Abstract
Age estimation is a complex issue of multiclassification or regression. To address the problems of uneven distribution of age database and ignorance of ordinal information, this paper shows a hierarchic age estimation system, comprising age group and specific age estimation. In our system, two novel classifiers, sequence knearest neighbor (SKNN) and rankingKNN, are introduced to predict age group and value, respectively. Notably, rankingKNN utilizes the ordinal information between samples in estimation process rather than regards samples as separate individuals. Tested on FGNET database, our system achieves 4.97 evaluated by MAE (mean absolute error) for age estimation.
1. Introduction
Age is a significant attribute of humans and plays an important role in interpersonal communication. It reveals people’s personal conditions and social background, as well as conducting human behaviors. Therefore, age estimation has become a hot research area in the field of computer vision. It has many potential applications in real applications, such as electronic customer relationship management, security control and surveillance monitoring, humancomputer interaction, biometrics, and criminal investigation [1, 2].
Over the past decade, many approaches have been proposed for age estimation. Geng et al. [3] defined aging pattern as the sequence of a particular individual’s face images sorted in time order, by constructing a representative subspace. Given an unseen face image, they projected it into a subspace that could best reconstruct it with minimum reconstruction error, and the position of the face image in that aging pattern was the age value. Guo et al. [4] proposed a manifold learning method for age estimation. The original face image space was mapped into a lowdimensional subspace by using special subspace learning method. Then, they designed a locally adjusted robust regression algorithm to learn and predict human age. Afterwards, Guo and Wang [5] found that there existed relations between age prediction and expression changes and also built a robust system to conduct crossexpression age estimation. Wu et al. [6] made emphasis on facial shapes, which were modeled as landmarks on Grassmann manifold. Then, these points were projected onto tangent space, and age estimation was fulfilled by TangentSpace Regression. Guo et al. [7] established hybrid features, combining features, such as shape feature, texture feature, and frequency feature. They utilized SVM and SVR to predict human age.
However, there are two problems confronted by researchers on age estimation at present. (1) Uneven distribution is common for facial aging databases. (2) Intrinsic correlation of ages is often overlooked, since samples are regarded as independent individuals and ordinal information is lost. The issues both have a great influence on the accuracy of age prediction, whereas they are often neglected. Most researches, including some of the papers listed above, enhance prediction performance mainly by extracting features containing more affluent age information and utilizing more robust and complex classifiers.
Till now, some works have considered the ordinal information of samples. Chao et al. [8] proposed “labelsensitive” concept. Instead of treating each class independently, they considered the samples with similar class labels. The weights of similar samples were also assigned based on the label similarity. In addition, they presented locality preserving projection algorithm to avoid overfitting and explore the connections between facial features and aging labels. Lu and Tan [9] aimed to use ordinary characteristics of age information to learn discriminative features in low dimensions. They proposed an ordinary preserving manifold analysis method to build lowdimensional subspace, on which samples were projected. Finally, to gain the relation between lowdimensional features and sample values for age estimation, a multiple linear regression model was formed through learning. Li et al. [10] tried to preserve the ordinal information among aging processes, so they employed ordinal discriminative feature learning. To solve uneven distribution of database, a combination of age group and specific age estimation was adopted. Li et al. [11] firstly divided test samples into two groups: birth to adulthood and adulthood to old age. In the next phase, they were detected within certain groups.
In this paper, we introduce a hierarchical system with two layers, age group and age estimation, to tackle uneven distribution problem and use ordinal information. We categorize samples into several age groups and then conduct age estimation within groups to avoid predicting ages on the whole database. Moreover, we employ rankingKNN to introduce ordinal information under ranking framework.
The rest of this paper is organized as follows. Section 2 describes our proposed system and related algorithms, including feature extracting, age group classification, and age value estimation. Section 3 presents experimental results on FGNET database [12] with inner and outer comparisons, and the conclusion is showed in Section 4.
