運用輪廓色彩和五官的人臉偵測技術

Abstract

由於人臉相較於其他身份識別系統如虹膜識別較不會有隱私權的問題，因此自動化人臉識別在物件識別的領域當中是一個非常重要的課題。但是，再做進一步的識別之前，我們首先需要將人臉從一張給定的影像中切割出來以供後端處理之用。而人臉偵測，在現今依舊是一個非常困難的問題。 人臉偵測系統大致分為兩類，以特徵為基礎或是以整體為基礎。在以整體為基礎的偵測方法中，給定的影像並未藉由尋找某種特定的特徵來偵測人臉而是做整體考量。這樣的方法具有較能找出較小人臉與解析度較差人臉的優勢。相反地，以特徵為基礎的方法先找出給定的影像中是否有人臉特徵的存在，再藉由這些人臉特徵做出是否有人臉出現在給定的影像之中的判斷。 在本篇論文中，我們對於靜態彩色影像的人臉偵測問題提出了一個以特徵為基礎的偵測方法。我們的方法先偵測出給定影像中類膚色的像素並基於空間關係做分群。每個互斥的類膚色區域將被檢測其邊界形狀是否呈現橢圓，若是，該區域將被進一步的判斷其內是否包含有合法的人臉特徵三角形。為了達到這樣的目標，我們結合了色彩、模糊邏輯與加伯小波以偵測眼睛與嘴巴的候選點。合法的三角形將被儲存起來並用來做為顯示人臉及其眼睛、嘴巴所在位置。 在我們依據加州理工大學之資料庫子集合所做的實驗結果中，我們的方法對於一張大小為 的彩色影像平均只需 秒即可完成偵測(使用英特爾 Q6600 2.4GHz 處理器)。有95%的人臉被成功找到且同時只有一個非人臉的區域被判斷為人臉。 本篇論文的架構如下，在第二章到第四章，我們將回顧一些物件識別的演算法，在第五章到第十章，將深入討論我們所提出的人臉偵測系統之方法。

Automatic human face recognition is a very important subject in pattern recognition fields. This is due to the fact that face has less privacy problems against other identification systems such as iris recognition. Before doing further processing, we need first to segment the face from a given image and it is still a very difficult problem. Face detection systems can be roughly classified into two approaches, feature-based and holistic-based. In holistic-based method, given image is processed without analyzing it into several smaller features. It has the advantage of finding small faces and faces in poor-quality images. On the other side, feature-based method is first to detect the existence of facial features in the given image and then based on the extracted features to declare whether a face is presented. In this thesis, we will concern the problem of finding a face in a still color image and proposed a feature-based face detection scheme. Our scheme is first to find skin-like pixels and group them based on spatial relation. Each disjoint skin-like region will be examined to see whether the boundary of it is like an ellipse or not. If the answer is yes, the region will further processed to check whether it contains a valid facial feature triangle. For doing this, we combined color, fuzzy logic and Gabor wavelets to detect eye candidates and mouth candidates. Valid triangles will be stored and used for displaying the location of the face with labeled eyes and mouth. In our experimental results based on a subset of Caltech database, our processing time for detecting faces in a color image with size is in average seconds (Intel Q6600 2.4GHz). The detection rate is 95% while only one false positive is found. The thesis is organized as below. In chapter 2-4, we will review some pattern recognition algorithms. In chapter 5-10, we will discuss our proposed face detection system in deep details.