Robust Multi-View Prototype Learning

Abstract

In vision and machine learning, information fusion from multiple sensors can be regarded as multi-view learning paradigm to make use of the pairwise complementary information. Due to disturbed variances by illumination, equipment and environment, the collected data is frequently smeared by noises. Although there have been outlier-against works proposed, most of them suffer from redundant parameters. Even worse, few of them have modeled the cross-view complementary information. Actually, in multi-view learning, not only the clean data points, but also those outliers share the same prototypes in common space, by contrast, the margin value of the latter is non-zero. Motivated by this, in this article, we construct a robust multi-view prototype (RMVP) learning model with auxiliary margin matrices. Specifically, in RMVP, the projection matrices corresponding to each view span the common space to characterize the prototypes, simultaneously. In addition, an auxiliary sparse margin matrix for corresponding view identifies the distance gaps of outliers, whose non-zero column vector represents the margin value. To improve the robustness, the lasso regularization is penalized on the auxiliary matrices. And, an alternating optimization algorithm is presented to solve the proposed model. Finally, extensive experiments are constructed to testify the effectiveness of the proposed method.