請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/83883
標題: | 描繪腦力激盪草圖之深度學習模型 A Deep Learning Framework of Sketching Brainstorms |
作者: | Ming-Min Hsu 許明敏 |
指導教授: | 李瑞庭(Anthony J. T. Lee) 李瑞庭(Anthony J. T. Lee | jtlee@ntu.edu.tw | 0000-0003-0320-7309), |
關鍵字: | 腦力激盪,產品設計與開發,佈局生成,深度學習,關係圖卷積網路, brainstorming,product design and development,layout generation,deep learning,relational graph convolutional network, |
出版年 : | 2022 |
學位: | 碩士 |
摘要: | 為了因應快速變遷的趨勢與增加顧客契合度,企業常使用腦力激盪開發創新點子、尋找解決方案與擬定策略決策,以滿足客戶需求。當進行腦力激盪發想時,企業常利用一序列的手繪草稿圖傳遞想法的概念與流程,將這些手繪草稿圖轉換成佈局線框圖,可以增進彼此的溝通,建立共識,並進一步描繪產品設計和實作的細節。但據我們所知,目前未有研究專注於從手繪草稿圖序列中生成線框佈局圖序列,因此,我們提出了一個研究架構從一系列手繪草稿圖中生成線框佈局圖。我們提出的研究架構包含四個階段:第一階段,我們偵測草稿圖序列中每個草稿圖上的元件,並得出元件之間的空間和方向關係;第二階段,我們運用關係圖卷積網路,並利用元件間空間和方向的關係聚合節點特徵;第三階段,我們設計了一個線框佈局模型,並使用狀態、內部草圖和跨草圖三種注意力機制來生成每個草稿圖的佈局;最後,我們開發了一個渲染模組,將生成的佈局、辨識出的文本內容與各種元件渲染成線框佈局圖序列。實驗結果顯示,我們所提出的研究架構優於比較方法,且生成出來的線框佈局圖不僅可準確傳達空間和方向對齊的資訊,亦可產生接近實際應用範例的佈局。我們的研究架構可幫助企業,在腦力激盪的過程中,將他們的想法從手繪草稿圖轉換為線框佈局圖,並及時且具有成本效益地加快設計流程,進而幫助企業快速反應市場需求,增進其市場競爭力。 To catch fast-moving trends and increase customer engagement, businesses often employ the brainstorming technique to develop innovative ideas, find appropriate solutions to problems, and make strategic decisions for satisfying the needs of customers. The outcome of brainstorming or idea generation may be often presented by a sequence of sketches for showing a concept or process flow. Deriving the layout wireframes from those sketches may bridge the gap in communicating with others during brainstorming sessions and depict the concept or process flow for further design and implementation. To the best of our knowledge, there is no study dedicated to generating the layouts from a sequence of sketches. Therefore, we propose a framework to generate the wireframes from a sequence of sketches. Our proposed framework contains four phases. First, we detect the components in each sketch of the input sequence and derive the spatial and directional relationships between components. Second, we utilize the relational graph convolutional networks to aggregate the node features through the derived spatial and directional relationships. Third, we devise a wireframe layout model to generate the layout of each sketch by using three attention mechanisms namely, state, inner-sketch, and across-sketch. Finally, we develop a rendering module to generate the wireframes for the input sketch sequence. The experimental results show that our proposed framework outperforms the comparing methods, and can generate the layout wireframes not only accurately convey the information of the spatial and directional alignment but are also close to the real applications. The proposed framework may help businesses convert their ideas from sketches to wireframes during brainstorming and speed up their design processes in a timely and cost-efficient manner, which in turn help business catch up with the opportunities in the market. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/83883 |
DOI: | 10.6342/NTU202201496 |
全文授權: | 未授權 |
顯示於系所單位: | 資訊管理學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-1607202210354600.pdf 目前未授權公開取用 | 3.28 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。