More Filters. Human-Centered Social Media Analytics. View 1 excerpt, cites methods. Facial age estimation by nonlinear aging pattern subspace. ACM Multimedia. View 1 excerpt, cites background. Face recognition with temporal invariance: A 3D aging model.
Highly Influenced. View 3 excerpts, cites methods and background. Age invariant face recognition: a survey on facial aging databases, techniques and effect of aging. Artificial Intelligence Review. Facial age estimation by multilinear subspace analysis.
View 2 excerpts, cites methods. Age estimation from face images: Human vs. Human age, as an important personal trait, can be directly inferred by distinct patterns emerging from the facial appearance. Computer-based age synthesis and estimation via faces have become … Expand. Computer Science, Medicine. Learning from facial aging patterns for automatic age estimation.
MM ' View 1 excerpt, references methods. IEEE Trans. Pattern Anal. Highly Influential. View 7 excerpts, references methods. Face Verification Across Age Progression. View 2 excerpts, references methods. Comparing different classifiers for automatic age estimation. View 6 excerpts, references methods. But this from an image-class problem based on complete data into method relies on well controlled high-quality face images an image sequence - class sequence problem based on in- and can only classify faces into one of three groups babies, complete data.
To solve this problem, an iterative learning young adults, and senior adults , which greatly restricts its algorithm is specially developed to learn a representative usage in real applications. In their teristics of facial aging variation is proposed based on the work, the aging pattern is represented by a quadratic func- learned aging pattern subspace.
Some re- lated works are briefly introduced in Section 2. During the tion 3. After that, the AGES algorithm is proposed in Sec- training procedure, a quadratic function is fitted for each tion 4. As and analyzed. Finally in Section 6, conclusions are drawn for age estimation, they proposed four approaches to deter- and several issues for the future work are indicated.
Both formation about various external factors, such as gender, of them have been widely used in the studies of interface health, living style, weather conditions, etc. Such informa- usability, machine-mediated human communication, and al- tion is very difficult to obtain in real applications, which ternative input devices for disabled users. The technology makes WPS impractical.
The Weighted Appearance Specific has become sufficiently advanced that several companies of- WAS approach is the best one among the methods that do fer working commercial systems. Later Lanitis et al.
Among them, the Appearance-Age Specific y architecture AAS achieved the best performance because it handles the face image clusters separately according to 8 t both the appearance and the age group. This 3 prevents them from getting more satisfying results. In de- 2 1 tail, there might be four weaknesses in such approaches.
There is no evidence suggesting that the relation between face and age is as simple as a quadratic function. Second, the temporal characteristic cannot be well utilized by the aging function. The dependence relation among the Aging Pattern x aging faces is monodirectional, i. However, the relation re- vealed by the aging function is bidirectional: any changes Feature Extractor on a particular face will affect all other faces.
Although people age in different 0 1 2 3 4 5 6 7 8 ways, there must be some commonality among all aging pat- m m b2 m m b5 m m b8 terns, i. Such commonality is Aging Pattern Vector also crucial in age estimation, especially when the personal training data is insufficient. Fourth, the new aging function for the unseen face image is simply a linear combination Figure 1: Vectorization of the aging pattern.
The of the known aging functions, rather than being generated ages are marked at the top-left to the corre- from a certain model of aging patterns. All of these prob- sponding positions and above the corresponding fea- lems can be solved, from an entirely new point of view, by ture vectors.
The missing parts in the aging pattern the AGES algorithm proposed in this paper. Suppose the extracted feature If age estimation is regarded as a conventional classifica- vector b is n-dimensional, the number of interested ages is tion problem, then one straightforward way is to model face p. The problem is that different persons dimensional feature vector x. Each age is allocated n con- age in different ways. If the face image at a particular age each person. Consequently, the aging torization of the aging pattern when the interested ages are pattern, rather than the separate ages, should be modelled.
Definition 1. An aging pattern is a sequence of personal By representing aging patterns in this way, the concepts face images sorted in time order. All and the temporal characteristic of aging variation can be face images in an aging pattern must come from the same well utilized. So long as the aging patterns are well sam- person, and they must be arranged by time. Suppose a gray- pled, i.
How- aging pattern can be represented by a three-dimensional ma- ever this brings two other challenges: 1 The learning task trix P, where P x, y, t is the intensity of the pixel x, y in has been changed from a conventional image-class problem the face image at the time t.
Take the aging pattern shown into a novel image sequence - class sequence problem; 2 in Figure 1 as an example. Along the t axis, each age from The learning algorithm applied to the aging patterns must 0 to 8 in this example is allocated one position. If face im- be able to handle highly incomplete training samples. The ages are available for certain ages in this case 2, 5 and 8 , following section mainly tackles these two problems.
If not, the positions are left blank dotted squares. If all positions are filled, the aging pattern is called a complete aging pattern, 4.
