一般觀眾對於「類神經網路」之藝術風格轉移認知模式先期研究

A Pilot Study on Audience's Cognitive Model of Neural Style Transfer

近年隨著人工智慧所引領的類神經網路深度學習技術日趨成熟，而其相關應用已經逐漸蔓延至包括藝術在內的各個領域，對藝術作品的創作、體驗、審美和欣賞將帶來新的機遇和挑戰。目前該領域多注重演算法的精進，對於演算法的精進主要牽涉到兩個問題：1.就科技（理性）而言，觀眾能否辨認經由電腦的藝術風格轉移？2.就藝術（感性）而言，哪些主要因素會影響經由電腦的藝術風格轉移？因此，本研究透過認知人因工程研究，分析人們如何感知創作者的編碼過程（色彩、筆觸、紋理）與閱聽者的解碼過程（技術、語意、效果）之影響，以期能對於構建人工智慧應用於藝術創作之研究具有積極的助益。本研究係屬一系列的相關研究，先期研究招募了31位藝術、美學與設計等專家學者參與，對野獸派畫作與對應轉換圖的契合度和認知效果進行評估。結果發現，不同繪畫的風格特性是可以被認知與分辨，風格轉移的整體效果會影響觀眾的喜好度；最後，提出一個探討藝術風格轉移的研究模式。除了有助於藝術風格轉移演算法的改進外，也為人工智慧應用於藝術創作相關研究提供參考。

In recent years, the use of AI is becoming more mature with the development of neural network technology, and access to various fields including art creation. Meanwhile, it brings new opportunities and challenges in the field of invention, experience, aesthetics and appreciation of art. At present, this field only focuses on the optimization of algorithms, but there are two key points in the improvement of algorithms. (1) As far as technology (rationality) is concerned, can the audience recognize the artistic style transfer through computers? (2) In terms of Art (sensibility), what are the main factors that affect the artistic style transfer through computers? Therefore, the purpose of this paper is to analyze the difference of audience’s cognition during the creator's encoding process (color, stroke, texture) and the audience's decoding process (technical level, semantic level, effectiveness level). This study is a series of related studies, in which 31 experts and scholars with a background in art, aesthetics and/or design, were recruited to participate in the previous study to evaluate the degree of fitness and the effect of cognitive between fauvism portraits and corresponding converted images. The results showed that the stylistic characteristics of different images can be recognized and distinguished by subjects. The overall effect of the style transfer will affect the audience’s preferences. At last, a research model of artistic style transfer is put forward. The results of the research not only provide suggestions for the optimization of the NST algorithm but also provide some references for the research on the application of AI into art creation.