Study on Illumination Insensitive Face Detection Based on Normalization Techniques

• Item type: SIG Technical Reports(1)
• 公開日: 2013-08-26
• タイトル: Study on Illumination Insensitive Face Detection Based on Normalization Techniques
• タイトル / 言語: en
• タイトル: Study on Illumination Insensitive Face Detection Based on Normalization Techniques
• 言語: eng
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: Department of Information Processing, Tokyo Institute of Technology
• 著者所属: Imaging Science and Engineering Laboratory, Tokyo Institute of Technology
• 著者所属(英): en
• 著者所属(英): Department of Information Processing, Tokyo Institute of Technology
• 著者所属(英): en
• 著者所属(英): Imaging Science and Engineering Laboratory, Tokyo Institute of Technology
• 著者名: Min, Yao; Hiroshi, Nagahashi
• 著者名(英): Min, Yao; Hiroshi, Nagahashi
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Face detection has been attracting much academic attention. Illumination problem is one of the most important aspects that decreases its performance. Conventional methods have focused on extracting illumination invariant features or utilizing skin color information. However, the features and colors are unreliable in adverse lighting conditions. In this paper, we investigate various illumination normalization techniques for learning-based face detection under varying illumination. They are four illumination insensitive face representation techniques and five histogram-based normalization methods including our proposed method SH (segmentation-based HHTS). The experimental results show that 1) effective illumination normalization techniques in face recognition are not necessarily useful in face detection; 2) histogram-based methods significantly outperform illumination insensitive face representation techniques in average; 3) SH obtains the best results.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Face detection has been attracting much academic attention. Illumination problem is one of the most important aspects that decreases its performance. Conventional methods have focused on extracting illumination invariant features or utilizing skin color information. However, the features and colors are unreliable in adverse lighting conditions. In this paper, we investigate various illumination normalization techniques for learning-based face detection under varying illumination. They are four illumination insensitive face representation techniques and five histogram-based normalization methods including our proposed method SH (segmentation-based HHTS). The experimental results show that 1) effective illumination normalization techniques in face recognition are not necessarily useful in face detection; 2) histogram-based methods significantly outperform illumination insensitive face representation techniques in average; 3) SH obtains the best results.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AA11131797
• 書誌情報: 研究報告コンピュータビジョンとイメージメディア（CVIM）; 巻 2013-CVIM-188, 号 1, p. 1-6, 発行日 2013-08-26
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会