視訊人物分群方法之分析與改進

Abstract

人臉分群是近年研究廣泛的課題，近三十幾年出現了許多此領域的論文與相關研究。用於保安方面的人臉辨識系統，以及監視器偵測人臉，也是這方面的主要應用。目前還沒有一個完美的方法被提出來，所以未來此領域仍是多媒體工程的發展主流。

近年許多的研究，偏重於一段影片而非一張張影像上，把影片中連續出現的人臉影像作成串列，有助於人臉分群。本文以人臉影像串列之間的相似度做為主要研究方向，對多種人臉分群方法進行實驗與分析，並提出各種情況適用哪一種分群方法與參數，以及改進這些分群方法的策略。

Face clustering has been widely studied in recent years. In the past thirty years, there have been developed many papers and applications about this subject. Face recognition for security system and face detection for monitors are primary applications of this subject. For now, there does not have a perfect method, so this topic is a hot development direction for multimedia engineering in the future.

Recent research focuses more on a video instead of images. We can extract facial images which appear sequentially in a video, and convert these images into image sequences. By this way, we can improve face clustering techniques. We focus on face image sequence similarity as the main research topic. We experiment on many face clustering algorithms and analyze experimental results. We illustrate how each method works in any cases, and some ways to improve these algorithms.