General structured sparse learning for human facial age estimation

Human facial age estimation has attracted much attention in the field of computer vision in recent decades, due to its potential applications in human-computer interaction and electronic customer relationship management. However, the insufficiency of labeled facial images has made age estimation a challenging problem. Being aware of this, we focus on learning the ordinal relationships among different age labels from limited training samples. To exploit the intrinsic property of ordinal relationships, the learning problem is formulated as a structured sparse multi-class classification model regularized by the structured sparsity. The structured sparse regularization encodes the ordinal relationships among different age labels, and ensures that the samples with similar age labels are close to each other in the feature space. In addition, by aggregating the prediction functions for different age labels into a matrix W, we introduce the grouped sparsity which constrains the matrix W to share a common subspace. The smoothing proximal gradient (SPG) method is employed to optimize the proposed model since the objective function is convex and non-smooth with respect to variable W. Additionally, we conduct experiments on two publicly available facial image datasets, the FG-NET and the MORPH-II, and the empirical results have demonstrated the effectiveness of the proposed method against the state-of-the-arts.