An Active Illumination and Appearance (AIA) Model for Face Alignment

Abstract

Illumination conditions have an explicit effect on the performance of face recognition systems. In particular, varying the illumination upon the face imposes such complex effects that the identification often fails to provide a stable performance level. In this paper, we propose an approach integrating face identity and illumination models in order to reach acceptable and stable face recognition rates. For this purpose, Active Appearance Model (AAM) and illumination model of faces are combined in order to obtain an illumination invariant face localization. The proposed method is an integrated Active Illumination and Appearance Model (AIA) which combines identity, illumination and shape components in a single model and allows us to control them, separately. One of the major advantage of the proposed AIA model is that efficient model fitting is achieved, whilst maintaining performance against illumination changes. In addition to model fitting, images illuminated from different directions can easily be synthesized by changing the parameters related to illumination modes. The method provides a practical approach, since only one image with frontal illumination of each person for training, is sufficient. There is no need to build complex models for illumination. As a result, this paper has presented a simple and efficient method for face modeling and face alignment in order to increase the performance of face localization by means of the proposed illumination invariant AIA method for face alignment, such as the Active Appearance Models, invariant to changes in illumination. From the experimental results, we showed that the proposed AIA model provides higher accuracy than classical Active Appearance Model for face alignment in a point-to-point error sense.

Elastic appearance models

“…the wavelet and wedgelet model by Larsen et al [13]. Other research groups have increased the robustness of AAM by combining it with the active shape model [23] or by making it light invariant [12]. Suggestions to improve the accuracy of the AAM include the maximum a posteriori formulation by Cootes et al [5] and inclusion of local image structure by Scott et al [21].…”

Contourlet appearance model for facial age estimation

“…In the second approach, due to local textures around feature points being more important than others at the model fitting stage, Cristinacce and Cootes [7] suggested replacing the global textures by local ones to avoid dealing with high dimensional texture. Kahraman et al [17] used two in dependent subspaces, illumination and identity, to combat variations in illumination. Christoudias and Darrell [2] pre sented a nonlinear modeling approach combining a Gaus sian Mixture Model (GMM) and a Nearest Neighbor Model (NNM) to model objects in nonlinear subspaces.…”

Face Image Synthesis and Interpretation Using 3D Illumination-Based AAM Models

“…We anticipate that this approach can be extended to face recognition under difficult conditions of lighting and can be generalized to the analysis and recovery of other types of sources of appearance variation such as age, gender, expression, etc., where lighting interferes seriously in the analysis process. Particularly, face interpretation has been faced through two paradigms: 3DMMs Blanz et al (1999;; Romdhani et al (2005; and AAMs Cootes et al (1998;; Dornaika et al (2003); Edwards et al (1998); Kahraman et al (2007); Legallou et al (2006); ; Sattar et al (2007); Xiao et al (2004). 3DMMs cover a wide range of information recovery but are slow and cannot model properly every type of lighting.…”

“…On the other hand, AAMs are fast but cannot model lighting and 3D information simultaneously. AAM models have been used for fast 2D face alignment under variable conditions of lighting Huang et al (2004); Kahraman et al (2007) 3D shape, albedo and illumination under non-uniform lighting conditions, which is still a challenging problem. In contrast, some authors Dornaika et al (2003); Sattar et al (2007); Xiao et al (2004) have proposed 3DA A M s for estimating 3D pose and shape but do not include illumination.…”