Hybrid Learning and Augmentation Techniques for Human Age Determination

• Main Authors: Yi-Ching Hsu, 許意晴
• Other Authors: Jian-Jiun Ding 丁建均
• Format: Others
• Language: en_US
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/9979hq

碩士 === 國立臺灣大學 === 電信工程學研究所 === 107 === Age play a critical role in our lives. It can help us solve many things in people’s life. It has wide range of application, such as unmanned store, monitoring,marketing, autonomous vehicle and so on. In marketing, it can analyze the age level of the purchased goods and help the store to increase the sales amount. The challenge of age estimation is that people have little change in face during middle age, so employing only deep learning methods is not enough. Therefore, we proposed an age identification system that combines many methods to improve the accuracy of the age. These methods combine deep learning and traditional techniques. In order to address the problem that face changes little in middle age, we classify the age and then make accurate estimates from each age range. Our proposed system is divided into three stages. The first step is pre-processing, including feature map、face detection and cropping. The second step is to classify ages into seven classes through deep learning and support vector machine (SVM) techniques. The third step is to train seven Deep Neural Networks (DNN) models, then predict the explicit age. The input for each model contains the probability of each age group and also increases the probability of each range by ten years. This probability is estimated by the previous step Deep Neural Networks (DNN). Increasing the features of the input can effectively reduce errors.