Analysis of Differences Between Western-Caucasian and East-Asian Basic Expressive Faces for Automatic Facial Expression Recognition

• Material Type: 博士論文
• Title: Analysis of Differences Between Western-Caucasian and East-Asian Basic Expressive Faces for Automatic Facial Expression Recognition
• Author/Editor: Benitez Garcia, Gibran de Jesus
• Author Heading: Benitez Garcia, Gibran de Jesus
• Alternative Title: 顔表情自動認識における西洋人と東洋人の基本的表情の違いに対する分析
• Degree grantor/type: 電気通信大学
• Date Granted: 2017-09-29
• Dissertation Number: 甲第911号
• Degree Type: 博士（工学）
• Text Language Code: eng
• Target Audience: 一般
• Note (General): Facial Expression Recognition (FER) has been one of the main targets of the well-known Human Computer Interaction (HCI) research field. Recent developments on this topic have attained high recognition rates under controlled and “in-the-wild” environments overcoming some of the main problems attached to FER systems, such as illumination changes, individual differences, partial occlusion, and so on. However, to the best of the author’s knowledge, all of those proposals have taken for granted the cultural universality of basic facial expressions of emotion. This hypothesis recently has been questioned and in some degree refuted by certain part of the research community from the psychological viewpoint. In this dissertation, an analysis of the differences between Western-Caucasian (WSN) and East-Asian (ASN) prototypic facial expressions is presented in order to assess the cultural universality from an HCI viewpoint. In addition, a full automated FER system is proposed for this analysis. This system is based on hybrid features of specific facial regions of forehead, eyes-eyebrows, mouth and nose, which are described by Fourier coefficients calculated individually from appearance and geometric features. The proposal takes advantage of the static structure of individual faces to be finally classified by Support Vector Machines. The culture-specific analysis is composed by automatic facial expression recognition and visual analysis of facial expression images from different standard databases divided into two different cultural datasets. Additionally, a human study applied to 40 subjects from both ethnic races is presented as a baseline. Evaluation results aid in identifying culture-specific facial expression differences based on individual and combined facial regions. Finally, two possible solutions for solving these differences are proposed. The first one builds on an early ethnicity detection which is based on the extraction of color, shape and texture representative features from each culture. The second approach independently considers the culture-specific basic expressions for the final classification process. In summary, the main contributions of this dissertation are: 1) Qualitative and quantitative analysis of appearance and geometric feature differences between Western-Caucasian and East-Asian facial expressions. 2) A fully automated FER system based on facial region segmentation and hybrid features. 3) The prior considerations for working with multicultural databases on FER. 4) Two possible solutions for FER with multicultural environments. This dissertation is organized as follows. Chapter 1 introduced the motivation, objectives and contributions of this dissertation. Chapter 2 presented, in detail, the background of FER and reviewed the related works from the psychological viewpoint along with the proposals which work with multicultural databases for FER from HCI. Chapter 3 explained the proposed FER method based on facial region segmentation. The automatic segmentation is focused on four facial regions. This proposal is capable to recognize the six basic expression by using only one part of the face. Therefore, it is useful for dealing with the problem of partial occlusion. Finally a modal value approach is proposed for unifying the different results obtained by facial regions of the same face image. Chapter 4 described the proposed fully automated FER method based on Fourier coefficients of hybrid features. This method takes advantage of information extracted from pixel intensities (appearance features) and facial shapes (geometric features) of three different facial regions. Hence, it also overcomes the problem of partial occlusion. This proposal is based on a combination of Local Fourier Coefficients (LFC) and Facial Fourier Descriptors (FFD) of appearance and geometric information, respectively. In addition, this method takes into account the effect of the static structure of the faces by subtracting the neutral face from the expressive face at the feature extraction level. Chapter 5 introduced the proposed analysis of differences between Western-Caucasian (WSN) and East-Asian (ASN) basic facial expressions, it is composed by FER and visual analysis which are divided by appearance, geometric and hybrid features. The FER analysis is focused on in- and out-group performance as well as multicultural tests. The proposed human study which shows cultural differences in perceiving the basic facial expressions, is also described in this chapter. Finally, the two possible solutions for working with multicultural environments are detailed, which are based on an early ethnicity detection and the consideration of previously found culture-specific expressions, respectively. Chapter 6 drew the conclusion and the future works of this research.; 2017
• Source: fulltext; https://uec.repo.nii.ac.jp/?action=repository_action_common_download&item_id=8563&item_no=1&attribute_id=20&file_no=1
• Access Restrictions: インターネット公開
• Data Provider (Database): 国立情報学研究所 : 学術機関リポジトリデータベース（IRDB）（機関リポジトリ）; https://irdb.nii.ac.jp
• Original Data Provider (Database): 電気通信大学 : 電気通信大学学術機関リポジトリ