Learning Face Age Progression: A Pyramid Architecture of GANs
Hongyu Wang Yang Di Huang Yunhong Anil K. Jain
Beijing Advanced Innovation Center for Big Data and Brain Computing, Beihang University, China
Michigan State University, USA
{Hongyuyang, dhuang, yhwang}@buaa.edu.cn [email protected]
Summary
Facial age development has two important requirements, namely aging and accuracy persistent identity, but in the literature are not well learned. In this paper, we propose a method of generating confrontation based on an innovative network. The method of the inherent characteristics of the specific objectives and constraints to change the model were constructed according to specific age face the passage of time, to ensure that the resulting face can represent the desired effect of age, and can also guarantee the stability characteristics of the characters. Furthermore, a high level of age-specific characteristics in order to generate a more realistic facial details, synthetic human face to convey will be a pyramid against the discriminator estimated at multiple scales, and with a finer way to simulate the effect of age. The proposed method is applied in different situations in a posture, facial expressions and makeup wait distinguish face samples, and the ability to achieve significant age effect is vivid. In both the visual fidelity and quality evaluation show that the method is superior to other pre-existing best method.
1 Introduction
The development of age is an aesthetic process, presenting a given facial image, to present the effects of aging. It is often used in the entertainment industry and forensics, for example, predicted after the children grow up facial expressions or generated as a contemporary photograph of the missing person. The inherent complexity of the physical aging, interference by other factors (such as changes in PIE) and the lack of data marked aging of the common age-progressed to a very difficult problem. In the past few years, people solve this problem made great efforts aging of accuracy and identity persistence is often considered [29] [36] [26] [14] successful two basic premises . Early attempts mainly based on the anatomical structure of the skin, they are mechanically simulated contours and facial muscle growth along with the passage of time changes [31] [35] [23]. These methods provide a human face for the first time to understand the aging synthesis. However, they usually work in a complex way, it is difficult to generalize. Then using a data-driven approach, primarily by using the prototype aging test details to the face [13] [29], or to change the longitudinal portion corresponding age [28] [34] dependency between [20] be built mold to grow facial age. Although the visible signs of aging are combined well, but they often can not express complex function of the aging of the aging mechanism accurately, limiting the diversity of the aging model.
The depth generation network in the image generation [8] [9] [11] [30] showed a significant aspect of capacity, and the progress of age were studied. Compared with the previous conventional solutions, these people face a more attractive method of aging effects and fewer ghost. However, this problem has not been fundamentally resolved. Jutilaiyue, these methods will focus more on modeling transitions between the two age groups, the age factor plays a dominant role, and identity plays a subordinate role, leading to aging of accuracy and durability of body while at the same time difficult to achieve, especially long-term progress in achieving the age [18] [19]. In addition, multi-stage training needs more than one face images of different ages of the same individual, which involves another difficult problem that aging of the human face within an individual sequence acquisition [33] [15]. Both facts show that there is room for improvement in the current generation of aging depth method.
In this paper, we propose an innovative line of progression facial age, it combines the advantages of generating confrontation Network (GAN) on a priori knowledge in the field of credible and synthetic vision images of human aging. Compared with the existing methods in the literature, it is more dealing with two key requirements for the development of the age, namely the identity of durability and aging accuracy. Specifically, the method using a generator based on convolutional neural network (CNN) learning age of conversion, and are modeled over time according to different facial attributes. Between the output - and therefore the input image contained in the training squared Euclidean space loss, loss function to encourage the GAN generated face difficult to distinguish the elderly face training set, the identity while minimizing losses will be embedded in the characters and their characterization of advanced distance. It ensures that the resulting face showing the expected effects of aging, while the identity of the property remain stable. By data density estimate for each target age clusters, our method does not require a corresponding method that, like most across two age-domain matching the same person face right. Technology In addition, before the main crop region for the face (not usually include the forehead) compared to, we emphasize the synthesis of the whole face is very important, because some can significantly affect the perception of the age of the forehead and hair. To achieve this and to further enhance the aging detail, our approach takes advantage of the inherent hierarchical structure of the underlying network, and designed discriminator pyramid structure to fine-grained way advanced the estimated age-related clues. Our approach overcomes the limitations of a single age-specific representation, handling local and global age transition. Thus, a more realistic images generated (aging results shown in Figure 1).
