From Canonical Face to Synthesis - An Illumination Invariant Face Recognition Approach

The need to further develop robust face recognition techniques to meet real world requirements is still an open research challenge. It is widely understood that the two main contributions of poor recognition performances are that caused by variations in face pose and lighting. We will deal with the problem of illumination in this chapter. Approaches addressing the illumination-related problems can be broadly classified into two categories; feature-based approach and exemplaror appearancebased approach. Feature-based approaches aim to define a feature space that exhibits some broad invariance over the lighting variations. Examples of these are (Adini & Ullman, 1997), (Belhumeur et al., 1997) and (Yang et al., 2004) which uses different image representations like 2D Gabor-like filters, first and second derivatives of the image, and the logarithmic transformation. Although these features may exhibit intensity immunity, none of these are found to be reliable to cope with significantly large variations in illumination changes (Manjunath et al.1992) (Yang et al., 2004). Exemplaror appearancebased approaches use a set of sample images taken of a class object (in this case a face) as a basis to compute an intermediary image. The intermediate image can then be used either directly as the probe image or be used to synthesize novel views of the face under different lighting conditions (Mariani, 2002). For example, (RiklinRaviv & Shashua, 2001) reported a method to compute the Quotient Image from a small sample of bootstrap images representing a minimum of two class objects. The illumination invariant signature of the Quotient Image enables an analytic generation of the novel image space with varying illumination. However, this technique is highly dependent on the types of bootstrap images used which has the undesirable effect of generating diversely looking Quotient Images even from the same person. (Sim & Kanade, 2001) used a statistical shapefrom-shading model to estimate the 3D face shape from a single image. The 3D recovery model is based on the symmetric shape-from-shading algorithm proposed by (Zhao & Chellappa, 1999). They used the 3D face model to synthesize novel faces under new illumination conditions using computer graphics algorithms. The approach produce high recognition rate on the illumination subset of the CMU PIE database (Sim et al., 2003). However, it was not evident how their synthesis technique can cope with extreme illumination conditions (Sim & Kanade, 2001). (Debevec et al., 2000) presented a method to