| 講演抄録／キーワード |
| 講演名 |
2007-01-26 16:50
Automatic Extraction of Facial Feature points from Orthogonal Color Images for Surgical Planning
○M. Shamsi（Univ. of Tehran/Univ. of the Ryukyus）・R. A. Zoroofi・Caro Lucas・M. Sadeghi Hasanabadi（Univ. of Tehran）・M. R. Asharif（Univ. of Ryukyus） |
| 抄録 |
（和） |
Recent advances in digital human technology combined with the broad availability of high performance computer graphics have created a unique opportunity to develop a novel set of applications meeting the demands of reconstructive and aesthetic plastic surgery. Detection of the facial feature points and measurements based on them plays an essential role in facial reconstructive surgical planning. The appearance of the skin color depends on the lighting conditions. One problem in extraction of feature points is that the image features can be severely corrupted due to illumination, noise, and occlusion, while shadows can cause numerous strong edges. In this paper we presented a fully automatic algorithm based on Gaussian mixture model for segmentation and extraction of facial color skin. Simultaneously, we correct the image shadows from varying illumination. This algorithm will be able to determine accurate contour of facial color skin from two orthogonal views. For this reason, mathematically the varying illumination effect can be expressed as follows:
where is the slowly varying shadow from varying illumination; and is the image noise. Typically, image classes are modeled by a Gaussian distribution with mean and variance . Certainly, replacing one of classes with uniform density probability and amending the EM algorithm appropriately, gives significantly better results. The first step in applying EM algorithm is initialization of the class parameters. This is done with k-means algorithm and then using of the EM parameter estimation. After facial skin localization, we directly locate eyes, mouth and nose based on their feature maps derived from both the luminance and chrominance of frontal image in YCbCr color space. We consider only the area covered by face mask. Finally, the edge information and feature map information combined together for detection of some feature points from frontal view. The combination of edge information with segmented color skin obtained from profile view detects accurately the external edge of face. Then we compute nose tip as a feature point, partition external contour of face in profile view, fit nth order curve and make derivation for feature points detection. 100 patients are selected from ENT section of the Imam Hospital, Tehran, Iran. We have two orthogonal color images from patients. The presented algorithm demonstrates the better results in extraction of face mask and facial features, and is highly robust against shadow from varying illumination in image. |
| （英） |
Recent advances in digital human technology combined with the broad availability of high performance computer graphics have created a unique opportunity to develop a novel set of applications meeting the demands of reconstructive and aesthetic plastic surgery. Detection of the facial feature points and measurements based on them plays an essential role in facial reconstructive surgical planning. The appearance of the skin color depends on the lighting conditions. One problem in extraction of feature points is that the image features can be severely corrupted due to illumination, noise, and occlusion, while shadows can cause numerous strong edges. In this paper we presented a fully automatic algorithm based on Gaussian mixture model for segmentation and extraction of facial color skin. Simultaneously, we correct the image shadows from varying illumination. This algorithm will be able to determine accurate contour of facial color skin from two orthogonal views. For this reason, mathematically the varying illumination effect can be expressed as follows:
where is the slowly varying shadow from varying illumination; and is the image noise. Typically, image classes are modeled by a Gaussian distribution with mean and variance . Certainly, replacing one of classes with uniform density probability and amending the EM algorithm appropriately, gives significantly better results. The first step in applying EM algorithm is initialization of the class parameters. This is done with k-means algorithm and then using of the EM parameter estimation. After facial skin localization, we directly locate eyes, mouth and nose based on their feature maps derived from both the luminance and chrominance of frontal image in YCbCr color space. We consider only the area covered by face mask. Finally, the edge information and feature map information combined together for detection of some feature points from frontal view. The combination of edge information with segmented color skin obtained from profile view detects accurately the external edge of face. Then we compute nose tip as a feature point, partition external contour of face in profile view, fit nth order curve and make derivation for feature points detection. 100 patients are selected from ENT section of the Imam Hospital, Tehran, Iran. We have two orthogonal color images from patients. The presented algorithm demonstrates the better results in extraction of face mask and facial features, and is highly robust against shadow from varying illumination in image. |
| キーワード |
（和） |
Facial feature extraction / EM algorithm / , Orthogonal images / Color space / / / / |
| （英） |
Facial feature extraction / EM algorithm / , Orthogonal images / Color space / / / / |
| 文献情報 |
信学技報, vol. 106, no. 509, MI2006-127, pp. 165-168, 2007年1月. |
| 資料番号 |
MI2006-127 |
| 発行日 |
2007-01-19 (MI) |
| ISSN |
Print edition: ISSN 0913-5685 |
| PDFダウンロード |
|
|