A study on the prediction of Asian's Age and Gender using deep neural network

Abstract

Quick and accurate prediction of Asian Age and Gender from face images is a challenging task in real-world applications. Researchers working on this task usually face challenges such as lacking high-quality data for Asian faces, imbalanced age distributions, and computational complexity. Previous studies on this problem use age-group as the final output. This approach is not feasible in practice: (i) lack of helpful information for precise age prediction; (ii) inconsistency in the division of age groups. This thesis proposes a compact yet efficient model for Asian Real Age, Gender prediction using deep-learning-based, along with a new soft label representation of age mechanism to accelerate the accuracy of the final actual age label. Focal loss and Kullback-Leibler divergence were applied to confront the skew in age distribution. The model will be trained on well-known large-scale datasets such as IMDB/Wiki and fine-tuned with Asian face-specific datasets. The experiments on these datasets show that accuracy is at an acceptable rate while time complexity is feasible for practical applications.