Please use this identifier to cite or link to this item:
標題: 隨機直線取樣法於人像辨識系統的應用
Face Recognition System using Random Line Sampling Method
作者: 楊佳明
關鍵字: Face recognition;人像辨識;line-based algorithm;Gabor filters;隨機直線取樣;Gabor filters
出版社: 電機工程學系
以非侵入式的技術來抽取出測試者的特徵所完成的即時人像辨識系統已經具有極高的辨識正確率。雖然即時人像辨識系統具有許多的優點,然而在一不受限制的辨識環境下進行人像辨識卻是一件極為困難的工作。許多的人像辨識系統必須對影像設下嚴苛的條件如拍攝環境的光線投射角度、影像格式大小、雜訊、以及臉部表情等才能正常運作。本篇論文提出一套使用直線取樣法結合Gabor filters所完成的即時人像辨識系統以降低系統對上述條件的限制。而此演算法對於測試者頭部的偏轉以及影像格式大小改變等具有一定程度的抵抗力。此外,計算的過程也十分有效率。在抽取特徵向量之前,我們先使用Gabor filters將臉部的重要特徵如眼睛、鼻子、嘴唇等予以強化,使其不易受影像採光所影響,接著再使用隨機直線取樣的方式來抓取臉部影像的特徵。最後,使用最鄰近分類器將測試影像正確分類為資料庫中的某一成員。在我們的實驗中,使用ORL人像資料庫來測試此人像辨識系統,測試的結果顯示其平均辨識正確率為99.6 %,平均辨識一張測試影像所需時間為1.784秒。

Real-time face recognition systems based on inoffensive feature extraction techniques have already produced very high identification rates. Although real-time face recognition system has many advantages, real-time face recognition in an unconstrained environment is a difficult task. Many real-time human face recognition systems operate under strict imaging conditions such as controlled illumination, image size, noises, and limited facial expressions. In this thesis, we propose a line-based face recognition algorithm combined with Gabor filters to alleviate the constraints. This algorithm achieves high recognition rates for rotations both in and out of the plane, is robust to sacling, and is computationally efficient. Before the feature vector extraction, we use Gabor filters to enhance the important facial features such as the eyes, the noise, and the mouth. After that, we use a set of random straight lines to extract the feature vector of the face image. Finally, the nearest-neighbor classifier is used to classify the test face image into one of the persons in the database. In our experiment, we use ORL face database to test this line-based face recognition system. Our method achieved an average recognition rate of 99.6 % using 1.784 seconds per view in average.
Appears in Collections:電機工程學系所

Show full item record

Google ScholarTM


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.