2. The Proposed System and Algorithms
Figure 1 depicts the outline of our system, which contains three main parts: sample preprocessing, feature extraction, and estimation. Age information of a facial image is encoded in both texture and shapes. In feature extraction phase, we use appearance features including shape ratio features, wrinkle features, and ULBP (uniform local binary pattern) features [13]. They are combined to describe facial images, acting as the input of following age estimation stages. As for estimation part, it consists of age group estimation and age value estimation. The result of age group estimation is used to guide the next step: age value estimation.
2.1. Sample Preprocessing
In many practical applications, it is difficult to ask users to keep frontal face postures, which will further affect the accurate localization of local facial regions and the computation of facial geometric measurements. To decrease the influence of nonfrontal facial posture, we conduct sample rotation inplane. First, the rotation angle inplane is computed using the location of two eyes. If the angle is larger than a specific threshold, rotation correction with will be added for the sample. Considering marking deviations of facial landmarks, we set the threshold as . The computation of rotation angle inplane is shown as follows: where and are the locations of the left and right eyes, respectively.
2.2. Feature Extraction
2.2.1. Geometric Feature
Craniofacial changes during aging process are obvious for the young, distinguishing juvenile group from other adult groups effectively [14]. These changes can cause difference between minors and adults in some facial point distances. Wu et al. [6] pointed that “although facial shape could be affected by many factors, such as expression, pose, and age, it still conveys much information about the age of the subject.” In the famous FGNET aging database, each sample has a corresponding file recording sixtyeight facial feature points. We choose eight distances, which can be calculated by fifteen landmarks from these sixtyeight points. However, sample size normally can influence the distance between certain points, introducing instability to shape description. A good solution to eliminate the dependency on the image scale is to apply shape ratios instead of the distances between facial feature points [15]. We calculate eight distances and then form six ratios as geometric feature.
Let be the Euclidean distance between facial landmark and landmark , and let be the width of cheek between left and right ears. Six shape ratios are defined as
2.2.2. Texture Feature
Good texture descriptors are very important for image appearance. In this paper, we choose ULBP to describe detailed facial local texture. The algorithm of local binary pattern is initially presented by Ojala et al. [13], which is a simple yet very efficient texture operator. It labels the pixels of an image by thresholding the neighborhood of each pixel and considers the result as a binary number [16]. A further modification is ULBP, introducing bit transitions (from 1 to 0 or vice versa) to reduce the dimension of traditional LBP. Then, an LBP pattern is recognized as a uniform pattern if it contains the bit transitions no more than twice.
Besides ULBP textures, we also extract wrinkle density (WD) [15], a simple and rough way to present wrinkle texture. This simple feature is used in age group estimation step. Let be a wrinkle region, let be the pixel amount of canny edge image of , and let be the area of . WD is just the ratio of and listed as follows:
2.2.3. Combined Feature
Aging is mainly reflected by local wrinkles. When extracting aging features, facial partition is widely used. Considering the influence of moustache, we extract texture features in three local regions: forehead, eye corner, and face cheek. We apply some of 68 facial landmarks shown in Figure 2 to get these regions.
(a) Local regions
(b) Some of the used landmarks
Assume that is the coordinate of the th facial landmark, and are the coordinates of the topleft and lowright points of a rectangle region, respectively, and is the function to get the maximum value of giving data. The detailed localization methods are listed in Table 1.
To reduce the influence of facial deflection in explane, only the larger regions (left or right) of eye corner and face cheek regions are used. So for one image, only three regions are used to extract ULBP texture features and WD features.
In the step of age group estimation, geometric ratios and WD are used. More subtle appearance feature, a combination of ratios and ULBP texture features, is used in the next age value estimation step to get a more accurate estimation within a special age group.
2.3. Age Estimation
Uneven distribution of aging dataset has an important influence on age estimation accuracy [17]. Actually, two issues lie in uneven distribution for aging database. For one thing, some algorithms use the images of the same person to build a specific aging model. If the samples belonging to this person in the train set are not sufficient to cover a wide age range, reconstruction of the missing ages will be added to assure the integrity. This process may introduce error, because different people may have different aging laws. For another, most samples fall into several age groups, while other age groups comprise much less samples. Unfortunately, sample collection for age estimation is very tough, so effective algorithms are needed to make up this problem. For example, FGNET is a widely used aging dataset. Sample distribution is displayed in Table 2. Obviously, most of samples fall into 0–29 age groups (85%). Besides, there are no samples for some certain ages.
