Learning Face age progression- A Pyramid architecture of GANs

Generic badge Generic badge Generic badge Generic badge Generic badge Generic badge Generic badge

Abstract


Since the face age progression deals with addressing the two main aspects, i.e. aging accuracy and identity permanence, a Generative Adversarial Network is employed to model restraint for the inherent subject specific characteristics as well as the age- specific changes in faces with time.
For generating more realistic facial details pyramidal adversarial discriminator is utilised which models the effects of aging in a more advanced manner.

Goals/Objectives

Research in the face age progression can be employed in finding out the subjects that went missing, in our case we will be analyzing on new born babies which can further be used by the cops to search for the missing baby even after years he/she went missing.

Introduction

Age progression is the method of altering the facial images to demonstrate the consequences of aging on the face of an individual. The main reason behind opting for GAN in age progression is as follows:

  1. It facilitates age estimation and face verification while keeping the issues of aging effect generation as well as identity preservation,
  2. In the previous study forehead and hair components of face is not kept under considerations.

Previous Work

While solving the problem of face age progression innate complexity of physical aging and interferences occurring due to the other factors makes it harder. There are many attempts that have been employed for tackling the above problem:

  1. A compositional and dynamic model was employed where compositional model represent the faces in distinct age groups by hierarchy of And-Or-Graph(And nodes fragments the face into finer details i.e. hair, wrinkles, etc; Or nodes depicts the variety of face by substitutes). It is modelled as Markov process or the parse graph on the basis of two important aspects of face age progression i.e. accuracy and identity preservation [4].
  2. Analyzing the skin’s anatomical structure, a 3 layered dynamic skin model to counterfeit wrinkles [5], There are many other methods that were employed for face age progression, but in our work we will be implementing face age progression with the help of GANs.

Method

Overview
A typical GAN consists of a generator G and a discriminator D which is trained with the help of adversarial nets in iterations. The generative network comprises of both encoder as well as decoder. When the infant face is given as an input the generator make use of strided convolution layers to encode it to a latent space.

Dataset

We collected the dataset from CACD [1] is a comprehensive face dataset with a wide age range 0-116. It consists of 20,000 facial images both augmented as well as unaugmentated which is labelled as age, gender and ethnicity.

Method

We will be making use of the method cited in Learning Face Age Progression:A pyramid Architecture of GANs [2].
Using CACD dataset, we will be heading towards preprocessing the image data for finding out missing images. After the images are preprocessed it will be fed to the discriminator. In generator transformation of age into a target image space is attained by a strided convolutional layers, by making use of instance normalization, orthogonal weight normalization and ReLU non-linearity activation. For making the input and output size uniform, paddings are added.
In discriminator the pyramidal architecture is employed, actual faces and the output generated by the Generator is fed to the discriminator for obtaining more intense age-specific facial details, throughout training. Meanwhile, Batch Normalization and LeakyReLU is utilized at each convolutional layer, except the last one in each pathway.
The training process decreases the loss of Generator to Discriminator along with loss of Generator to Generator, to generate more photorealistic image in a fine-grained manner. For identity preservation we will be using the work done in Towards Open-Set Identity Preserving Face Synthesis [3], as it incorporates asymmetric loss function for training the input face image to generate an identity vector along with any other input face image to withdraw an attribute vector containing pose, emotions and illumination,etc.

References

[1] CACD database. https://www.v7labs.com/open-datasets/cacd.
[2]Hongyu Yang, Di Huang, Yunhong Wang, Anil K. Jain. 2019 Learning Continuous Face Age Progression: A Pyramid of GANs: 32-38

[3] Jianmin Bao, Dong Chen, Fang Wen, Houqiang Li, Gang Hua. 2018 Towards Open- Set Identity Preserving Face Synthesis.

[4] 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
[5] Y. Wu, N.M. Thalmann, and D. Thalmann. A plastic-visco-elastic model for wrinkles in facial animation and skin aging. In PG, pages 201,214, 1994. 1,2