Before the aging pattern can be further processed, the 4. A representative model for the aging patterns can be built In this paper, the feature vector is extracted by the Appear- up by the information theory approach of coding and de- ance Model [7]. The main advantage of this model is that coding. Each line shows one aging pattern from the age 0 to The ages are marked above the corresponding faces. The faces learned by the algorithm are surrounded by the dashed squares.
Note that the aging variation in the data set. The projection in the subspace is patterns are highly redundant, it is possible to estimate yk computed by only based on part of xk [14], i. The problem is that the aging pattern vector that correspond to the positions of xak. Suppose the subspace x is highly incomplete. There exist several approaches to is d-dimensional, then there are d unknowns in yk.
Thus PCA with missing data [10]. However, the statistical distri- at least d available features elements in xak are needed to bution of the aging patterns is unlikely to be normal, thus over-constrain the problem. The whole procedure repeats until the not suitable.
The convergence of this algorithm learn a representative subspace. Any sample in this set can be written as Proof. But xk by Wi. So tor b is first extracted by the feature extractor. Step 1 is to select an aging pattern suitable for I the 2nd characteristic , and step 2 is to find a proper During the learning procedure of AGES, the missing faces position for I in the selected aging pattern the 3rd char- in the training aging patterns can be simultaneously learned acteristic.
Step 1 can be achieved by finding a projection by reconstructing the whole aging pattern vectors through in the aging pattern subspace that can reconstruct b with Equation 3. Figure 2 gives some typical examples of the minimum reconstruction error. For clarity, only the faces in the ror cannot be actually calculated. Thus I is placed at every most changeable age range from 0 to 18 with 2 years as possible position in the aging pattern, getting p aging pat- interval are shown.
Noticing that b is the only available feature in zj , the patterns. Thus this learning algorithm can also be used to projection yj of zj can be estimated by Equation 5 , and simulate aging effects on human faces. Then the projection yr that The procedure of the learning algorithm is actually a pro- can reconstruct b with minimum reconstruction error over cedure of interaction between the global aging pattern model all the p possible positions is determined by and the personalized aging patterns.
Then, the global model is further refined by the updated Thus the suitable aging pattern for I is zr. Step 2 afterward personal aging patterns. In this way, the commonality and becomes trivial because r also indicates the position of I in the personality of the aging patterns are alternately utilized zr. Finally the age associated to the position r is returned to learn the final subspace. However, it fails ure 3. To make it more understandable, the face images to utilize the main advantage of age range estimation that both the original and reconstructed ones instead of fea- it is possible to obtain a better aging pattern training set ture vectors are show in the aging patterns.
It is interesting in both quantity and quality completeness. This advan- to note that when the test image is placed at a wrong po- tage can be exploited in the second way, i.
In this approach, each age range is represented by structed faces become ghost-like twisted faces. On the other an integer ar. Suppose different age ranges have no inter- hand, if the test image is placed at the right position, the section and each age range contains w different continuous aging pattern subspace can not only reconstruct the original ages.
If the age of a face image is age, then the new label of face very well, but also reasonably conjecture all other faces this image is derived by in the aging pattern. Such age range labels share ferent ages of the subject in the test image can be simulated similar characteristics with ages. The difference mainly lies at the same time without additional computation.
This aging pattern for the test image is generated based on both brings the advantage that more images might be available the aging pattern subspace and the face image feature.
The for each position. On the one hand, aging patterns with less subspace defines the general trend of aging, and the face or even no blank positions can be composed. On the other image feature indicates the personalized factors relevant to hand, large number of aging patterns for each individual can facial aging. By placing the feature vector at different posi- be generated through the combination of multiple images at tions, candidate aging patterns specified to the test face are different positions.
After the new aging pattern training set generated. Among these candidates, only one is consistent for age range estimation is generated, the learning and test with the general aging trend, which can be detected via min- procedures of AGES can be executed as those used in age imum reconstruction error by the aging pattern subspace. In estimation. The 5. There are totally face images from 82 different In real applications, sometimes the age range is more subjects in this database.
Each subject has face images meaningful than the exact age of a person. For instance, labeled with ground truth ages. The ages are distributed in the police are more often to describe an unknown suspect a wide range from 0 to Lanitis et al. In this experiment, the algorithms age into the corresponding age range. The images are collected under totally uncontrolled conditions. Besides the aging variation, most of the age-progressive im- age sequences display other types of facial variations, such as significant changes in 3D pose, illumination, expression, etc.
This greatly increases the difficulty of age estimation on these face images. Some typical aging face sequences in this database are shown in Figure 4. Only the face region not including hair is used as training data.
The face feature extractor used in the experiments is the Appearance Model, which is a combined model of shape and intensity. The shape model is trained on 68 manually labeled key points on each face image. For the intensity model, each face image is aligned to the mean face shape i. It has been shown that the Appearance Model is robust against many facial varia- tions such as illumination, view angle, and facial hair.
