- 學位論文
適用於不同光線之下的人臉辨識系統
Face Recognition System under Varying Lighting Conditions
摘要
隨著科技的進步，自動身分確認已成為一個愈來愈重要的問題。如何建構一套既安全又便利的身分確認系統是目前學界與業界都很熱衷的研究課題,本篇論文旨在設計一個可適用在不同光線環境之下的人臉辨識系統,我們主要架構分做兩個部分,一個是光線的處理,人臉辨識有幾個較大的問題例如光線變化. 姿勢變化…. 姿勢變化的影響可以藉由規範使用者而減輕,但是如何去固定光線仍然是個問題,所以我們在這裡希望能夠對光源部份做特別的處理,一個是入侵者的偵測.在大部份的人臉辨識系統的研究裡較少去討論如何去針對入侵者做偵測,但在一個完整的人臉辨識系統裡這一部份是不可獲缺的. 在光線的處理部份,我們利用Quotient image演算法得到一個人在不同光線之下的不變部分,並結合多重分類器跟單一分類器,利用多重分類器的結果達到辨識的目的,在考慮可能有入侵者的情況下,利用單一分類器的特性去作到入侵者的偵測。然後依照訓練資料分別得到兩個不同分類器的信任值,並作多方的搭配,找出最佳的搭配方式 (例如:分別給予不同的権重),並將権重的找尋規劃為一個新的分類問題,並利用這樣的組合增加系統的強建性。
並列摘要
The automatic person authentication becomes more and more important as technology advances. How to build a safe and convenient identity recognition system is a hot research topic in academia and business. In this thesis, we introduce a face recognition system which is suitable under varying light conditions. The main framework can be divided into two parts: the lighting process, the intruder detection. Face recognition based on computer vision suffers from three problems: variations due to changes in pose, viewpoint, and illumination. In the practical face recognition system, we can control the viewpoint, and force the user pose manually. But it is not easy to control the lighting condition as the pose and the viewpoint. In the intruder detection, it is less discusses how to detect the intruders in face recognition, but it is essential to a complete face recognition system. In the lighting process part, we use the quotient image algorithm to obtain the illumination part of the people. We use multi-class classifier to predict the result of the new people, and in consider of the intruder we use one-class classifier to detect intruder. Then we can obtain the confident values of the multi-class classifier and one-class classifier by training data. , and we use this two confident values to find the best match (i.e. We can combine the two confident values by the different weighting). Therefore we design a problem which is to find the weighting combination to be a classification problem. We can use this combination to enhance the robustness of our system.