Structure-Based Learning in Sampling, Representation and Analysis for Multimedia Big Data

This paper presents disruptive insights and techniques on structure-based learning for multimedia big data. Along this viewpoint, significant technical challenges for multimedia big data are investigated, including sampling and reconstruction, representation, and analysis. For multi-dimensional compressive sampling, the union of data-driven subspace is addressed via subspace learning with structured sparsity. To enrich the correlated reconstruction, spatio-temporal regularity is presented within various multimedia data. Inspired by this insight, multi-scale dictionary learning is proposed to leverage spatio-temporal structures for sparse representation and make learning-based structured prediction and analysis.