正如我們所知，在影像中，顏色是一個主要的特徵，且影像中的顏色，將會決定影像傳遞給人的感覺。而攝影師和設計師通常藉由加強作品中的整體顏色，傳遞他們想表達的感覺。所以，更改圖片中的顏色將可以改變圖片傳遞的感覺。 本論文中，提出了一個情感色彩轉換的機制，利用參考圖的顏色分佈，去轉換原圖的整體顏色，使得原圖的顏色主題和參考圖相似，進而改變原圖所傳遞的感覺。轉換過後的結果圖，也可利用本論文中所提出的情感分析來評量，結果圖的情感分類將會和參考圖的分類一致。另外，為了使情感色彩轉換機制適用於所有圖片，本論文提出了一個膚色保存步驟，針對在其他色彩轉換方法下有問題的人像圖片類型作處理，避免在膚色上過度修改造成色彩轉換結果失真，同時，在色彩轉換過程中，著重保護圖片上顏色的和諧度，將原圖和結果圖的透明度結合，減少結果圖片視覺上失真的問題。總言之，本論文提出適用於任何圖片的情感色彩轉換機制，包括一般圖片和包含人像的圖片。同時，論文提出的另一個機制為圖片情感分析。首先，擷取圖片中的主要顏色，再利用事先定義好的情感類別來對圖片做分類。為了確保擷取的主要顏色能代表圖片的整體顏色，進而提升我們情感分析的準確率，在分析過程中，我們使用顯著圖來輔助圖片色彩的擷取。 最後，情感分析和情感色彩轉換的結果將驗證我們提出方法的可行且有效性。情感分析結果與實際圖片傳遞給人的感覺一致，且拿我們提出的情感色彩轉換機制和其他色彩轉換方法相比較，我們的方法可以得到比較好的結果。 As we know, color, one of important features for composing images, can affect people on emotional level. Photographers and designers usually enhance desired color in their works to convey feeling. Editing color theme of images achieves the goal to change the emotion evoked by images. In this thesis, a method to change the emotion in images by editing color content called affective color transfer is proposed, we focus on implementing color transfer in images to make overall color theme of input image be similar with reference image. The result can be evaluated by affective analysis proposed in our work, and the affective class of result image is consistent with reference image. The proposed skin color preserving phase prevents skin color from over-modification in original color transfer. Furthermore, we attach importance about the harmony of output image, combing both the opacity of input and output image in color transfer to reduce visual distortion. The proposed method performs well in not only general images but also images which exit human. Another mechanism is affective analysis in images. First, we define affective classes and then extract affective colors in the image to classify the affective class. By using saliency map, we extracted the affective color in the image exactly, and result of affective analysis is improved. Finally, experiment results of affective analysis and affective color transfer have confirmed the effectiveness of proposed method. Affective classes predicted by affective analysis accord with the emotion evoked by images. We compare the algorithm of color transfer between proposed method and other methods, where our method performs better.
