Face age estimation using wrinkle patterns

• Main Author: Ng, Choon Ching
• Published: Manchester Metropolitan University 2015
• Subjects: 006.3
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.680215

Face age estimation is a challenging problem due to the variation of craniofacial growth, skin texture, gender and race. With recent growth in face age estimation research, wrinkles received attention from a number of research, as it is generally perceived as aging feature and soft biometric for person identification. In a face image, wrinkle is a discontinuous and arbitrary line pattern that varies in different face regions and subjects. Existing wrinkle detection algorithms and wrinkle-based features are not robust for face age estimation. They are either weakly represented or not validated against the ground truth. The primary aim of this thesis is to develop a robust wrinkle detection method and construct novel wrinkle-based methods for face age estimation. First, Hybrid Hessian Filter (HHF) is proposed to segment the wrinkles using the directional gradient and a ridge-valley Gaussian kernel. Second, Hessian Line Tracking (HLT) is proposed for wrinkle detection by exploring the wrinkle connectivity of surrounding pixels using a cross-sectional profile. Experimental results showed that HLT outperforms other wrinkle detection algorithms with an accuracy of 84% and 79% on the datasets of FORERUS and FORERET while HHF achieves 77% and 49%, respectively. Third, Multi-scale Wrinkle Patterns (MWP) is proposed as a novel feature representation for face age estimation using the wrinkle location, intensity and density. Fourth, Hybrid Aging Patterns (HAP) is proposed as a hybrid pattern for face age estimation using Facial Appearance Model (FAM) and MWP. Fifth, Multi-layer Age Regression (MAR) is proposed as a hierarchical model in complementary of FAM and MWP for face age estimation. For performance assessment of age estimation, four datasets namely FGNET, MORPH, FERET and PAL with different age ranges and sample sizes are used as benchmarks. Results showed that MAR achieves the lowest Mean Absolute Error (MAE) of 3.00 ( 4.14) on FERET and HAP scores a comparable MAE of 3.02 ( 2.92) as state of the art. In conclusion, wrinkles are important features and the uniqueness of this pattern should be considered in developing a robust model for face age estimation.