Comparative study between machine and deep learning methods for age, gender and ethnicity identification

Facial recognition is becoming increasingly used in real-world applications, such as video surveillance, human-computer interaction applications. One of the most popular applications is the automatic age, gender, and ethnicity (race) classification. The extraction of these facial attributes has an important role in social interactions this study aims to compare different machine and deep learning techniques and propose a novel deep learning architecture to automatically classify age, gender, and ethnicity from a person's face image. We used UTKFace dataset, each one is labeled with age, gender, and ethnicity. Machine learning methods did not give more than 58.04%, 86.25%, and 72.78% as test accuracy for age, gender, and race respectively. Transfer learning gave 63.53%, 89.14% and 72.39% as best results when our proposed CNN architectures outperform the previous results, they gave 65.92%, 90.3% and 78.88 % with a modified version of age prediction that gave 80.46 % using three classes (Child, Teenager, and Adult) instead of five.