For age estimation, traditional solutions regard it as a multiclassification or regression problem. However, multiclassification approaches neglect the inherent relationship between labels, assuming that labels are independent. By contrast, regression methods may make use of ordinal information, but, due to difference in personalized aging pattern, kernel functions would be unstable, leading to overfitting problem in learning process easily [18].
In our paper, we treat age estimation as a multiclassification issue and try to utilize ordinal information to construct classifiers. A twolayer estimation system (shown in Figure 3) is employed to conduct age group detection and then age estimation. First, in the age group estimation step, we set all age values into an ordinal category sequence and then map category labels into different age groups by Sequence KNN. For age value estimation step, age value is estimated within a given age group under ranking model with serials of binary KNN classifiers combination.
2.3.1. Sequence KNN
To reduce the influence of uneven distribution, we bring out an original simple algorithm, SKNN (sequence KNN), to predict age group. Based on the age order of samples, a mapping is built to gain new labels from age value to ordinal category label. The mapping makes samples belonging to several age labels classified into one group. By this way, the gap between age labels can be reduced, leading to several better distributed subsets of aging dataset. Then, we can conduct age estimation within certain age groups to avoid the uneven distribution confronted by the whole dataset. The description of SKNN is listed in Algorithm 1.

Given a dataset (), where is a sample, refers to its age label, and is the number of total samples, we divide all the age labels into categories (); namely, . For each category, the age span is .
After the output of rough age value , we then map it into a special age group which is used to get the more precise age value in the following estimation step by rankingKNN. Suppose there are age groups divided beforehand; for each group , the boundary age values are and ; the mapping from an age value to a group is shown as follows:
2.3.2. RankingKNN
To utilize the inherent ordinal information between samples, we propose rankingKNN to predict age value. Under ranking framework, age value estimation is viewed as a series of binary enquiry problems. Each classifier outputs 1 or 0 by checking whether a test sample looks older than the images within a given subset which is formed in age label order. Thus, the ordinal information is used. In the ranking framework, KNN is used as the base classifier.
We consider the nonrepeating age labels as the rank order. Suppose denotes a subdataset with age labels larger than and denotes a subdataset with age labels no more than . The outline of rankingKNN is described in Algorithm 2.

3. Experiments and Analysis
We evaluate the effectiveness of the proposed system on the widely used FGNET aging dataset using leaveonepersonout (LOPO) testing strategy. Two popular evaluation criteria, namely, mean absolute error (MAE) [18, 19] and cumulative score (CS) [1, 2], are adopted. In the following experiments, neighbor number is set to 30 in age group estimation module and 15 in age value estimation.
3.1. Grouping Evaluation
Group division may influence MAE. Actually, to simplify grouping process, grouping with equal age span is widely used. We also adopt this way to divide all ages into several groups. Considering few ages over 50, to divide ages over 50 into several groups makes no sense. Given group span , we divide samples into . Figure 4 presents MAE curves changing with different group divisions.
From Figure 4, we can see that when group span is more than 10, MAE rises quickly over 6.0. For the group span from 2 to 7, MAE changes slightly. The best performance can be achieved by . The reason may be that if group span is too large, more samples with very different age labels will drop into the same group, which will further affect the estimation accuracy of next step.
3.2. Feature Evaluation
In order to test the role of different features used in this paper, we exclude geometric ratio features, wrinkle density features, and ULBP texture features separately in feature extraction stage. Meanwhile, age estimation part is still a hierarchical structure. Thus, we get different fusions of features: WD plus ULBP, ratio plus ULBP, and ratio plus WD. Then, we evaluate MAE under the hierarchical framework with group span . Table 3 gives the evaluation results.
