Skip to Main content Skip to Navigation

Segmentation du masque capillaire dans un visage

Abstract : This thesis focuses on segmentation of the hair mask in a face. This report presents the study on the localization and the characterization of hair representation in a picture in a picture. The aim of the thesis is to propose a automatic method to segment the hair with best possible reliability in order to take into account the variability of different hairstyles. The segmentation is performed by a Matting treatment which is a region-contour based approach. This algorithm extract a foreground object from an image. The picture is divided into 2 plans : the first one represents the area of hair information and the second one represents the background plan. This approach is initialized by the definition of some markers which will be diffused into the entire picture. Obtaining a good segmentation depends directly on the precision of these markers placement. We define their position by a combined analysis of three parameters which are characteristics of hair : its texture, its color and its position around the face. In a first time, we set up a frequential analysis to characterize the hair texture. We performed a filtering of the image using a Gaussian band pass isotropic filter. We define a localization mask based on the detection of the frequential zones which are similar to hair. In a second time, we set up a hair color analysis. We define a color classifier which represents the distribution of the color model by a Gaussian distribution on each component of chrominance in YCbCr color space. Common information of hair localization are combined by a method of data fusion based on the transferable belief model. This approach allows to take into account various degrees of ignorance by the modelization of a state of "uncertainty". The addition of this new state is well adapted to our algorithm of segmentation since it makes possible to control the position of the pixels whose state is estimated during the Matting treatment. Then this approach is improved by the addition of a discounting function based on the third hair parameter which is the localization of hair around the face. This function leads to balance the reliability of our sources compared to the distance of the face. Indeed the probability for a pixel to belongs to the hair mask decreases as the pixel is far from the face. The hair segmentation is evaluated thanks to a quantitative analysis by a comparison with a baseline of references hair masks obtained by a semi-manual segmentation. Finally the hair mask is characterized by three descriptors for each hair parameter. This classification allows hair description by an approach similar to a cognitive description made by a human observer.
Document type :
Complete list of metadatas

Cited literature [76 references]  Display  Hide  Download
Contributor : Cedric Rousset <>
Submitted on : Friday, October 29, 2010 - 3:21:15 PM
Last modification on : Thursday, November 19, 2020 - 12:59:39 PM
Long-term archiving on: : Sunday, January 30, 2011 - 2:57:49 AM


  • HAL Id : tel-00530635, version 1



Cedric Rousset. Segmentation du masque capillaire dans un visage. Interface homme-machine [cs.HC]. Institut National Polytechnique de Grenoble - INPG, 2010. Français. ⟨tel-00530635⟩



Record views


Files downloads