Perceiving human faces and modeling the distinctive features of human faces that contribute most towards face recognition are some of the challenges faced by computer vision and psychophysics researchers. Human faces belong to a special class of 3-D objects, modeling which involves developing accurate characterizations that account for illumination, head pose variations, facial expressions, etc. Moreover, human faces also undergo growth related changes that are manifested in the form of shape and textural variations. Hence, the robustness to variations due to factors such as illumination, pose, facial expressions, aging, etc., is a significant metric in evaluating face recognition systems. Though psychophysical studies have contributed significantly towards the perception of growing faces and towards understanding craniofacial growth, age progression in human faces has largely been ignored while developing face recognition systems.
Facial aging effects are manifested predominantly in the form of shape variations in faces during formative years (0 - 18 yrs) and in the form of textural variations during adulthood. Factors such as gender, ethnicity, age group etc. are often attributed to playing a significant role in affecting facial growth rates. From a computer vision perspective, developing models for facial aging has significant implications to performing face recognition / face verification across age progression (homeland security), automatic age estimation from face images (parental control, age based Human-Computer Interaction), appearance prediction across age progression (finding missing individuals), etc.
1. A Computational Model for Adult Age Progression (Current work)
During adulthood, facial aging effects are often perceived by means of subtle variations in facial shape and facial texture. It is generally observed that individuals belonging to the same age group, gender and ethnicity exhibit similar traits of facial aging effects. Drooping facial features, forehead wrinkles, nasolabial folds, broken jawline etc. are some of the typical manifestations of aging effects in adult faces. Such characteristic traits of aging effects are often attributed to factors such as the prolonged effects of gravity, the longitudinal pull of facial muscles, facial fat atrophies, gradual loss of skin elasticity etc. Hence, modeling facial aging in adults involves characterizing the typical manifestations of facial aging effects (both in terms of shape and texture) during different ages, in a manner that accounts for the factors that induce aging effects. We propose a facial aging model for adult faces that comprises of a shape variation model and a texture variation model. Facial shape variations are characterized by means of the drifts observed on facial features across ages. The shape variation model implicitly accounts for the physical properties and geometric orientations of different facial muscles and characterizes facial feature drifts across ages as that governed by the changing physical properties of the underlying facial muscles. Facial texture variation model is developed such that facial wrinkles and other skin artifacts are accounted for. From a dataset that comprises of 1200 pairs of age separated face images, we compute facial growth statistics that play a crucial role in developing the shape and texture variations models.
Figure 1 : Facial aging in adult faces
Face recognition across age progression, appearance prediction with increase in age etc. are some of the intended objectives of this initiative.
2. Modeling Age Progression in Young Faces (CVPR 2006)
We propose a craniofacial growth model that draws inspiration from the ‘revised’ cardioidal strain transformation model proposed by Todd et al. [21]. Craniofacial growth observed during formative years, is often attributed to the internal forces (a combination of bio-mechanical stress and graviational forces) that act on the craniofacial complex. Drawing analogies between the remodeling of a fluid-filled spherical object with pressure and the remodeling of the human head with growth, Todd et al. performed a hydrostatic analysis on the effects of internal forces on craniofacial growth. In developing the ‘revised’ cardioidal strain transformation model, they essentially developed a pressure-driven growth transformation model that is applicable to objects that are nearly spherical in shape.
Face anthropometric studies provide measurements extracted across key fiducial features across different ages (1 to 18 years). Such quantitative characterization of facial growth implicitly captures the different rates of growth observed during different stages of growth and hence plays a crucial role in developing our proposed model. Further, factors such as gender, ethnicity etc. are inherently accounted for. Characterizing facial growth by means of age-based facial growth parameters, we study the transformations observed in facial measurements across different ages and propose methods to compute the facial growth parametes. The following figure provides an overview of the proposed approach :
ds that
Figure 2 : Overview of the proposed craniofacial growth model
Figure 3: Appearance prediction across age progression
3. Face Verification across Age Progression (CVPR 2005):
Human faces undergo gradual variations due to aging and since periodically updating large face databases would be a tedious task, a better alternative would be to develop face recognition systems that verify the identity of individuals from pairs of age separated face images. Understanding the role of age progression in affecting the similarity between two face images of an individual is important in such tasks. We wish to address the following problem: How similar is a pair of age separated face images of an individual? How do inherent changes in a human face due to aging, affect facial similarity? Given a pair of age separated face images of an individual, what is the confidence measure associated with verifying his/her identity? Our database comprises of pairs (younger and most recent) of face images retrieved from the passports of 465 individuals. The age span of individuals in our database is 20 years to 70 years.
We develop an age difference classifier designed primarily for the purpose of establishing the identity between a pair of age separated face images and for estimating the age separation between the pair of face images. During adulthood, aging effects in faces are more commonly observed in the form of textural variations such as wrinkles and other skin artifacts. In our formulation the age difference based classification of face images is based on textural variations that are commonly observed in faces due to aging. Across each pair of face images, we compute the difference image by subtracting the more recent image from the older image. The difference image, when computed between age separated images of the same individual (intrapersonal images), captures facial variations due to aging effects. Intuitively, the difference images obtained from the intrapersonal image pairs (image pairs of the same individual) with smaller age separation would be less exaggerated than that obtained from the intrapersonal image pairs with larger age separation. Further, one would expect the difference images obtained from images belonging to different individuals (extrapersonal images) to be different from the ones obtained from intrapersonal images due to the large mismatch in facial features. We propose a Bayesian age-difference classifier that is built on a probabilistic eigenspaces framework to separate the intrapersonal and extrapersonal images and further classify the intrapersonal images based on age differences. The following figure provides an overview of the proposed approach and further illustrates facial similarity as a function of time.:
Figure 4 : Overview of the proposed age-difference classifier & Similarity scores with increase in age differences
II. Facial Similarity across Disguises (ICIP 2004):
How similar are disguised faces ? How do we circumvent non-uniform illumination across face images while computing facial similarity measures ? How sensitive are facial similarity measures to different kinds of facial disguises, illumination variations etc. ? These are some of the questions that we inteded to answer in this work. We began with suggesting a framework to compensate for pose variations (contour based approach) and with introducing the notion of ‘Half-faces’ to circumvent the problem of non-uniform illumination. A set of disguised face images of an individual (acquired from National Geographic) and the set of disguised images from the AR Face database, were used in our experiments. The disguises included dark glasses, facial hair, hats, false teeth and combinations of such props. We observed that illumination variations and severe disguises had comparable effects on facial similarity scores . Further, we observed that on those images where the illumination is non-uniform, that half of the face with better illumination performs better in terms of recognition than the regular full-faces. Our experimental results reflected the fact that disguises that significantly alter the pixel intensities greatly affect facial similarity scores.
Figure 5 : Similarity scores on disguised images of the same individual