LRTI: landmark ratios with task importance toward accurate age estimation using deep neural networks

• Published in: Neural computing & applications 2022-06, Vol.34 (12), p.9647-9659
• Main Authors: Badr, Marwa M., Elbasiony, Reda M., Sarhan, Amany M.
• Format: Article
• Language: English
• Subjects: Age; Artificial Intelligence; Artificial neural networks; Chronology; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Continuity (mathematics); Data Mining and Knowledge Discovery; Feature extraction; Image Processing and Computer Vision; Labels; Machine learning; Neural networks; Original Article; Probability and Statistics in Computer Science
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-022-06955-6

Nowadays, age estimation systems have become a pressing need in several vital fields such as security services and health systems. Over the past decade, there have been introduced several efforts to build accurate and robust age estimation systems, where deep networks have proved to be the superior leader of machine learning tools. From this point, we propose a system named landmark ratios with task importance (LRTI), which accurately estimates a person’s age using deep neural networks. The proposed system extracts more precise information using the facial landmarks—rather than using only the extracted features inferred by the convolutional neural network —to estimate the age. The proposed system is based on defining the purposeful characteristics that distinguish the different age classes. As a result, LRTI computes the ratio of distances between the facial landmarks to represent the facial stretching through aging. These distance ratios are added to the network to precisely differentiate the age classes. The proposed system takes into account the in-between relation of age labels which enhance the accuracy of the age estimation process. The in-between relation of age labels is addressed by generating an importance vector, which gives a weight for each class label according to the degree of neighborhood to the target label. From the conducted experiments, the LRTI system adequately models the ordering and continuity properties of the aging process; thus, it has outperformed other state-of-the-art approaches when applied onto MORPH II, FGNET, CACD, AFAD, and UTKFace datasets. LRTI achieved the best mean absolute error, reaching 2.58 with MORPH II, 2.51 with FGNET, 5.39 with CACD, 3.44 with AFAD, and 5.14 with UTKFace.