[Image dump the chain fails, the source station may have security chain mechanism, it is recommended to save the picture down uploaded directly (img-WZLYDV5o-1582194544982) (file: /// C: / Users / JINXIL ~ 1 / AppData / Local / Temp /msohtmlclip1/01/clip_image002.png)]
The main contribution of this paper is:
· Propose a new age of progress GAN-based method, combined with age estimation and face verification technology to solve the coupled effect of age and identity to generate leads preservation;
· We emphasize the importance of closely related to the perception of age but was ignored in other studies of facial hair, forehead and components; it does improve the accuracy of the synthesis of age;
· Based on the existing experiments, we have established a new verification experiment, including the evaluation of commercial facial analysis tools and insensitivity evaluation based on facial expressions, posture, makeup changes
Our approach is not only proven to be effective and robust to age.
2. Related Work
In the initial exploration of the development of facial age, the skull to simulate the mechanisms of aging and facial muscles using a physical model. Todd et al [31] describes a modified cardioidal-strain converter, wherein the head is on a growth model geometry program built calculable. Wu et al. [35] Based on the anatomical structure of the skin, proposes a three-layer model to simulate the dynamic skin wrinkles. Ramanathan, Chellappa [23] and Suo [28], who also used the method of mechanical aging.
The method then mostly data-driven, less dependent on the a priori knowledge of the biology of aging is to learn from the training mode in the face. Wang et al [34], the mapping relationship between the sample and the corresponding lower face in the high resolution space tensor, and added details aging tensor space. Kemelmacher-Shlizerman [13] proposed a method based on the prototype, and illumination factors were considered further. Yang [36], who first solved the problem of multi-attribute decomposition, is achieved by component into the age of the target age group. These methods do improve results, but often appear in ghost synthetic human face.
More recently, it attempts to generate the depth of the network. In [33], Wang et al., By modeling in the intermediate transition state model RNN smoothly transformed different ages face. But each target in the training stage requires more face images of different ages, in the testing process requires accurate facial age label, which greatly limits its flexibility. From the encoder against the condition [37] of the frame, due to the aging of facial muscles simulated relaxation, but because of lack of skills training classifiers, presents only coarse wrinkles. Using Temporal Non-Volume Preserving (TNVP) aging method according to [18], [10] the density data of consecutive two age groups by mapping ResNet block, to achieve short-term development of age, long-term aging of the final synthesis stage by linking short . However, its main drawback is that it only takes into account the probability that a group does not have any identity information of the face of distribution. Therefore, aging results in a complete sequence of the synthetic human face has changed a lot in color, facial expressions and even identity.
Our research also makes use of GAN image generation capability, and presents a different but effective method. Age-related loss for the age of transformation, based on individual criticism is used to ensure the identity of clues stability, as well as multi-channel discriminator is applied to improve the details of the aging generation. This solution is the core issue deal with age, that the accuracy of age and identity preservation is more powerful.
3. Methods
3.1. Overview
A typical generator G comprises a GAN network discriminator and D, by opposing the training process interact. Generating function G important to try to capture the data density D and confused discriminant function, for optimizing to obtain distinguishability D, D can be distinguished so that the dummy natural face image and the face G generated from the generator. G and D can be approximated to the neural network, such as the Multilayer Perceptron (MLP). Risk function is:
[Image dump the chain fails, the source station may have security chain mechanism, it is recommended to save the picture down uploaded directly (img-kw0C9OFo-1582194544983) (file: /// C: / Users / JINXIL ~ 1 / AppData / Local / Temp /msohtmlclip1/01/clip_image003.png)]
z is the noise from the sample prior probability distribution of Pz, x represents a real face image from a distribution of Pdata. Convergence, the distribution Pg composite image would be equivalent to Pdata.
In recent years, the conditions GANs (cGANs) gained increasing importance, preprocessing image generation model approximates G (or control attributes) and the correspondence between the target dependencies. cGANs video prediction [17], the text image synthesis [24], the translation image to the image [11] [38], etc. have achieved good results. In our example, the generator will be based CNN young faces as input, and learn to map the elderly face corresponding domain. In order to achieve the effect of aging, while maintaining information associated with a user, using a composite criticism critic, that is a blend of traditional square Euclidean loss in image space, GAN loss to encourage generated face with training in the elderly face in terms of age indistinguishable, while the identity of the loss function by entering personal characteristics of the advanced features embedded characterization to minimize the - output distance. Figure 2 is an overview.
