Impact of Data Augmentation on Age Estimation Algorithms

This paper presents and analyzes algorithms for age estimation of the person in an image. Theoretical analysis and practical implementation of the experimental system were performed. The emphasis was on neural network-based algorithms. The impact of increasing the training set on the accuracy of the network was analyzed and different techniques of data augmentation in real-time and out of real-time were presented. Some simpler augmentation techniques were used to enlarge the training set in real-time, while augmentation out of real-time was made possible by merging several different age estimation datasets. In addition to merging existing datasets, the possibilities of increasing the training set with artificially generated images were also examined. For that purpose, a generative adversarial network was used, which can generate artificial images that belong to the appropriate age groups. At the end of the paper, the results on the test set were evaluated. It has been shown that increasing the training set achieves about 6% higher accuracy compared to the case without augmentation and about 3% higher accuracy compared to the results from the published papers.