Object Class Invariant Model for Age Estimation and Gender Classification in Pose Variant and Occluded Faces

• Main Authors: Wen-Lung Liu, 劉文龍
• Other Authors: Wei-Ta Chu
• Format: Others
• Language: en_US
• Published: 2010
• Online Access: http://ndltd.ncl.edu.tw/handle/19565113766890336909

碩士 === 國立中正大學 === 資訊工程所 === 98 === Many face applications have been developed. For example, a cigarette vendor machine should allow only adults to purchase. A surveillance system in a female dorm warns if males try to enter. In previous researches, global features are used for age estimation and gender classification. Hence, the analysis results are often wrong for faces with different head poses or occlusions caused by sun-glasses and hat. To deal with these problems, we exploit the object class invariant (OCI) model for age estimation and gender classification. The OCI model consists of a set of local features (scale-invariant features are adopted in this thesis). With the OCI model, we first localize faces from images captured in arbitrary views, and then determine the most distinctive features on faces. Relationships between distinctive features and the invariant vector are described based on geometry and appearance information, in the form of a probabilistic model. We demonstrate that these distinct features convey age/gender information. Face localization (i.e. finding the OCI), age estimation, and gender classification can be integrated into the same framework. In age estimation experiments, the method of active appearance model (AAM) combined with a multi-layer perceptrons classifier is compared with the OCI approach. The method proposed by Aghajanian et al. is compared with our method in gender classification experiments. We verify that performance of the OCI approach is promising both in age estimation and gender classification.