[Image dump the chain fails, the source station may have security chain mechanism, it is recommended to save the picture down uploaded directly (img-MtD1hbvg-1582194544984) (file: /// C: / Users / JINXIL ~ 1 / AppData / Local / Temp /msohtmlclip1/01/clip_image005.png)]
3.2. Builder
Age development of synthetic human face only needs to be spread through G before. Generating a network is a combination of an encoder and a decoder. After entering young face, it first layer encoding using the convolutional three steps to a potential space, capturing face attribute, it is possible with the passage of time tends to stabilize, and then a residual block by four [10] Construction of Input and Output the face can share a common structure, similar to [12]. Finally, to achieve transformation of the age of the target image space through three layers transpose convolution, it has been the development of age results in a given conditional young face. We do not use upsampling and max-pooling layer calculated feature map, but the use of the step 2 of 3 × 3 convolution kernel, to ensure that the contribution of each pixel and an adjacent pixel synergistic transformation. Convolution of all layers have been Instance Normalization and ReLU nonlinear activation function. Padding added to the layer, the input and output have exactly the same size. G architecture shown as supplementary information.
3.3. Discriminator
Age group critic bind the target system data density of the human face prior knowledge is introduced discriminator D, scalar scalar output D (x) represents the probability of data from x. After optimization, distribution Pg adult students face (We distributed the young face of x ~ Pyoung, adult students face distribution G (x) ~ Pg) should be equal to Pold distribution. Suppose we use the typical GAN [9], which uses the binary cross-entropy classification, the training process is to minimize the loss:
[Image dump the chain fails, the source station may have security chain mechanism, it is recommended to save the picture down uploaded directly (img-xWIdkFPN-1582194544985) (file: /// C: / Users / JINXIL ~ 1 / AppData / Local / Temp /msohtmlclip1/01/clip_image006.png)]
D and G always desirable uniform convergence; however, often realized in practice D faster resolvability and feedback to learn disappearance gradient G, as JS divergence local saturation. Recent studies, i.e. Wasserstein GAN [5], least squares GAN [16] and for loss sensitive GAN [22], discloses a basic problem is how to accurately define the probability distribution of the distance between the sequences. Here, we use least squares loss instead of the negative log-likelihood target, the target sample from the sample to punish the distance from the decision boundary in measurement space, thereby minimizing divergence Pearson X 2. In addition, in order to obtain more convincing facial detail and more vivid certain age, the young faces and old faces actual age of development are generated as a negative sample input D, while the real image of the elderly as a positive sample.
Therefore, the training process in turn minimizing the following:
[Image dump the chain fails, the source station may have security chain mechanism, it is recommended to save the picture down uploaded directly (img-MqdHja54-1582194544986) (file: /// C: / Users / JINXIL ~ 1 / AppData / Local / Temp /msohtmlclip1/01/clip_image007.png)]
Age Note (3) and (4) is connected Φage G and D function for extracting face images is conveyed correlation characteristics, as shown in FIG. 2. Considering the face of various age groups share the a common configuration and the same texture characteristics, and therefore a feature extractor Φage independently and D designed, the output advanced characterization to that generated human face more easily in terms of age real the elderly face separate. Especially Φage using VGG-16 structure [27], pre-trained to estimate the age of multi-label classification task, after the convergence, we removed the full link layer and connect it to our framework. Because natural images with multi-scale features and along the layered architecture, Φage capture value from the property gradually accurate pixel to a specific high-level semantic information age, this study internal pyramid hierarchy. Pyramid facial features characterizing a number of scales are estimated jointly D, in a manner to produce fine-grained aging effect.
Φage 2,4,7,10 convolution of the output layer is to be used. Through the D channel thereof, the finally obtained 12 × 3 series represented. In D, the base layer are all volumes followed Batch Normalization and LeakyReLU activation function, except that each rear left channel layer. D detailed structure of visible additional information on the advanced features in common estimate will be explained in Figure 3:
[Image dump the chain fails, the source station may have security chain mechanism, it is recommended to save the picture down uploaded directly (img-VzTPZnVY-1582194544987) (file: /// C: / Users / JINXIL ~ 1 / AppData / Local / Temp /msohtmlclip1/01/clip_image009.png)]
3.4. Identity preservation
A central issue facial age is to keep people dependent on stable property. We therefore measured by an appropriate input feature space - be introduced from the output-related constraints, which is sensitive to changes identity feature space, robust relative to other changes. Specifically, the face depth description of the network [21], using φid said personalized information is encoded, further defined identity loss function. φid are using Face Dataset one million people face figure from thousands of individuals to training. It was originally identified by N = 2,622 independent individuals to start; layer and then removing the last layer of the classification is adjusted φid (x) to improve the ability to verify the use of the training program triplet loss Euclidean space. In our example, φid cut into layer 10 the base layer volume, and then develop the identity loss:
[Image dump the chain fails, the source station may have security chain mechanism, it is recommended to save the picture down uploaded directly (img-jS7OQjPT-1582194544988) (file: /// C: / Users / JINXIL ~ 1 / AppData / Local / Temp /msohtmlclip1/01/clip_image011.png)]
d is the Euclidean distance between the feature characterization. More details of the implementation of the depth of the face descriptor may reference [21] (OM Parkhi, A. Vedaldi, and A. Zisserman. Deep face recognition)
3.5. The purpose
In addition to specially designed age-related GAN critic and identity persistence punishment, there is a gap in the pixel L2 loss based on a further narrowing the input and output of image space, such as color, etc., is expressed as:
[Image dump the chain fails, the source station may have security chain mechanism, it is recommended to save the picture down uploaded directly (img-HPiHwgjX-1582194544989) (file: /// C: / Users / JINXIL ~ 1 / AppData / Local / Temp /msohtmlclip1/01/clip_image013.png)]
x represents the input face, W, H, and C represents the shape of the image
Finally, training the loss of function of the system is:
[Image dump the chain fails, the source station may have security chain mechanism, it is recommended to save the picture down uploaded directly (img-wySNU2Q4-1582194544990) (file: /// C: / Users / JINXIL ~ 1 / AppData / Local / Temp /msohtmlclip1/01/clip_image015.png)]
We alternating training G and D, until the optimal solution G will finally learn the age of desired conversion, D will serve as a reliable estimator
4. Experimental Results
4.1. Data collection
Data sources are used for training GANs imaged MORPH mugshot standardized data set [25] and the celebrity datasets across ages relates PIE variables (Cross-Age Celebrity dataset, CACD) [7].
MORPH extended aging a database [25] contains a color image 52,099, having a proximal frontal pose, neutral face, and uniform illumination (that there are some minor changes in the pose and expression). Subjects ranged in age from 16 years to 77 years, with an average age of about 33 years old. A longitudinal study of the age span ranging from 46 days to 33 years old. CACD is collected through a Google image search [7] public data sets, including 2000 celebrities 163,446 faces between 10 years, ranging in age from 14 years old to 62 years old. The image data set having the largest number of changes with age, showing the change of pose, illumination, expression, and control of acquisition is less than MORPH. We mainly use MORPH and CACD training and validation. Using FG-NET [4] for the test, a fair comparison with previous work. The data set in the more popular prior to aging in the analysis, but the image 82 contains only 1,002 individuals. More nature of these databases can be found in the supplementary materials.
4.2. Implementation details
Before the network input image, using the data set itself provides eye location (CACD) or face ++ [3] Online recognition API (MORPH) aligned human face. MORPH removed image is not detected, the final concentration of the two data respectively 51,699 and 163,446 sheets of face images, cut into 224 × 224 pixels. Since these two databases, the number of people over age 60 face are very limited and do not contain images of children, so we only consider the age adults. Before we follow the many studies reported in [36] [29] [37] [37] in [18] for each age group of 10-year time span, the age of progress applied under 30 face, synthesized a series of 30 years old, 40 years of age and 50 years of age progression renderings. Therefore, for different target age group, there are three separate training session.
Supplementary material G shows the architecture of the network and the network D. For MORPH, weigh parameters λp, λa and λi are set empirically to 0.10,300.00 and 0.005; the value CACD 0.20,750 and 0.005 respectively. During the training phase, we use Adam, the learning rate is 1 × 10-4, the right of every 2,000 iterations heavy attenuation factor of 0.5.
