Three-Band Modeling Using Prominent Edges for Face Alignment

Fundamental difficulty in face recognition systems is mainly related to successful human face alignment from the input image. In recent years, model based approaches get attention among others. Most powerful method among model-based approaches is known as Active Appearance Model. The method accomplishes this by constructing a relation between shape and texture. Face alignment methods are required to work well even in the presence of illumination and affine transformation. Classical AAM extracts texture and shape information from the training image by using RGB color space. Classical AAM can only handle images having the same or similar color distribution to the images in the training set. Classical AAM cannot align images obtained under different lightning conditions from the training images even if the same person exists in the training database. In this study, we propose to use features which are shown to be less sensitive to illumination changes instead of directly using RGB colors. The proposed AAM is called three-band AAM. The bands are Hue, Hill, and Luminance. Prominent edge detection constitutes the most important part of the model. Experimental studies show that prominent edges do not depend on illumination changes much when compared the original color space, and the three-band AAM based face alignment outperforms the classical AAM in terms of alignment precision.