用於人臉資訊分析的視訊資料集和視訊相似度之分析

Abstract

人臉辨識一直是多年來許多人們研究的課題。早期從心理學方面開始研究人類如何辨別不同的人臉，到現在我們試圖找出一個可靠的方法來讓電腦辨識人臉，一直是一個很大的挑戰，且至今為止也沒有一個完美的方法被提出來。大部分人臉辨識的演算法都是以影像為基礎的，但是在很多情況下，我們卻需要應用在一段影片上而不是單一的影像。 比起單一影像，一段影片能夠提供更多的資訊，有利於提升人臉辨識的可靠性。因此本文主要以人臉影像串列之間的相似度為主要研究方向。本文在四種資料集中搭配一些前處理以及不同環境下比較了幾種人臉影像串列相似度計算方法，說明各方法的優劣和幾種可能會影響效能的因素，並且分析各種資料集的特性。

Face recognition has been studied for many years, but it has stayed a challenging problem as no one perfect method has been proposed. Most face recognition algorithms are image-based. However, in many cases, it is useful and beneficial to apply face recognition algorithms to video data rather than single images. Compared to a single image, a video can provide more information, thus improving the reliability of face recognition. This thesis focuses on facial image sequence similarity as the main research topic. We compare four different datasets under several different environments to analyze algorithm for computing face image sequence similarities. We illustrate the pros and cons of each method and also discuss several factors that may affect the performance. In addition, we also analyze the characteristics of these data sets.