人工智慧衣著配色生成與形象感知之研究

Abstract

在這個社群媒體平台發達的時代，虛擬角色已成為虛擬平台的重要識別，其中個人化設計更是不可或缺。色彩為角色外觀最直接的心理感受，而人們對於角色外觀或形象上多以一般口語化的語句描述，同樣地設計師在挑選色票時，也會先確定產品預期要給予的情緒感受，而後才來搭配適合的色彩，可見詞彙對應色彩是未來設計領域的趨勢。本研究將使用自然語言處理技術，搭配所收集的49個形象形容詞與描述句以及小林重順的色彩意象尺度上174個情感詞彙的描述句，開發形象分類器以及色彩意象分類器，使得輸入的描述句能自動分類到適當的類別，接著使用混和色彩意象主題生成方法生成特定的色彩主題，最終能生成出與各形象適配的色彩主題，並評估此方法與形象感知是否一致。研究結果顯示，形象分類器的分類結果有良好的成效，而色彩意象分類器的分類結果大多也都可以分類到適當的類別，只有少部分的結果不適合。問卷結果顯示在16個形象中有8個形象和虛擬角色的衣著配色之間的感知相符，且具設計背景的受測者較能準確感知到形象與衣著配色間的關係。由此可見，本研究所建立的分類器與色彩主題生成方法有一定的可行性，有助於幫助設計師建立情感與詞彙間的關係，進而挑選到適合的顏色。In this era of flourishing social media platforms, virtual characters have become crucial identities on virtual platforms, with personalized design being indispensable. Color, being the most direct psychological perception of a character's appearance, is often described in everyday language when discussing the visual aspects of a character. Similarly, when designers choose a color palette, they first determine the emotional impression the product is expected to convey before selecting suitable colors. It is evident that the correlation between vocabulary and color is a trend in the future of design.This study will utilize natural language processing techniques along with 49 image adjectives and descriptive sentences collected,as well as 174 emotional terms from Kobayashi Shigenobu's color image scale. The goal is to develop image and color image classifiers that automatically categorize input sentences into appropriate categories. Subsequently, a hybrid color image theme generation method will be employed to generate specific color themes, ensuring compatibility with various images. The final step involves evaluating the consistency of this approach with image perception. The research results indicate that the image classifier performs well in classification, and the classification results of the color image classifier mostly fall into appropriate categories, with only a small portion of results being less suitable. Survey results show that out of 16 images, 8 images exhibit perceived congruence between clothing color schemes and virtual character representations. Moreover, participants with a design background demonstrate a more accurate perception of the relationship between images and clothing color schemes. This suggests that the classifier and color theme generation method established in this study have a certain feasibility, assisting designers in establishing connections between emotions and vocabulary, thereby aiding in the selection of suitable colors.