使用高維度描述子的快速大量人臉辨識系統

Abstract

本論文的目標在於以大量的人臉影像建立高維度的特徵資料庫，來解決人臉2D影像常見的角度與光影問題。為了能處理大量的高維度資料，本研究使用一種近似最近點的演算法稱為Locality-sensitive hashing，並提出針對時間複雜度的改良版本，以增加檢索的效能。整個人臉辨識系統分成三個步驟，第一步驟運用Active Shape Models進行人臉區域的擷取；第二步驟使用phased-based Zernike Moment做為特徵區域描述；第三步驟則使用改良式Exact Euclidean LSH建立特徵向量的索引結構。建立索引結構時，會將特徵區域分為形狀與內容兩種進行處理，在保有準確度的前提下，加速檢索的過程。

The goal of this thesis is to solve the 2D face images pose and illumination problems by using the large-scale face images to create high dimensional feature database. For this reason that we need to handle the large-scale high dimensional database, we propose an approximation nearest neighbor algorithm calls Locality-sensitive hashing. However, we propose a new method that improves the time complexity and retrieval efficacy. This system is composed of into three steps. In the first steps, Active Shape Models is used to extract the regions of human face. In the second step, the phased-based Zernike Moment is used to describe the regions. In the third step, the new Locality-sensitive hashing is used to establish the index of Zernike Moment feature vectors. When constructing the database, the regions are divided into two parts: shape and context. We used those two parts to improve the speed of retrieval.