基於多尺度集成增強樹的層級瞳孔中心定位系統

Abstract

在這篇論文中，作者提出了一個新的用於瞳孔中心定位的層級系統。首先， 作者提出了一個用於識別各種頭部姿勢下的眼睛位置的新眼睛定位框架。實驗結果顯示出，和當前主流的眼睛定位方法相比，這篇論文中提出的眼睛定位框架具有更高的穩健性。論文作者接著介紹了改進目前節點訓練方法速度的方法，以增加模型的訓練速度。接著，作者推導了目前的瞳孔定位增強式學習算法的變化版本，並進行了驗證。驗證顯示出變化後的版本在效能上超過了目前最先進的瞳孔定位方法。接著，作者調查了兩個之前提出的改進技巧，並展示了他們在瞳孔定位上的效用。到本論文的最後，一個新的穩健的能夠比目前最先進方法更精確地預測各種情況下的臉部照片中的瞳孔中心位置的瞳孔定位系統被提出了。

In this thesis, the author proposes a new cascaded system for pupil center localization. A new eye detection framework is proposed at first to detect eye locations in large head poses. Experiments demonstrate that the eye detection framework proposed in this thesis shows more robustness against different head poses when compared with the current mainstream method. The author then introduces an improved node training method to increase the training speed of pupil localization models. Next, the author develops new variations to the existing pupil localization boosting algorithms, which are then validated and show superior performance over the current state-of-the-art method. Additionally, the author investigates two previously reported improvement techniques and shows their effects on pupil localization. To the end of this thesis, a new robust pupil localization system, which can estimate the pupil centers from a variety of facial images more accurately than the current state-of-the-art method, is established.