Hierarchical sparse representation based Multi-Instance Semi-Supervised Learning with application to image categorization

Recent studies have shown that sparse representation (SR) can deal well with many computer vision problems. In this paper, we extend a hierarchical sparse representation algorithm into Multi-Instance Semi-Supervised Learning (MISSL) problem. Specifically, at the instance level, after investigating the properties of true positive instances in depth, we propose a novel instance disambiguation strategy based on sparse representation that can identify the instance confidence value in both positive and unlabeled bags more effectively. At the bag level, in contrast to the traditional k-NN or @e-graph construction methods used in the graph-based semi-supervised learning settings, we propose a weighted multi-instance kernel and a corresponding kernel sparse representation method for robust @?^1-graph construction. The improved @?^1-graph that encodes the multi-instance properties can be utilized in the manifold regularization framework for the label propagation. Experimental results on different image data sets have demonstrated that the proposed algorithm outperforms existing multi-instance learning (MIL) algorithms, as well as the MISSL algorithms with the application to image categorization task.