Development of age estimation algorithm by determination and classification of wrinkle from face images

Aging resulting from the nature of living beings is defined as irrevocable changes in cell, tissue and systems with increasing age. With developing technology, many researchers try to find solutions to increase average lifetime of humans and decelerate aging in every way. Among studies in aging deceleration, the most popular one seems to be anti-wrinkle products which also increase humans' life quality. The upcoming analysis on wrinkles will also provide contribution with exploration of the efficiency of anti-wrinkle products. Faster and efficient treatment processes seem to be possible via adapting products to face age instead of body age. In the literature, there are pioneering works on age estimation by wrinkle determination and classification. However, there is a need for a standard system which determines wrinkles before and after clinical treatments automatically. In this thesis, a decision-support system on age estimation via wrinkle analysis on people's face area is developed. To observe the efficiencies of planned operations and medications on wrinkle treatment, a methodology is presented. The accuracy of acquired image processing and classification algorithms coming from this study is tested by MORPH, FG-Net database and new database generated throughout this thesis. Firstly, wrinkle segmentation is provided by applying image processing algorithms on available images. Then, Gabor filter values belonging to acquired wrinkle regions, Local Binary Pattern (LBP) histograms and feature vectors coming from wrinkle levels are constituted based on weighting of the relation of each region with aging. After that, the classification models generated by artificial neural networks, decision trees, support vector machines and k-nearest neighbour clustering methods are trained. Tests are done with four used individual classification models, a system giving information about people's age based on test images is designed. Furthermore, a fusion classifier which takes a vote among the accuracy rate of these individual classifiers. As a result of these experiments, the fusion classification method gives highly accurate outcomes comparing with individual classifiers.