(I) updated in each iteration is determined from the
(Ii) the use of age-related and identity critic at each iteration Builder
(Iii) using the pixel level every 5 critic generator iteration
Network using batch size is 5,000 times the total iteration 8, trained eight hours on the GTX 1080Ti GPU machine
4.3. Performance Comparison
4.3.1 Experiment 1: Age Development
It was five fold cross-validation of. On CACD, comprising 400 individuals each fold, respectively, from the [14-30], [31-40], [41-50] and [51-60] four age groups, and a total of nearly 10,079,8,635,7,964 6,011 sheets of face image (the other four data range is not as a fold); MORPH in each fold of 4467,3030,2205 and 639 face four age groups from the composition, a total of nearly 2586 test By. Each time you run, four fold for training, and the rest for evaluation. Examples of the results of progress of age shown in FIG. As we have seen, although these examples are covered on race, sex, posture, makeup and face widespread, but visually, it seems credible, and be able to achieve convincing aging effect.
[Image dump the chain fails, the source station may have security chain mechanism, it is recommended to save the picture down uploaded directly (img-P5dJouCA-1582194544991) (file: /// C: / Users / JINXIL ~ 1 / AppData / Local / Temp /msohtmlclip1/01/clip_image017.png)]
4.3.2 Experiment II: Age model assessment
We recognize that progress in the face of age should predict the future of personal appearance on aesthetic, not just the appearance of wrinkles and identity preservation, therefore, in this experiment, by visual and quantitative analysis, age progression results provide a more comprehensive assessment.
Experiment II-A: Visual Fidelity: FIG. 5 (a) is displayed with glasses, occlusion and postural changes sample face images. With age, the human face is still realistic, with the original input is true; but in the past the prototype-based approach [29] [32] In this case not suited to the parameters of the aging model [26] [28 ] may also lead to ghosting. Figure 5 (b) shows some examples of hair aging. To our knowledge, almost all in the literature [36] [26] [13] [33] [33] [37] The method of the aging [15] are proposed without considering the case of aging of hair cut face study, mainly because the hair is not as structured as the face area. In addition, hair texture, shape and color are varied, making it difficult to model. However, the present method the whole face is input, and as expected, the simulated aging process, thin hair becomes thin. FIG 5 © demonstrated the ability to retain the necessary facial features in the aging process, FIG. 5 (d) shows the smoothness and consistency of aging changes, thin lips, bags and more obvious, deeper wrinkles.
Experiment II-B: Aging Accuracy: With the aging of the face, the estimated age should be increased. Accordingly, aging is measured by objective age estimation accuracy. We'll Face ++ [3] the face of online analysis tools applied to the synthesis everyone's face. Negative is not detected, the data set studied MORPH 22,318 faces aging test samples (in the cross-validation 5fold average face each test run 4,464). Table 1 shows the results:
[Image dump the chain fails, the source station may have security chain mechanism, it is recommended to save the picture down uploaded directly (img-BoAS4Tgc-1582194544992) (file: /// C: / Users / JINXIL ~ 1 / AppData / Local / Temp /msohtmlclip1/01/clip_image019.png)]
The average age of three age groups were 42.84 years and 50.78 years of age 59.91 years old. Desired viewing age range [31-40] [41-50] and [51-60]. It is true that lifestyle factors may speed up or slow down the rate of aging individuals, resulting in deviation estimated age and the actual age, but the overall trend should be relatively strong. Because of this inherent fuzziness, further data sets for all ages face an objective estimate as a reference. From Table 1 and Figure 6 (a), can be seen in Figure 6 ©, age estimated age of synthetic face with the real image of good agreement, and over time showed a steady growth trend clearly validates our method.
[Image dump the chain fails, the source station may have security chain mechanism, it is recommended to save the picture down uploaded directly (img-ZHjbqNDB-1582194544993) (file: /// C: / Users / JINXIL ~ 1 / AppData / Local / Temp /msohtmlclip1/01/clip_image021.png)]
On the CACD, we used 50,222 young faces of the aging synthesis results (average face each run 1044 tests). Although the clustering of different age distribution MORPH not have as good separation, but still indicate the age of the progress of the proposed method does capture the data density of a given age, the face subset. Specific results in Table 1 and FIG. 6 (b), FIG. 6 (d)
Experiment II-C: Identity Preservation: the objective of using the face ++ for face verification, to check the identity of the original property is preserved during the age. For each test face, we will input image and the corresponding aging simulation results are compared: [test Face, aging Face 1], [test Face, aging Face 2], [test Face, aging Face 3]; synthesis and statistical analysis of the face, i.e., [1 facial aging, aging face 2], [1 facial aging, aging face 3], [2 face aging, aged face 3]. And II-B experiment similar to this assessment uses 22,318 sheets of deformed young faces and their age progress renderings total verified 22,318 × 6 times. As shown in Table 2, the average authentication rate three age-progressed clusters obtained respectively 100%, 98.91%, and 93.09%; for CACD, there 50222 × 6-authentication, the average authentication rate is 99.99%, and 99.91%, respectively, and 98.28%. This is clearly demonstrated the ability to save the identity of the method.
