Age classification using Hierarchical Support Vector Machine based on characteristics of upper facial area

• Main Author: Dahlan, Hadi Affendy
• Format: Thesis
• Language: English
• Published: 2013
• Online Access: http://psasir.upm.edu.my/id/eprint/56153/1/FK%202013%20105RR.pdf; http://psasir.upm.edu.my/id/eprint/56153/

Facial aging classification is a growing research in pattern recognition area, where it can be used in many applications. Most of the digital image feature extractor needs the whole facial area to be used for the age classification. This however causes disadvantage to the people who may unable to show their full face because of a certain condition such as Muslim woman who wears ‘purdah’ to cover their ‘aurah’. Furthermore, only a few have performed feature extraction on the upper facial area, which an approach that may improve feature used and classification performance. Additionally, not many researchers have study the classification effect on different genders when using the features on the upper region. This study aimed to classify age that focused on wrinkle features at the Region of Interest (ROI) on the upper facial area using specific orientation of Gabor wavelet filter. The region is detected using a robust eye detection method. The Gabor wavelet filter is used for the wrinkle extractions together with the employed 2-step Hierarchical Support Vector Machines (SVM) as the classifier. The first step of the method classifies the sample between age groups 20-39 and above 40, while the second step classified it again into more specific age groups whether the sample is between 20-29 and 30-39 or 40-49 and above 50. The captured facial image of Malaysian database is used in this study, as to compare with the normal use of the Caucasian databases. For the Malaysian database using the hierarchical-SVM on the upper region, the best results obtained for the overall average accuracy and Mean Absolute Error (MAE) for the male were 59.34% and 0.5968 respectively when using the 3 upper regions in both stage; while the female obtained 59.08% and 0.5308 respectively when using the 5 upper regions for both stage. To gained better performance, modification were made in the final test, by combining the used of the full facial region in the first stage and the 3 upper region in the second stage of the hierarchical-SVM classification. The final results obtained were 62.96% for the male and 62.09% for the female in the overall average accuracy; while the MAE obtained for the male is 0.4573 and 0.4857 for the female. The result for the Malaysian database has shown that the full facial ROI usage does provide good age classification in the first step. However, for distinguishing specific age group in the second step, the full ROI will not be effective anymore. It is more appropriate to use the upper facial ROI to distinguish the more senior age. The main finding in this study suggests that age classification for different gender can be detected using upper facial wrinkle, which thus complements the biometric information field.