Age estimation using local direction and moment pattern (LDMP) features

An automatic estimation of age from face images is gaining attention due to its interesting applications such as age-based access control, customer profiling for targeted advertisements and video surveillance. However, age estimation from a face image is challenging due to complex interpersonal biological aging process, incomplete databases and dependency of facial aging on extrinsic and intrinsic factors. The published literature on age estimation utilizes multiple existing feature descriptors and then combines them into a hybrid feature vector. There is still an absence of specially designed aging feature descriptor which encodes facial aging cues. To address this issue we propose aging feature descriptor; Local Direction and Moment Pattern (LDMP), which capture directional and textural variations due to aging. We encode the orientation information available in eight unique directions. The texture is embedded into the magnitudes of higher order moments which we extract using local Tchebichef moments. Next, orientation and texture information is combined into a robust feature descriptor. To learn the age estimator, we apply warped Gaussian process regression on the proposed feature vector. Experimental analysis demonstrates the effectiveness of the proposed method on two large databases FG-NET and MORPH-II.