Using Generative Adversarial Networks to Construct Different Age Groups of Oriental Faces

• Main Authors: HUANG, YU-MIN, 黃鈺珉
• Other Authors: YANG, SHIH-HSUAN
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/9az294

碩士 === 國立臺北科技大學 === 資訊工程系 === 107 === The aesthetic medicine industries are booming in recent years. However, current research on face imaging pays more attention to the western faces and few studies specifically aim for oriental faces. This thesis develops techniques that automatically generate face images of different age ranges for oriental faces, using the Conditional Generative Adversarial Networks (CGAN) model. Three major improvements against the conventional CGAN are proposed. First, the training data set is replaced with age-labeled oriental faces with adequate pre-processing. Second, the CGAN model is modified for handling both aging and rejuvenation to accommodate the need of medical aesthetics. Third, the CGAN parameters of age classifiers are adjusted to manifest the age characteristics of oriental faces. Experimental results show that the proposed method can more effectively generate oriental faces of different ages than the conventional CGAN. Judging by the Microsoft Face API, the proposed method produces images closer to the target age and gaining higher recognition similarity.