From Table 3, it is obvious that different features play different effects, because they depict a face from different views. Exclusion of ratio features from the feature combination will lead to a dramatic rise of MAE. The reason may be that there are a lot of young samples in the database. This is also one of the reasons that the combination of ratio and WD can outperform the combination of ratio and ULBP. ULBP is more suitable to describe adult. Only with the combination of ratio, WD, and ULBP texture under our hierarchical framework can the best system performance be achieved.
3.3. Contribution Evaluation of System Components
Our hierarchy system consists of age group estimation (SKNN) and age value estimation (rankingKNN). Experiments also are carried out to evaluate their roles. The detailed tests include the following four experiments.
Whole system (our system): system performance is tested by using the overall hierarchy system (shown in Figure 3), consisting both age group and age value.
System without rankingKNN (replaced by KNN): similar to the whole system, the system is also hierarchical, but in the age value estimation stage rankingKNN is replaced by traditional KNN.
System without hierarchy system (rankingKNN): age group estimation component is excluded, so the system just uses rankingKNN to estimate age. Obviously, this is a nonhierarchical structure.
System without hierarchy system (SKNN): rankingKNN estimation component is excluded, so the nonhierarchical system just includes SKNN, namely, age group estimation part.
The results are shown in Table 4 and Figure 5. Actually age estimation system with an error more than 10 years is not acceptable, so we just test the CS with error level just within 10 years old.
As is showed in Table 4 and Figure 5, the experiment of “system without rankingKNN” shows an increase in MAE and a slight decrease in CS curves, because we use traditional KNN to replace rankingKNN. It proves that rankingKNN performs better than KNN for age estimation. For the system without hierarchy framework (rankingKNN), MAE drops to 7.65 from 4.97, which proves that age group estimation plays an important role in age value estimation. The reason may be that there exist some “confused” samples with similar aging features but very different age. Limiting age value estimation within a specific group, instead of the whole database, can reduce the influence of these “confused” samples. For the system just with SKNN to get the age value, system performance also drops quickly. Clearly, the two experiments of “system without hierarchy framework” gain apparent MAE decrease and low CS. It well testifies that hierarchic framework can boost estimation accuracy greatly.
3.4. Performance Evaluation of Overall System
Finally, we measure the performance of our proposed system on different age ranges and compare it with some other published systems.
Table 5 reports the results on different age ranges of our proposed system.
The MAE of the whole database is 4.97. From Table 5, it is clear that the sample volume of an age group greatly affects the final prediction result. With the decrease of the sample number within a group subset, MAE increases accordingly. This also can testify that uneven distribution is a key factor that affects the final estimation accuracy. The collection of aging database with wide span and enough higher ages is a very tough and urgent issue in the future.
The performance comparisons between several recently published systems and ours are shown in Table 6. Especially, the four systems listed in the left column are partly similar to ours, either with KNN or ranking model, using LOPO test rule. Table 6 shows that our system performs better than some of the published estimation systems.
4. Conclusions
This paper brings up a twolayer system for age estimation, combining age grouping and age value calculation. First, we use sequence KNN algorithms to classify an unknown sample into a special age group by fusing geometric ration and wrinkle density features. Then, appearance features binding geometric ration and ULBP textures are employed to estimate the age value under rankingKNN framework. Especially, by translating age label into category label at age group estimation stage, the whole age span is divided into some ordinal subspans, which can help to reduce somewhat the influence of uneven distribution problem confronted by current face datasets. For ranking model, it takes the age evaluation task as the combination of many ordinal binary classifiers which can utilize the ordinal information hidden among the aging processes.
In the future, we will put more effort on sample preprocessing, considering the great influence from facial occlusion, expression, and pose. At the same time, we will try to refine our age group estimation algorithm, since its result has a crucial effect on the whole system’s performance.
Conflict of Interests
The authors declare that there is no conflict of interests regarding the publication of this paper.
Acknowledgments
This research is partially sponsored by the Natural Science Foundation of China (NSFC) under Contracts nos. 61105120, 61170115, 61170117, and 61372090 as well as the National Key Development Plan of Fundamental Research no. 2011CB505402. The authors also would like to thank Dr. A. Lanitis for providing FGNET aging database.