請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/69688
標題: | 運用耳鼻等五官特徵擷取之多角度人臉偵測技術 Advanced Face Detection Algorithm for Arbitrary Rotation, Head-up, and Head-down Cases Using Prominent Facial Features and Hybrid Learning Techniques |
作者: | Chien-Yu Chen 陳建宇 |
指導教授: | 丁建均 |
關鍵字: | 多角度人臉偵測,Adaboost演算法,五官特徵擷取,邊緣偵測,型態學影像處理,深度學習, multi-view face detection,Adaboost machine learning algorithm,facial features extraction,edge detection algorithm,morphological operations,deep learning algorithm., |
出版年 : | 2018 |
學位: | 碩士 |
摘要: | 人臉偵測在計算機視覺領域中是很常被研究的主題,且其在多種的臉部分析演算法上也扮演很重要的角色,包括像是臉部辨識、人臉之年齡估測及表情辨識等等之應用,人臉偵測可以被稱為這些演算法的基石。人臉偵測的最終目標,即為給定一張任意的圖像,並且去檢測此張圖像是否有人臉的存在,倘若存在著人臉,即回傳人臉在此張圖像中的位置及範圍。
近年來,有越來越多人著重於解決比較特殊且難以被偵測的人臉例子,包括像是姿勢、照明情況及含有遮蔽物的情況等等,而此篇論文主要針對多角度的人臉偵測的例子提供解決的辦法。一直以來,總有許多人嘗試修改知名的‟Viola & Jones(Adaboost)”此人臉偵測的方法,此方法在人臉偵測研究的歷史中可被視為一個里程碑,然而,其卻只在正臉的偵測上有顯著的效果,在面對側臉或是其他多角度人臉的例子,其效果相對於近期的演算法較為略遜一籌,因此,在此篇論文中,我們針對彩色人臉的影像提出了一套偵測的方法,基於Adaboost正臉偵測之演算法的架構,加入五官特徵擷取的條件,去解決多角度的人臉偵測之問題。 在此篇論文中,首先,我們先透過膚色濾波器(Skin-filter)及Viola-Jones正臉偵測器找出正臉的例子,再針對無法成功找到人臉的例子進行五官特徵擷取,包含嘴巴、鼻子及耳朵偵測。我們藉由顏色、輪廓及邊緣的資訊去找到嘴巴及鼻子的候選區域,再者,透過近來熱門的深度學習的方式 ( Faster R-CNN ) 來抽取耳朵的特徵並且找到耳朵的候選區域,最後藉由三者特徵的相對位置去分別判定是否為鼻子、嘴巴及耳朵的正確所在位置,再進一步偵測側臉。除此之外,針對仰頭的例子,我們使用邊緣偵測以及型態學影像處理相關的演算法,再加上一些色彩的資訊,去找出眼睛及鼻孔的候選區域,透過計算眼睛及鼻孔的中心點位置,去找出眼睛及鼻孔的正確所在位置,再進一步偵測人臉。最後透過非極大抑制演算法(Non-Maximum Suppression)去改善偵測率。經由這些處理,此篇論文的方法可以解決許多透過Adaboost的演算法無法偵測的例子,並且大幅提升辨識率,更可以成功解決許多多角度人臉偵測的問題。 Face detection is one of the most research topics in the computer vision field and it also plays an important role on many applications of the face analysis algorithms, such as face recognition, age identification, facial expression recognition, and so on. It is the foundation of many applications. The final goal of face detection is given an arbitrary image, and to detect whether the face exists in this image or not. If the image contains the face, the position and range of the face in the image will be returned. Many people contribute to solving particular cases which cannot be detected easily because of pose variation, illumination variation, and occlusion. Therefore, we will provide the solutions on multi-view face detection in this thesis. In recent years, many researchers have intended to modify the well-known Viola and Jones (Adaboost) face detection algorithm. This Viola-Jones detector can be regarded as a milestone in the history of face detection. Nevertheless, its sufficient effectiveness is confined to frontal face detection. It is unable to get better detection rates on multi-view face detection. Hence, in this thesis, we propose a robust face detection algorithm based on the “Adaboost” machine learning algorithm and novel methods of facial features extraction to solve the multi-view face detection problems. First, we apply a skin-filter and Viola-Jones detector to conduct frontal face detection. Second, we extract the facial features of other face images which cannot be found the locations of the faces by first step through our proposed methods, e.g., mouth detection, nose detection, and ear detection. We make use of information of color, edge and contour to extract the facial features such as mouth and nose. Then, we propose a novel method based on the popular deep learning algorithm by improving the techniques of the Faster R-CNN to conduct ear detection. By these proposed methods of facial feature detection, we can obtain the locations of these prominent facial features and proceed to detect the correct locations of the profile faces which contain head-up and head-down cases. In addition, for some cases with head-raised, we apply the edge detection algorithm, morphological operations, and color information to detect eye and nostril candidates. By calculating the locations of the center points of eye and nostril candidates, we can obtain the correct locations of eye and nose to detect the face. Finally, we use the non-maximum suppression algorithm to improve the detection rates. We perform the proposed system on some popular multi-view face databases (e.g., FEI database, CVL database, Pointing’04 database, and so on). Our proposed methods can attain higher detection rates in this novel system, and the effectiveness will be demonstrated in this thesis. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/69688 |
DOI: | 10.6342/NTU201800922 |
全文授權: | 有償授權 |
顯示於系所單位: | 電信工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-107-1.pdf 目前未授權公開取用 | 8.16 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。