混和色彩意象主題生成之研究
No Thumbnail Available
Date
2023
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
當今元宇宙以及相關的多媒體產業逐漸興起，相較於傳統的人工配色，自動化的色彩主題生成，將能更加的符合市場需求。本研究設計了一種具備色彩意象可預測性的色彩主題生成方法，用來符合設計中色彩計畫的決定因素，試圖解決人工生成色彩主題資料數量有限性的問題，在多媒體應用中將數個須自動決定顏色的對象物產生個別的配色。本設計運用自編碼器的特性，以Kobayashi所提出的色彩意象尺度中色彩意象及色彩主題作為訓練資料，建構新的色彩主題生成模型，同時利用色彩意象的色彩意象情緒座標(CIEC)作為參數，進行符合意象的色彩主題微調。本研究提出「單一色彩意象主題生成」及「混合色彩意象主題生成」兩種生成方法，在兩種生成方法中又可以各自在分出兩個不同的組成模式。希望此方法能將有限資料發揮更大的效用，甚至是預測未知資料的數據。研究結果顯示生成色彩意象及色彩主題在單一色彩意象及混和色彩意象都有良好的成效，經問卷調查的結果也證實90%的題目在機器學習生成的色彩意象之色彩主題與人類的感知之間是相似的，並且可以應用在生成式藝術畫作中，有助於多媒體領域的設計師進行色彩計畫。
As the current metaverse and related multimedia industries are gradually emerging, automated theme generation is becoming more aligned with market demands compared to traditional manual color matching. This study presents a theme generation method that incorporates the predictability of color image to fulfill the determining factors of color scheme in design, aiming to address the limitation of manually generating a limited amount of theme data. In multimedia applications, this method generates individual color schemes for multiple objects that require automatic color determination.The design employs the characteristics of autoencoders and utilizes the color image and themes proposed by Kobayashi as training data to construct a new theme generation model. Additionally, the color image emotional coordinates (CIEC) are used as parameters to fine-tune the themes according to the desired image. This research proposes two generation methods:"single color image theme generation" and "mixed color image theme generation." Each of these methods can further be subdivided into two different composition modes. The goal of this approach is to maximize the utility of limited data and even predict data from unknown sources. The research results demonstrate that the generated color image and themes perform well in both single color image and mixed color image. The results of a questionnaire survey also confirm that 90% of the respondents perceive the color themes generated by machine learning to be similar to human perception. The generated themes can be applied in generative art paintings, aiding designers in the multimedia field with color scheme.
As the current metaverse and related multimedia industries are gradually emerging, automated theme generation is becoming more aligned with market demands compared to traditional manual color matching. This study presents a theme generation method that incorporates the predictability of color image to fulfill the determining factors of color scheme in design, aiming to address the limitation of manually generating a limited amount of theme data. In multimedia applications, this method generates individual color schemes for multiple objects that require automatic color determination.The design employs the characteristics of autoencoders and utilizes the color image and themes proposed by Kobayashi as training data to construct a new theme generation model. Additionally, the color image emotional coordinates (CIEC) are used as parameters to fine-tune the themes according to the desired image. This research proposes two generation methods:"single color image theme generation" and "mixed color image theme generation." Each of these methods can further be subdivided into two different composition modes. The goal of this approach is to maximize the utility of limited data and even predict data from unknown sources. The research results demonstrate that the generated color image and themes perform well in both single color image and mixed color image. The results of a questionnaire survey also confirm that 90% of the respondents perceive the color themes generated by machine learning to be similar to human perception. The generated themes can be applied in generative art paintings, aiding designers in the multimedia field with color scheme.
Description
Keywords
色彩意象, 生成式AI, 神經網路, Color Image, Generative AI, Neural Networks