Enhancing the Accuracies of Age Estimation With Heterogeneous Databases Using Modified CycleGAN

Abstract

Age estimation using face images has been widely employed across various fields. Because the characteristics of face images usually vary greatly depending on race, camera type, lighting, and other environmental factors, the recognition ability of untrained heterogeneous face image databases is not accurate by previous methods. Therefore, various attempts have been made where different heterogeneous databases were combined to enable training; however, the training time is extended and diverse environmental variables in databases cannot be sufficiently trained. To address these issues, this study proposes modified cycleconsistent generative adversarial network (CycleGAN) that generates an even distribution of heterogeneous face data, and an age-estimation method that is effective for heterogeneous data based on comparative CNN for age estimation (CCNNAE). In addition, we propose the method of reducing the errors caused by image transformation in modified CycleGAN through adaptive selection between transformed and original face images for the input to CCNNAE, which is based on age similarity between the transformed face image and the original face image. Experiments with two open databases, MORPH and MegaAge databases showed that our method outperformed the state-of-the art methods. INDEX TERMS Age estimation, heterogeneous database, modified CycleGAN, CCNNAE.

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

“…Both the MegaAge Asian [16] and Asian face age dataset (AFAD) [11] include only the Asian face images. In previous research [10], they confirmed that the age estimation accuracies with Morph database were higher than those with AFAD, and the age estimation accuracies with Morph database were also reported to be higher than those with MegaAge Asian in [20]. From these results, we can estimate that the more distinctive changes including facial shape and texture occur in Caucasian race than those in Asian according to face aging, which causes the age estimation to be easier with the face images of Caucasian compared to those of Asian.…”

“…A modified CycleGAN-based age estimation method based on race transformations of facial images has been proposed [20]. This study used a generator similar to the conventional CycleGAN [37].…”

“…It is designed so that the output image maintains the input image's resolution and information to some degree. The modified CycleGAN [20], which applies this technique and focuses on race transformations of facial images, was designed to enable smooth bidirectional training through a new generator loss.…”

“…The generator is typically constructed using an autoencoder [25] with a U-net [26] or residual block structure. The residual block structure was used in the CycleGAN [18] and modified CycleGAN [20]. Two outputs were used in the modified CycleGAN; the output from the middle layer focused on maintaining age information in the generated image as compared to the input image, whereas the output from the last layer focused on image generation.…”

“…the AFAD data were higher quality than the MegaAge data, the generated data shown in Figure 11 showed better race transformation results than the 1st sub-dataset samples in Figure 8. Figure 12 shows the loss graphs for Equations (11) and (20), as obtained during the training process of the enhanced CycleGAN using these databases. As shown in Figure 12 In addition, Figure 13 shows the training loss graph for the age estimator, based on using the images generated by the enhanced CycleGAN as input.…”