基於格狀特徵值之推舉式人臉分類器

Abstract

偵測是人眼視覺中的一項重要功能，但對於電腦視覺仍是一大挑戰。在本論文中，我們提出一種用於推舉式人臉偵測器的新的特徵:「格狀特徵」，它有三項特點：一、特徵以格子網狀的形式來表現，藉此減少特徵的數量與特徵之間的重疊性。二、以漸進的方式擴大特徵空間，將簡單的特徵結合成一更具分辨能力的新特徵。三、加入「變化量」做為測量方法，用來獲取使用「總合」測量所不能獲取的資訊。實驗中，我們訓練了一個由五級分類器串接而成的人臉分類器，總共使用了兩百個弱分類器（特徵＋閾），前面四級分類器我們使用對稱的特徵使得偵測更為穩定強健。在兩組測試圖檔中，使用格狀特徵的分辨能力比傳統Harr特徵效果來得更好。

Detection is an important function of human vision. However, it is still a big challenge for computer vision. In this thesis, we propose a new grid feature for boosting-kind face classifiers. The grid features contain three major properties: (1). they use a grid representation to reduce the number and redundancy of features; (2). they adopt a progressive way to combine simple features together to form more complex and discriminative features; and (3). they add the variance measure to discover more information than the sum measurement. We train a face classifier cascaded by 5 layers and use 200 weak learners in total. In the first four layers, we only use symmetric features for the sake of robustness. Based on the experiments over two test patterns, our grid feature performs better than the commonly used traditional Harr-like feature.