基於增量學習之人臉辨識研究

Abstract

人臉在生物驗證中是非常重要的特徵，在過去十幾年來，人臉辨識於電腦視覺的研究上也是非常熱門的議題，人臉辨識的技術也廣泛的運用在各方面，例如使用於監視系統或是安全控管系統上。本論文使用了增量學習的方法，設計了一個於課堂環境中可以自動辨識學生的一個點名系統，由於學生的造型在每一次的上課中會與之前有些微的變化，因此在辨識的同時，將這些新的影像加入原有的訓練資料中訓練，對於後續所訓練出的人臉模型將會越來越好。在論文中，我們使用了二維線性鑑別法(2DLDA)作為人臉訓練及辨識所用的分類器，並且使用了影片辨識上常用的投票，以及課堂所能利用的互斥資訊來提升辨識率；在增量學習上，也提出了一個驗證方式由測試影像中選擇出適當的影像重新訓練，並且進行了許多實驗來評估增量學習使用於人臉辨識上的效能。

Face is one of the most important features in biometric verification. During past decades, face recognition has been a very active research issue in computer vision. It is widely to apply face recognition to many applications such as security control and surveillance system. This thesis employs the incremental learning approach to design a roll-call system that can automatically identify students in a classroom environment. Since student faces may be a little change in different classes, it could be better to involve current recognized face images as our training set. In this thesis, our roll-call system designs a two-dimensional linear discriminant analysis (2DLDA) classifier for face recognition. An exclusive method is adopted to improve the recognition results in multi-face environment. Then, an incremental model is proposed to validate what recognized face images can be involved in the next training face images. We also perform several experiments to demonstrate the performance of our approach.