| 摘要: | 人臉是人類對於誰是誰的最基本判定方法, 但是電腦要像人類一樣分出誰是誰段目前而言仍然是項挑戰, 如何快速並準確的由連續的Video Stream中找到人臉, 並分析這個人是誰, 這取決於理論的正確選擇. 本文由一些不同的方向討論人臉辨識的理論, 實作, 實驗及結果, 這些方向分別是人臉的特性, 類神經網路, 及分析統計, 每個章節利用許多代表性的發表著作來討論其優缺點, 和其改善方法, 實驗的數據是利用Yale的人臉資料庫, 收集團體為 [C.V.C.], 資料庫包括15個人, 165張彩色影像, 每人十一張影像, 內容包含表情, 眼鏡, 與光線上的變化, 測試結果由Component數與錯誤率做比較. 由於區域特徵與廣與特徵對於人類五官或是整體的辨識率有某種程度影響, 所以文內舉出區域特徵的擷取辦法, 像是Local Feature Analysis與Local PCA, 做為討論, 而因為光線對於大多理論與方法都具有強烈的影響, 所以文內亦舉出由3-D Model面, 理論對光線適應面, 及影像修改面等三種大方向著手討論解決. The human face is the basic way for human beings to recognize who a person is. Currently it is a challenge right now if we want the computer to recognize faces like humans do. How to detect and recognize faces in a video stream, quickly and accurately, depends on what algorithms you select and how you combine them. In this paper we gather and compare algorithms, implementations, experiments, and results for face recognition from different sources. We compare three techniques: the constraint approach, artificial neural network approach, and statistic approach. We use lots of representative papers in each chapter to discuss their advantages, disadvantages, and how to improve them. The database we use is the Yale face database, which was created by the Computer Vision and Control (CVC) center at Yale Unversity. It includes 15 people, 11 images per person, for a total of 165 color images. The images in the database consist of various illumination conditions and facial expressions. The final test results are computed from their component numbers and error rate. Local and global features have some influences for facial feature classification and recognition rate. Therefore we enumerate several local feature extraction algorithms, such as Local PCA and Local Feature Analysis and discuss how they work. Illumination conditions also affect recognition rate for most algorithms. We discuss the problem and try to solve it in three ways. The three ways are: 3-D model based, less sensitive algorithm based, and image modification based methods. |