[Image dump the chain fails, the source station may have security chain mechanism, it is recommended to save the picture down uploaded directly (img-m5vW7YXL-1582194544993) (file: /// C: / Users / JINXIL ~ 1 / AppData / Local / Temp /msohtmlclip1/01/clip_image023.png)]
Further, in Table 2 and FIG. 7, the face authentication performance increase over time between the two images is reduced, consistent with the physical properties of human facial aging [6], which may explain this assessment of CACD performance is better than MORPH.
[Image dump the chain fails, the source station may have security chain mechanism, it is recommended to save the picture down uploaded directly (img-1kY7BM4E-1582194544994) (file: /// C: / Users / JINXIL ~ 1 / AppData / Local / Temp /msohtmlclip1/01/clip_image025.png)]
Experiment II-D: Contribution of Pyramid Architecture: One model assumes that the discriminator D pyramid structure facilitates the generation of aging effects, so that more natural human aging face. Therefore, we compared single-channel discriminator, resulting in single-channel discriminator in the face directly input to the estimator, and not as the characteristic pyramid. Comparative Experiment employed the determined structures of the first path corresponds to the network link and the proposed pyramid Φage D is.
Figure 8 shows the results:
[Image dump the chain fails, the source station may have security chain mechanism, it is recommended to save the picture down uploaded directly (img-JAbLjzvp-1582194544995) (file: /// C: / Users / JINXIL ~ 1 / AppData / Local / Temp /msohtmlclip1/01/clip_image027.png)]
Visually, the details of the synthesis of the results of aging is not obvious. In order to make more specific and reliable comparison, further experiments using the quantitative evaluation and II-B and II-C provided similar statistical results as shown in Table 3.
[Image dump the chain fails, the source station may have security chain mechanism, it is recommended to save the picture down uploaded directly (img-6bEIw9uX-1582194544996) (file: /// C: / Users / JINXIL ~ 1 / AppData / Local / Temp /msohtmlclip1/01/clip_image029.png)]
Table, MORPH CACD and the estimated age is usually higher than the benchmark (see Table 1), and the mean absolute error for these two databases for each of the three clusters beyond 2.69 and 2.52 years old, showing the use of a larger pyramid structure deviation, respectively, and 0.79 large 0.50 years. This may be because the synthesis wrinkles comparative experiment not so clear, the face looks relatively messy. This may also explain the faces in Table 3 verify the cause of the decline of confidence. Based on visual fidelity and quantitative estimates, we can draw an inference, compared with the pyramid structure, before the widespread use of gan-based framework of the single-channel discriminator, aging changes in a complex modeling backward.
Experiment II-E: Comparison to Prior Work: For comparison with the previous work, we CACD training set, were tested on the FG-NET and MORPH database, previous studies of these are [26] [28] [ 33] [36] [19] [37] [18] [20] [15], they are the most advanced methods. In addition, one of the comparison of the most popular mobile applications aging, i.e. Agingbooth [1], and e-aging tool Face of the future [2]. Figure 9 shows some examples of the human face:
[Image dump the chain fails, the source station may have security chain mechanism, it is recommended to save the picture down uploaded directly (img-V8pF75ZC-1582194544997) (file: /// C: / Users / JINXIL ~ 1 / AppData / Local / Temp /msohtmlclip1/01/clip_image031.png)]
As can be seen, Face of the future and Agingbooth uses a prototype-based approach, in which the same aging mask directly applied to any given person's face, like most aging applications. Although the concept of this method is simple, but with age, is not a realistic human face. For literature, parametric methods have been published [28] and dictionaries reconstruction technique based on [36] [26] inevitably will ghosting. Technological advances can be generated in deep model [33] [37] [15] observed, but they are only concerned about the facial area to trim, but with age the facial aging lacked the necessary details. In further experiments, we collected 138 pairs of 54 individual pictures from published papers and invited the 10 human observers to assess what Zhang's face with age perform better in paired comparison. In 1380 votes, 69.78 percent of people support this method, 20.80% of people support the preliminary work, 9.42 percent said about the same. Furthermore, the present method does not require pretreatment such as cumbersome as before work, requires only two marker points can be aligned with the pupil. In summary, we can say that the proposed method is superior to the corresponding method.
