基於跨領域映射達成非監督式內容保留風格導向的圖片轉換

Abstract

在不同領域間的非監督式圖片轉換在電腦視覺中是個很重要的問題。這問題最大的挑戰是要保留原圖的內容(content preserving)同時又要能做到風格轉換(style translation)。我們定義了這兩個問題的矛盾點(trade-off)並且定義了我們所要保留的內容(content)是什麼。現有的方法都會受到這個矛盾點(trade-off)的限制。舉例來說，在貓轉狗的問題上，主要的問題是貓跟狗的領域間差異很大，所以現有的方法就會無法生出一個合理的結果。我們提出了一個模型去保留屬於各自領域的內容(domain-specific content feature)來解決這個困難的問題，並且證明了我們提出的方法比起其他現有表現最佳的方法有更優異的表現。

Unsupervised Image-to-Image across domains is an important problem in the field of computer vision. The most challenging task in this works is to preserve content information while performing style translation. We define what the content information we need to preserve in the cross domain translation and explore the trade-off between content preservation and style transfer. Current methods suffer from this trade-off and fail to strike a balance between content and style. For instance, the cat-to-dog translation task, the main problem is that there is a large gap between two domains such that current methods failed to generate a reasonable result. We propose a content-preserving model which learn the domain-specific content feature to control this trade-off and demonstrate that the proposed method achieves superior performance to the state-of-the-art methods.