ChildGAN: Face aging and rejuvenation to find missing children. (September 2022)

• Record Type: Journal Article
• Authors: Chandaliya, Praveen Kumar; Nain, Neeta
• Is Part Of: Pattern recognition. Volume 129(2022)
• Publication Date: 2022-09
• Subjects: Child face aging and rejuvenation -- Child datasets -- Face recognition -- Age estimation -- Gender preservation -- Child trafficking; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108761
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We present a large-scale longitudinal Indian child dataset (ICD), to address face aging, cross-age face recognition, age estimation, and gender preservation in children. ChildGAN, a new method for age-progression and rejuvenation, which can automatically generate visually realistic face photos, while attaining enhanced face-recognition, age-estimation, and gender-preservation rates, compared to state-of-the-art methods. We also investigate the generalization of ChildGAN by presenting Multi-Racial Child Dataset (MRCD) containing 64, 965 face images of four races (Asian, Black, White, and Indian). An encoder and discriminator architecture, inspired by the self-attention GAN, which exhibits a better balance between the ability to model long-range dependencies and computational and statistical efficiency. The model and the MRCD web crawled images are available at: https://github.com/praveenkumarchandaliya/ChildGAN_Tamp1. Abstract: Child-face aging and rejuvenation have amassed considerable active research interest, owing to their immense impact on a broad range of social and security applications, e.g., digital entertainment, fashion and wellness, and searching for long-lost children using childhood photos. All current face aging approaches based on generative adversarial networks (GANs) focus on adult images or long-term aging. We present a new large-scale longitudinal Indian child (ICD) benchmark dataset to facilitate face age progression and regression, cross-age face recognition, age estimation, gender prediction, and kinship face recognition to alleviate these issues. Furthermore, we propose an automatic child-face age progression and regression model, namely, ChildGAN, that generates visually realistic images for enhanced face-identification accuracy while preserving the identity. Consequently, we have trained state-of-the-art (SOTA) face aging models on ICD for comprehensive qualitative and quantitative evaluations. We also present a multi-racial experiments dataset named Multi-Racial Child Dataset (MRCD) containing 64, 965 child face images. The images are selected from publicly available datasets and web crawling. Finally, we investigate the generalization of ChildGAN by experimenting with White, Black, Asian, and Indian races. The experimental results suggest that the proposed ChildGAN and SOTA models can aid in reconnecting young children, who were lost at a young age as victims of child trafficking or abduction, with their families. The model and the MRCD web crawled images are available at https://github.com/praveenkumarchandaliya/ChildGAN_Tamp1/ .