5 Conclusion
In this study, compared with the conventional method of face progression of age, its key issues, namely the conversion accuracy of age and identity preserved, presents a different but more effective solution, and proposed a novel based on the GAN method. The method involves face recognition technology and age estimation techniques, used a simple merger pixel-level punishment, age-related loss of GAN achieve TRAINING critic age of transition, as well as to maintain a stable individual identity rely critic. In order to produce detailed signs of aging, a kind of pyramid discriminator, a more accurate way to assess high-level face representation. Through a lot of experiments, the aging images and quantitative evaluation of the results obtained show that the effectiveness and robustness of the method.
Acknowledgments
This work partially funded by China Development Plan (2016YFC0801002) by the national research, the National Natural Science Fund (61673033 and 61421003), and the China Scholarship Council (201,606,020,050) funding.
references
[1] AgingBooth*. PiVi & Co.* https://itunes.apple.
com/us/app/agingbooth/id357467791?mt=8*.* 8
[2] Face of the future*. Computer Science Dept. at*
Aberystwyth University. http://cherry.dcs.aber.
ac.uk/Transformer/index.html*.* 8
[3] Face++ Research Toolkit*. Megvii Inc.* http://www.
faceplusplus.com*.* 5
[4] The FG-NET Aging Database*.* http://www.
fgnet.rsunit.com/ and http://www-prima.
inrialpes.fr/FGnet/**. 5
[5] M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein GAN*.*
arXiv preprint arXiv:1701.07875, 2017. 4
[6] L. Best-Rowden and A. K. Jain. Longitudinal study of automatic
face recognition. PAMI the IEEE *,, 40, (1): 148162, Jan *
2018. 6
[7] B.-C. Chen, C.-S. Chen, and W. H. Hsu. Face recognition
and retrieval using cross-age reference coding with cross-age
celebrity dataset. IEEE TMM * 17 (6): 804-815, 2015. * 5
[8] A. Dosovitskiy and T. Brox. Generating images with perceptual
similarity metrics based on deep networks. In NIPS*,*
pages 658–666, Dec. 2016. 1
[9] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu,
D.Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative
adversarial nets. In NIPS*, pages 2672–2680, Dec.*
2014. 1*,* 3
[10] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning
for image recognition. In CVPR*, pages 770–778, Jun. 2016.*
2*,* 3
[11] P. Isola, J. Zhu, T. Zhou, and A. Efros. Image-to-image translation
with conditional adversarial networks. arXiv preprint
arXiv:1611.07004*, 2016.* 1*,* 3
[12] J. Johnson, A. Alahi, and L. Fei-Fei. Perceptual losses for
real-time style transfer and super-resolution. In ECCV*, pages*
694–711, Oct. 2016. 3
[13] I. Kemelmacher-Shlizerman, S. Suwajanakorn, and S. M.
Seitz. Illumination-aware age progression. In CVPR*, pages*
3334–3341, Jun. 2014. 1*,* 2*,* 5
[14] A. Lanitis. Evaluating the performance of face-aging algorithms.
In FG*, pages 1–6, Sept. 2008.* 1
[15] S. Liu, Y. Sun, W. Wang, R. Bao, D. Zhu, and S. Yan. Face
aging with contextural generative adversarial nets. In ACM
Multimedia*, pages 82–90, Oct. 2017.* 1*,* 5*,* 8
[16] X. Mao, Q. Li, H. Xie, R. Y. Lau, Z. Wang, and S. P. Smolley.
Least squares generative adversarial networks. arXiv
preprint ArXiv:1611.04076*, 2016.* 4
[17] Mr. Mathieu, C. Couprie, and Y. LeCun. Deep multi-scale
video prediction beyond mean square error. arXiv preprint
arXiv:1511.05440*, 2015.* 3
[18] C. Nhan Duong, Gia Quach K., K. Luu, N. Le, and M. SAVVIDES.
Temporal non-volume preserving approach to facial
age-progression and age-invariant face recognition. In ICCV*,*
pages 3755–3763, Oct. 2017. 1*,* 2*,* 5*,* 8
[19] C. Nhan Duong, K. Luu, K. Gia Quach, and T. D. Bui. Longitudinal
face modeling via temporal deep restricted boltzmann
machines. In CVPR*, pages 5772–5780, Jun. 2016.* 1*,*
8
[20] U. Park, Y. Tong, and A. K. Jain. Age-invariant face recognition.
