Facial Age Estimation
|Andreas Lanitis (2010), Scholarpedia, 5(1):9701.||doi:10.4249/scholarpedia.9701||revision #91250 [link to/cite this article]|
Facial Age Estimation
Age estimation is the determination of a person’s age based on biometric features. Although age estimation can be accomplished using different biometric traits, this article is focused on facial age estimation that relies on biometric features extracted from a person’s face. The main issues presented in the article involve typical applications where facial age estimation can be used, problem and challenges associated with facial age estimation, typical approaches reported in the literature and future research directions.
The appearance of a human face is affected considerably by aging (see Figure 1). Facial aging effects are mainly attributed to bone movement and growth and skin related deformations associated with the introduction of wrinkles and reduction of muscle strength (Albert 2007, Rhodes 2009). Usually bone growth takes place during childhood whereas during adult ages the most intense age-related deformations are linked with texture changes. The observation of aging-related features on faces allows humans to estimate the age of other persons just by looking at their face. However, researchers who carried out work in studying the process of age estimation by humans (Rhodes 2009) conclude that humans are not so accurate in age estimation hence the possibility of developing automatic facial age estimation methods poses an attractive direction.
In automatic facial age estimation the aim is to use dedicated algorithms that enable the estimation of a person’s age based on features derived from his/her face image. The facial age estimation problem shares several similarities with other typical face image interpretation tasks where the execution stage includes the process of face detection, location of facial characteristics, feature vector formulation and classification. According to the application for which an age estimation system is intended to be used, the output of the classification stage can be an estimate of the exact age of a person or the age group of a person or even a binary result indicating whether the age of a subject is within a certain age range. Among the three variations listed above age-group classification is the most widely used as in most applications it is only necessary to obtain a rough estimate of a subject’s age rather than his/her exact age. Another important factor pertaining to the age estimation problem is the range of ages considered. This parameter is an important aspect of the problem as different aging characteristics appear in different age groups; hence a system trained to deal with a specific age range may not be applicable to more diverse age ranges.
An important aspect of the age estimation problem is the formulation of suitable metrics for assessing the performance of age estimators. The most widely used error metric is the Mean Absolute Error (MAE) between actual and estimated ages of faces in a test set. Geng et al. (Geng 2007) also propose the use of the cumulative score (CS) that shows the percentage of cases among the test set where the age estimation error is less than a threshold. The CS measure is regarded as a more representative measure in relation with the performance of an age estimator. In the case that age-group classification is considered, the percentage of correct age-group classifications can also be used for performance evaluation.
The facial age estimation problem shares similarities with the age progression problem. Age progression is the prediction of the future facial appearance of a subject based on images showing his/her previous facial appearance. Both age estimation and age progression need to take into account age-related facial deformations encountered during the lifetime of a subject. However, the two problems are in effect inverse problems since in age estimation information extracted from face images is used for determining the age of a subject whereas in age progression given a target age a face image that displays typical aging characteristics associated with the target age group is synthesized. In some cases the age estimation problem is treated individually (Kwon 1999, Guo 2008, Fu 2008, Wang 2009) whereas in other cases age estimation and age progression are both treated using similar methodologies (Lanitis 2002, Geng 2007, Suo 2008).
The process of age determination could figure in a variety of applications ranging from access control, human machine interaction, person identification and data mining and organization. Typical applications for each category mentioned above include:
Age-Based Access Control: In some cases age-based restrictions apply to physical or virtual access. For example age-related entrance restrictions may apply to different premises, web pages or even for preventing the purchase of certain goods (e.g. alcoholic drinks or cigars) by under aged individuals. In most cases age–based restriction access control is enforced based on the judgment of humans, the presentation of documentation papers or based on data provided voluntarily by the user. As an alternative automatic facial age estimation can be applied in an attempt to provide objective, accurate and non-invasive determination of the age of a person seeking access to a specific physical or virtual domain.
Age Adaptive Human Machine Interaction (HCI): Persons belonging to different age groups have different requirements and needs related to the way they interact with computers or other machines. Automatic age estimation can be used for determining the age of a computer/machine user and automatically adjust the user interface in order to suit the needs of his/her age group. For example icon-based interfaces can be activated for young children whereas text with large font can be activated when dealing with older users. Age adaptive HCI is particularly useful for publicly available resources such as information kiosks.
Age Invariant Person Identification: Age invariant identity verification can be developed by applying age progression techniques for deforming the face of a subject in order to predict how the subject will look like in the future. Age progression algorithms often require information related to the current age of a person, hence an accurate facial age estimation system can play a key role in developing automatic age progression systems, supporting in that way age-invariant identity verification.
Data mining and organization: Age estimation systems can be used for age-based retrieval and classification of face images enabling in that way automatic sorting and image retrieval from e-photo albums and the internet.
