided structure information and automatically predicted 3D structures into a novel curve extrusion optimization method. We evaluated these methods using perceptual comparisons through online crowdsourcing. The results showed that our proposed methods were able to accurately capture various aspects of human prior and provide meaningful supports for various design activities; thus, the user using our methods are able to obtain better visual content than other methods. We envision that these methods and the experiences we learned in this study will provide a good foundation for future research on computational assistive design system to generate more complicated visual content. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/71416 |
DOI: | 10.6342/NTU202004361 |
全文授權: | 有償授權 |
顯示於系所單位: | 資訊網路與多媒體研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-2611202009262100.pdf 目前未授權公開取用 | 65.98 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。