IEEE PAMI*, 32(5):947–954, 2010.* 1*,* 8
[21] O. M. Parkhi, A. Vedaldi, and A. Zisserman. Deep face
recognition. In BMVC*, volume 1, page 6, Sept. 2015.* 4
[22] G. Qi. Loss-Sensitive Generative Adversarial Networks on
Lipschitz Densities*. arXiv preprint arXiv:1701.06264, 2017.*
4
[23] N. Ramanathan and R. Chellappa. Modeling shape and textural
variations in aging faces. In FG*, pages 1–8, Sept. 2008.*
1*,* 2
[24] S. Reed, Z. Akata, X. Yan, L. Logeswaran, B. Schiele, and
H. Lee. Generative adversarial text to image synthesis. In
ICML*, pages 1060–1069, Jun. 2016.* 3
[25] K. Ricanek and T. Tesafaye. Morph: A longitudinal image
database of normal adult age-progression. In FG*, pages 341–*
345, Apr. 2006. 5
[26] X. Shu, J. Tang, H. Lai, L. Liu, and S. Yan. Personalized age
progression with aging dictionary. In ICCV*, pages 3970–*
3978, Dec. 2015. 1*,* 5*,* 8
[27] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. arXiv preprint
arXiv:1409.1556*, 2014.* 4
[28] J. Suo, X. Chen, S. Shan, W. Gao, and Q. Dai. A concatenational
graph evolution aging model. IEEE PAMI*,*
34(11):2083–2096, Nov. 2012. 1*,* 2*,* 5*,* 8
[29] J. Suo, S. C. Zhu, S. Shan, and X. Chen. A compositional and
dynamic model for face aging. IEEE PAMI*, 32(3):385–401,*
Mar. 2010. 1*,* 5
[30] Y. Taigman, A. Polyak, and L. Wolf. Unsupervised crossdomain
image generation. arXiv preprint arXiv:1611.02200*,*
2016. 1
[31] J. T. Todd, L. S. Mark, R. E. Shaw, and J. B. Pittenger.
The perception of human growth. Scientific American*,*
242(2):132, 1980. 1*,* 2
[32] J. Wang, Y. Shang, G. Su, and X. Lin. Age simulation for
face recognition. In ICPR*, volume 3, pages 913–916, Aug.*
2006. 5
[33] W. Wang, Z. Cui, Y. Yan, J. Feng, S. Yan, X. Shu, and
N. Sebe. Recurrent face aging. In CVPR*, pages 2378–2386,*
Jun. 2016. 1*,* 2*,* 5*,* 8
[34] Y. Wang, Z. Zhang, W. Li, and F. Jiang. Combining tensor
space analysis and active appearance models for aging effect
simulation on face images. IEEE TSMC-B*, 42(4):1107–*
1118, Aug. 2012. 1*,* 2
[35] Y. Wu, N. M. Thalmann, and D. Thalmann. A plastic-viscoelastic
model for wrinkles in facial animation and skin aging.
In PG*, pages 201–214, 1994.* 1*,* 2
[36] H. Yang, D. Huang, Y. Wang, H. Wang, and Y. Tang. face
aging effect simulation using hidden factor analysis joint
sparse representation. IEEE TIP*, 25(6):2493–2507, Jun.*
2016. 1*,* 2*,* 5*,* 8
[37] Z. Zhang, Y. Song, and H. Qi. Age progression/regression by
conditional adversarial autoencoder. In CVPR*, pages 5810–*
5818, Jul. 2017. 1*,* 2*,* 5*,* 8
[38] J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros. Unpaired imageto-
image translation using cycle-consistent adversarial networks.
tion.* IEEE TIP*, 25(6):2493–2507, Jun.*
2016. 1*,* 2*,* 5*,* 8
[37] Z. Zhang, Y. Song, and H. Qi. Age progression/regression by
conditional adversarial autoencoder. In CVPR*, pages 5810–*
5818, Jul. 2017. 1*,* 2*,* 5*,* 8
[38] J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros. Unpaired imageto-
image translation using cycle-consistent adversarial networks.
arXiv preprint arXiv:1703.10593*, 2017.* 3
