以人臉為依據建立視訊影片中人物出現時間之索引

Abstract

本篇論文中，我們提出一套視訊中的人物分群流程，從人臉偵測、演員串列的建立、串列間關係的描述、分群方法的設計、到最後的串列擴張，都有清楚地介紹及討論，並針對投影基底、影像前處理、分群方法、描述資訊的選擇等多項議題進行討論與分析，當中又以身體資訊的使用為主要討論對象。 由於演員的臉部姿勢會跟著劇情變化或鏡頭位置而有所不同，造成臉部辨識的困難度，因此在進行演員串列間相似度╱相異度之描述時，除了臉部資訊的使用外，我們加入演員的身體（衣服）資訊，希望藉由兩資訊的結合，達到提升描述力之目的。然而在使用身體資訊時，可能會有演員更換服裝之情形發生，因此我們提出“根據串列間的時間差距決定身體資訊權重”進行臉部資訊與身體資訊的結合，藉由可變權重的使用，降低不同服裝所可能產生的錯誤描述。

In this thesis we propose a procedure for human face clustering from video. The procedure consists of human face detection, the construction of actor-sequences, the evaluation of similarity between actor-sequences, face clustering, and actor-sequence extension. The following issues are analyzed and discussed in this paper: the selection of basis vectors for subspace projection, the preprocessing of face images, various clustering algorithms, and the selection of descriptive information used in the clustering processing. Among them we have emphasized on the use of body appearance in the process. The face pose of an actor/actress changes a lot within a video, and this causes much difficulty for face recognition. To overcome this problem and improve the description of individual actors, while computing the similarity/dissimilarity between actor sequences, we utilize the information of body/clothing appearance in addition to the face images. However, as an actor may change his/her clothes within the video, we propose a weighted summation method that adjusts the relative weighting of face and body appearances according to the time difference between actor-sequences. We find that this can reduce errors caused by changed clothing in the video.