nextuppreviouscontents
Next:Focus of Attention Up:Face Recognition Previous:The Smart Door Project

Finding Faces in Color Images Just Using Hue

M.V. Ignatova, M.D. Levine

The detection, localization and extraction of faces from images is a challenging problem in computer vision. Its applications include criminology, security systems, content-based image retrieval etc. Faces of subjects with different racial characteristics, in arbitrary size, position and orientation, under varying illumination conditions or partially occluded have to be detected and localized. This project is a study of how well human faces can be detected and localized in color images by using color information alone, hue in particular. The typical way of using color for face localization is simple thresholding. In contrast to this, this work explores the color histogram intersection method. Simple thresholding processes color information pixel by pixel, whereas histogram intersection operates on groups of pixels. It thus captures more information. We investigate the possibility to localize faces based only on the use of color, as opposed to the post-processing of the thresholding image, usually required with simple thresholding. A color space is sought that minimizes the variations in facial color due to races and illumination conditions. We take advantage of the CIE XYZ color space because of both its perceptual superiority to the RGB space, and because a normalization of the space takes place prior to computing hue. Normalization of color space with respect to intensity is shown to minimize the variations of facial color. We choose to specify facial color using just hue and we further model it through 1D hue histogram. We compute a generalized facial color model by accumulating facial color histograms across many images. Testing was accomplished on a database of 200 color images downloaded from the Internet. We have observed that the generalized facial color models we compute are largely independent of the racial characteristics of the subjects. To accommodate the specificity of working with accumulation histograms instead of with histograms of just one object, we propose modifications to the color histogram intersection measure. An iterative algorithm scanning the image at multiple scales is proposed. At each scanning position the algorithm computes a histogram intersection measure to evaluate the color similarity between the sub-image at this position and the generalized facial color model. The algorithm produces a list of locations that have high similarity with the generalized facial model. A threshold value on the similarity value is used to separate the locations containing faces. The algorithm is invariant to different facial orientation, partial occlusion, and to translation and rotation about an axis perpendicular to the image plane, due to the use of integral image characteristics such as a histogram. Multi scaling enables the algorithm to detect faces of different size.


nextuppreviouscontents
Next:Focus of Attention Up:Face Recognition Previous:The Smart Door Project
Annual Report

Fri Nov 26 23:00:32 GMT 1999