Age estimation shares many problems encountered in other typical face image interpretation tasks such as face detection, face recognition, expression and gender recognition. Facial appearance deformations caused by different expressions, inter-person variation, lighting variation, face orientation and the presence of occlusions have a negative impact on the performance on automatic age estimation. However, when compared to other face image interpretation tasks, the problem of age estimation displays additional unique challenges that include:
Limited inter-age group variation: In certain cases differences in appearance between adjacent age groups are negligible, causing difficulties in the process of age estimation. This problem is escalated when dealing with mature subjects.
Diversity of aging variation: Both the rate of aging and type of age-related effects differ for different individuals. For example the amount of facial wrinkles may be significantly different for different individuals belonging to the same age group. As a result of the diversity of aging variation, the use of the same age estimation strategy for all subjects may not produce adequate performance. Several factors could influence the aging process including race, gender and genetic traits. For this reason different age estimation approaches may be required for different groups of subjects.
Dependence on external factors: External factors influence the rate and the aging pattern adopted by an individual affecting in that way the process of age estimation. Typical factors that affect aging patterns include health conditions, lifestyle, psychology and deliberate attempts to intervene with the aging process through the use of anti-aging products or cosmetic surgeries.
Data availability: The development of accurate age estimation systems requires the existence of appropriate datasets suitable for training and testing. Suitable datasets should contain multiple images showing the same subject at different ages covering a wide age range. Since aging is a type of facial variation that cannot be controlled directly by humans, the collection of such datasets requires the use of images captured in the past. Although currently there are two publicly available datasets (MORPH (Ricanek 2006) and FG-NET (Lanitis 2008)) that aim to support experimentation in the area of facial aging, none of them fulfills all requirements for a dataset suitable for age estimation experiments because the MORPH database contains only few samples per subject whereas the FG-NET database contains images displaying significant non-aging related variation.
Age estimation approaches fall within two main streams. According to the first stream the problem is treated as a standard classification problem, solved using standard classifiers where age estimation is performed by assigning a set of facial features to an age group. Within this context facial features used may be associated with the general appearance of a face or may be associated to age-related features (e.g. wrinkles). As an alternative age estimation approaches that rely on the modeling of the aging process have been developed. In this section typical approaches described in the literature are briefly presented. The aim of this review is not to present an exhaustive literature review of the topic but rather to highlight the evolution of the topic. A more detailed presentation of the related literature is presented by Ramanathan et al. (Ramanathan 2009) and Fu et al. (Fu 2010).
One of the first attempts to develop facial age estimation algorithms was reported by Kwon and Lobo (Kwon 1999). Kwon and Lobo use two main types of features: Geometrical ratios calculated based on the distance and the size of certain facial characteristics and an estimation of the amount of wrinkles detected by deformable contours (snakes) in facial areas where wrinkles are usually encountered. Based on these features Kwon and Lobo (Kwon 1999) classify faces into babies, adults and seniors.
Lanitis et al. (Lanitis 2002, 2004) use an Active Appearance Model based coding scheme for projecting faces into a low dimensional space. Aging functions in the form of quadratic equations are used for relating the coded representation of faces to the actual age allowing in the way the estimation of the age of a subject. According to the results the use of person specific aging functions produced improved age estimation results when compared to the use of a common aging function for all subjects.
Geng et al. (Geng 2007) generate aging patterns for each person in a dataset consisting of face images showing each subject at different ages. Each collection of temporal face images is considered as a single sample, which can then be projected to a low dimensional space. Given a previously unseen face, the face is substituted at different positions in a pattern and the position than minimizes the reconstruction error indicates the age of the subject. Experimental results based on publicly available datasets prove that this method outperformed previous approaches reported in the literature and also performed better than widely used classification methods. The results of this work suggest that methods that aim to deal with the unique characteristics of aging can yield better results when compared to standard classification approaches.
Fu and Huang (Fu 2008) represent aging patterns using manifold learning. A discriminative subspace learning based on manifold criterion is developed for low-dimensional representations of the aging manifold. Regression is generally applied on the aging manifold patterns, which shows significant improvements on age estimation. Along these lines Guo et al. (Guo 2008) use a Support Vector Machine Regressor (SVR) for learning the relationship between coded face representations and age. A key aspect of Guo’s work is the use of a global SVR for obtaining a rough age estimate, followed by refined age estimation using a local SVR trained using only ages within a small interval around the initial age estimate.
Wang et al. (Wang 2009) presented an age categorization method that applies Error-Correcting Output Codes (ECOC) to the fused Gabor and LBP features of a face image to categorize a person into one of four possible age groups (child, teen, adult and senior adult). Age categorization (a multiclass learning problem) is solved using the combination of ECOC with AdaBoost or SVM. Experimental results on the FG-NET and Morph databases are reported to demonstrate its effectiveness and robustness in age categorization. The results show that the fused features are better than the one based on Gabor alone or LBP alone.
Most age estimation methodologies described in the literature use information from the overall face. As an alternative Suo et al. (Suo 2008) use a three-level hierarchical face model as the basis for age estimation. The first level is the global face representation; the second level refers to multiple local facial regions corresponding to different features and the third level involves the use of fine details such as wrinkles and hairline information. Experimental results indicate that the use of local features is important for achieving improved performance.
