Enhanced Cycle Generative Adversarial Network for Generating Face Images of Untrained Races and Ages for Age Estimation

• Published in: IEEE access Vol. 9; pp. 6087 - 6112
• Main Authors: Kim, Yu Hwan, Nam, Se Hyun, Park, Kang Ryoung
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2021; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Age; Age estimation; Chronology; Coders; Datasets; enhanced CycleGAN; Estimation; Faces; Generative adversarial networks; Generators; heterogeneous database; Image enhancement; Measurement; Race; Testing; Training; transformation of heterogeneous race and age images
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2020.3048369

The datasets used in recent age estimation studies largely consist of two races (i.e., Asians or Westerners), and despite the large amount of data available, the problems regarding age-class imbalances still arise, owing to different age distributions. This causes overfitting in training process, reducing the generality of the age estimation. This problem typically occurs in homogeneous datasets, e.g., using the same Asian database for training and testing or using a database with the same age range for training and testing. Consequently, the problems arise in heterogeneous datasets, e.g., using an Asian database in training and a Westerner database in testing or using databases of different age ranges in training and testing, and the accuracy inevitably degrades when heterogeneous datasets are used for training and testing. To solve these problems, we proposes an enhanced cycle generative adversarial network (CycleGAN)-based heterogeneous race and age image transformation technique, which can transform the images of one race and age range to those of different race and age range. The encoder and decoder of the generator in the proposed enhanced CycleGAN include residual connections, thereby preventing information loss as much as possible as the layer deepens. In addition, the generator of the enhanced CycleGAN uses identity loss and age loss functions between the generator-produced image and a multi-channel input image obtained through 3D one-hot encoding. Through this, the training is directed to increasing the similarity not only between the images but also between the age class labels. And the enhanced CycleGAN uses a second discriminator in addition to the existing discriminator, thereby addressing a problem in which training is not properly performed when the discriminator converges too fast relative to the generator in a conventional CycleGAN. Experiments with three open databases demonstrated that our method outperforms state-of-the-art methods for facial image transformation and age estimation.