隨著科技的進步，電影的製作也日益更新，逐年增加，要如何在這龐大的資料量中幫助使用者快速尋找所欲瀏覽的影片內容，成為一個值得探討的議題。過去針對影片內容分析主要包含物件分類、類型(genre)分類與事件分類，但隨著情意計算(affective computing)的興起，情緒分類也逐漸為人重視。尤其在電影的拍攝手法中，從視覺的色彩、光線明暗等到聽覺的音樂曲調等，往往包含了導演所想表達的情感和場景氣氛，適合作為情緒分類上的輸入特徵。 本研究旨在透過影片的內容式分析，建置一個電影場景的情緒自動分類法。先以人工標記119部電影場景，且不同於傳統單一情緒標記方式，改以多情緒方式加標，試圖達到多情緒分類。之後自所有場景中擷取事先定義的視覺與聽覺特徵值，共50維的特徵向量。利用自我組織特徵映射圖網路演算法作場景分群，再以階層式聚合演算法合併相近群聚，改善群聚過多情況。最後以上述分類方式，實作於影片檢索系統上，在使用者瀏覽影片時，同時回饋內容情緒相似的場景，以達快速觀看同類型的影片。 透過實驗發現，本研究所提出的分類方式，最後所得的分類結果其平均recall、precision超過70%，為不錯的表現。 With the development of technology, digital video collections are growing rapidly in recent years. More and more movies are released around the world and play an important role in our life. How to analyze the huge content to help viewers search a specific type of video effectively becomes one of major issues. In general, earlier video content-based analysis includes object-based classification, genre-based classification and event-based classification. With the growing of affective computing, emotion-based classification is also emphasized because the audiovisual cues in movies are helpful for affective content. The purpose of this study is to construct an affective classification of movie scenes through video content-based analysis. First, a dataset of 119 different scenes from eleven movies were labeled manually and each scene can be described by multiple emotional labels, instead of single label as earlier studies. Fifty audiovisual features were extracted from all scenes for our classifier, self-organizing feature map. Then the hierarchical agglomerative algorithm was employed to merge similar clusters into groups. We implement the classification result to construct a retrieval system such that users can view movie scenes with similar emotion content. The experiments showed that the average recall and average precision achieves 70%. It was turned out our study is an efficient way.
