Image classifier learning from noisy labels via generalized graph smoothness priors

When collecting samples via crowd-sourcing for semi-supervised learning, often labels that designate events of interest are assigned unreliably, resulting in label noise. In this paper, we propose a robust method for graph-based image classifier learning given noisy labels, leveraging on recent advances in graph signal processing. In particular, we formulate a graph-signal restoration problem, where the objective includes a fidelity term to minimize the lo-norm between the observed labels and a reconstructed graph-signal, and generalized graph smoothness priors, where we assume that the reconstructed signal and its gradient are both smooth with respect to a graph. The optimization problem can be efficiently solved via an iterative reweighted least square (IRLS) algorithm. Simulation results show that for two image datasets with varying amounts of label noise, our proposed algorithm outperforms both regular SVM and a noisy-label learning approach in the literature noticeably.