Refer to [7] for more details about this face model. All HumanB. The first line shows the gray-scale face of the algorithms are tested based on these dimensional regions used in HumanA. The second line shows feature vectors. However, it is discovered in this experiment that mation is also tested. Each ob- tions, the algorithms will always perform better than using server is asked to assign an age from 0 to 69 to each image GA for example, when using GA, the mean absolute error based on their estimation.
The results are calculated based of WAS is 8. The same image set is and the training procedure will be much faster. Thus in this presented to the observers in two different ways.
After they determine the quadratic aging functions. It is worth mention have made their estimations, the whole color images are then that the output of the aging function is a real number, but presented to them. The first test is denoted by HumanA, the ages in this paper are represented by non-negative inte- and the second one is denoted by HumanB.
Figure 5 gives gers. In reality, AAS is rounded to the nearest non-negative integer. There people usually estimate age based on multiple cues, such as are some additional parameters in AAS. In the experiment, face, skin, clothes, hair, body build, voice and movement. When the best performance is observed, the tion purely based on face, while HumanB intends to test error threshold in the appearance cluster training step is set the human ability of age estimation based on multiple cues to 3, and the age ranges for the age specific classification are including face, hair, skin color, clothes, and background.
To set as , , and Af- provided to the algorithms, while in HumanB, additional ter 82 folds, each subject has been used as test set once, and information is available to the observers.
In this way, the algorithms are tested in the case similar to 5. Meanwhile, the relatively limited data set can First the algorithms and the human ability are evaluated be adequately utilized. With additional information provided to humans in age estimation at error levels from 0 to the observers, HumanB gets remarkably lower MAE than 10 years. HumanA does. It is The cumulative scores of the algorithms and humans at a bit surprising that AAS gets a very poor result since it the error levels from 0 to 10 years are compared in Fig- was proposed as an improved version of WAS.
The reason ure 6. The situation at higher error levels is not shown might be that in this experiment the possible ages are because in general, age estimation with an absolute error much more than those tested in [12] Thus the cluster higher than 10 a decade is not acceptable. As expected, specific training scheme of AAS is much easier to overfit the the cumulative scores of HumanB are significantly higher training set and performs poorly on the test set.
This justifi- than those of HumanA at all error levels. But even Hu- cation will be further verified in the later experiment on age manB loses its leading position this time. Instead, AGES range estimation, where the possible classes are much less.
This is impressive since more information is provided in same information as that fed into the algorithms. Considering that even comparison between the algorithms and HumanA is more human beings cannot achieve very high accuracy in age es- meaningful. Al- age estimation by face regions?
Perhaps not. It does not provide enough information error levels. Not until the error level increases to 8 does on how accurate the estimators might be. But at least at the relatively important low error ceptable level, rather than the mean absolute error. Those levels from 0 to 10, the average accuracy of WAS is worse estimations with higher absolute error than the acceptable than that of HumanA.
Since the acceptable error As mentioned in Section 4. In this experiment, the per- els. At each level, a cumulative score can be range estimation are tested through the re-labeling method. In the experiments, AGES 1. The success of HumanA 1. In particular, the following three are the most important ones: 1 The aging patterns are regarded The performance of age range estimation can be evaluated for the first time as training samples and represented as by the same measurements used in age estimation, i.
MAE sparse temporal data which naturally integrate the concept as an indicator of average performance, and cumulative score of identity and time; 2 An effective learning algorithm is as an indicator of accuracy. However, since the class label developed to learn a representative subspace from the highly itself represents a range, the cumulative scores are only com- incomplete aging patterns; 3 A two-step age estimation ap- pared at the error level 0, i. If wider age range corre- proposed based on the aging pattern subspace.
It even does not gorithms can still be evaluated at the error level 0. It is need to known the age of the input image, which is required worth mention that the hit rate of the age range estimation by most other aging face simulation technologies. The per- with 5 continuous ages in each range is much tougher a cri- formance of AGES in aging face simulation still needs to be terion than the cumulative score at the error level 4 of the compared with the existing methods [3, 19, 20, 25, 30] in the age estimation.
If the estimated age range is correct, then future. Also the morphing technology might be integrated the absolute age error is definitely no higher than 4. But with AGES for better visual result. The advantage of multi-cue age estimation has the same. The work age range estimation are compared in Table 3. It can be described in this paper focus on age estimation by face im- seen that AGES achieves the highest hit rate and the sec- ages. One remarkable changes in pose, illumination and expression.
If the feature difference with the results of age estimation is that AAS per- extractor in Figure 1 could filter out all facial variations ex- forms much better this time than WAS does in either MAE cept for the aging variation, then better performance could or hit rate.
0コメント