Özet:

Abstract In computer vision and pattern recognition, estimating age from a single Image of human face is a crucial but challenging job. The quantity of training data gathered has a direct impact on how well a learning algorithm performs. Research in this area is primarily concentrated on enhancing results by training and testing using a single dataset. Despite the high accuracy results on this task using recent deep learning approaches, due to the diversity of human characteristics such as race and nationality, and variations in capture circumstances, these approaches lack generality when applied to unseen Images. Poor image quality, insufficient image counts, and low precision data limit the effectiveness of current learning methods. In this process of age prediction using Mean Absolute Error (MAE), we adopted CNNs with VGG-16, Resnet-50 and DenseNet-201 architectures to estimate age of a person by treating it as classification problem. As part of this investigation, we extensively analyzed the UTK Face, FGNET, CACD, and AS-23 datasets. In the last stage, merged dataset is cross validated with an additional dataset that have not been investigated before. As a result of this process, it is discovered that VGG-16 has the best accuracy with an MAE value of 2.2 with cross data learning, whereas the MAE for the Merged dataset associated with VGG-16 was1.71. The MAE values achieved with VGG-16 model are the best among all the experimental values for estimating age. In comparison to training on an independent dataset, the results demonstrate that multi-dataset simultaneously training network results in a more notable performance. The proposed method, according to experimental findings, demonstrated that the it outperforms almost all previous methods on age estimation with an MAE of 1.71 years.

Anahtar Kelimeler: