$k^{3}$-Sparse Graph Convolutional Networks for Face Recognition

In recent years, deep learning networks have substantially improved the performance of face recognition. Although deep learning networks have been very successful, there are limited to underlying Euclidean structure data. When dealing with complex signals such as medical imaging, genetics, social networks and computer vision, recently there has been a growing interest in trying to apply learning on non-Euclidean geometric data. Graph convolutional networks are a new deep learning architecture for analyzing non-Euclidean geometric data. In computer vision, a human face image is modeled as a graph in the irregular domain. A major technical challenge is how to optimize the structured face graph. Because, classification performance critically depends on the quality of the graph. In this paper, we explore an undirected graph convolutional network called $k^{3}$-SGCNs ($k^{3}$ - sparse graph convolutional networks). The main idea is to use sparsity-constrained optimization that obtain connected sparse subgraphs. A sparse graph of face image is composed of connected sparse subgraphs. Experiments demonstrate that the learned sparse graph has better performance than mutual $k$-nearest neighbor graph and $l1$ graph.