Instead of estimating the age of a subject, Ramanathan and Chellappa (Ramanathan 2006) estimate the age-difference between a pair of faces belonging to the same individual. The problem is treated as a classification task where difference vectors between pairs of age-separated faces are used for establishing the statistical distributions for different age range separations which are subsequently used during the age-separation classification problem.
Open Research Questions
The last few years several authors reported successful age estimation methods that produce performances comparable with the abilities of humans in the age estimation task. However, one of the main reasons that dictate the need for developing automatic age estimation systems is the failure of humans to perform the age estimation task precisely thus in the future it is necessary to develop systems that convincingly outperform the age estimation performance achieved by humans.
Ideally age estimation should operate on unconstrained face images in order to support the use of this technology in real life applications. For example age estimation based on faces captured by surveillance cameras or based on images captured by low resolution web cameras need to be investigated. Such scenarios support the application of non-invasive age estimation for access control.
So far experimentation in facial age estimation was limited to static images. The possibility of using temporal features for age estimation is an area for possible future research. This approach falls in the general area of behavioural biometrics (Yampolskiy 2008) where human actions are used for identity verification tasks.
A key issue pertaining to the future development of facial age estimation is the availability of suitable publicly available datasets. Future research efforts in age estimation should include efforts for generating suitable datasets to support both the training and comparative evaluation of different age estimation approaches reported in the literature.
- A.K. Albert, K. Ricanek and E. Patterson (2007). ‘A review of the literature on the aging adult skull and face: Implications for forensic science research and applications’. Forensic Science International, 172(1), pp. 1-9.
- Y. Fu, G. Guo and T.S. Huang (2010). ‘Age Synthesis and Estimation via Faces: A Survey’.IEEE Transactions on Pattern Analysis and Machine Intelligence. (in press)
- X. Geng, Z.H. Zhou and K. Smith-Miles (2007), ‘Automatic Age Estimation Based on Facial Aging Patterns’. IEEE Transactions of Pattern Analysis and Machine Intelligence, 29(12), 2234-2240.
- G. Guo, Y. Fu, C. R. Dyer, and T.S. Huang (2008). ‘Image-Based Human Age Estimation by Manifold Learning and Locally Adjusted Robust Regression’, IEEE Transactions On Image Processing, 17(7), pp 1178-1188.
- Y. Fu and T.S. Huang (2008). ‘Human Age Estimation with Regression on Discriminative Aging Manifold’, IEEE Transactions on Multimedia, 10(4), pp 578-584.
- Y. H. Kwon and N. da Vitoria Lobo (1999). ‘Age Classification from Facial Images’, Computer Vision and Image Understanding Journal 74(1), pp. 1–21.
- A. Lanitis, C.J. Taylor and T.F. Cootes (2002). ‘Toward Automatic Simulation Of Aging Effects on Face Images’. IEEE Transactions of Pattern Analysis and Machine Intelligence, 24 (4), pp 442-455.
- A. Lanitis, C. Draganova C. and C. Christodoulou C (2004). ‘Comparing Different Classifiers for Automatic Age Estimation’. IEEE Transactions On Systems Man and Cybernetics, Part B, 34(1), pp 621-629.
- A.Lanitis (2008), ‘Comparative Evaluation of Automatic Age Progression Methodologies’. EURASIP Journal on Advances in Signal Processing, Article ID 239480.
- N. Ramanathan and R. Chellappa (2006). “Face Verification Across Age Progression”. IEEE Transactions on Image Processing, 15(11), pp 3349-3361.
- N. Ramanathan, R. Chellappa, and S. Biswas (2009). Computational methods for modeling facial aging: A survey. Journal of Visual Languages & Computing 20(3), pp. 131-144.
- M. G. Rhodes, ‘Age Estimation of Faces: A Review’(2009), Appl. Cognitive Psychology. 23, pp 1–12.
- K. Ricanek and T. Tesafaye (2006). ‘MORPH A Longitudinal Image Database of Normal Adult Age-Progression’. Procs. of the 7th IEEE International Conference on Automatic Face and Gesture Recognition.
- J. Suo, T.Wu, S. Zhu, S. Shan, X. Chen, W. Gao (2008). ‘Design Sparse Features for Age Estimation using Hierarchical Face Model’. Procs. of the 8th IEEE International Conference on Automatic Face and Gesture Recognition.
- J.-G. Wang, W.-Y. Yau and H. L. Wang (2009), 'Age Categorization via ECOC with Fused Gabor and LBP Features'. Procs. of the IEEE Workshop on Applications of Computer Vision (WACV), pp.313-318.
- R.V. Yampolskiy and V. Govindaraju (2008). “Behavioural Biometrics: A Survey and Classification”, International Journal of Biometrics, 1(1), pp.81-113.
- Sheng Zhang and Matthew Turk (2008) Eigenfaces. Scholarpedia, 3(9):4244.