CIF: Small: Beyond Sparsity - Exploiting Saliency in Compressive and Adaptive Sensing

The presence of salient features or anomalous behavior is a characteristic shared in many modern application domains. For example, an anomalous region in an MRI image could indicate the presence of a tumor or lesion; accurately locating conspicuous features in perhaps very high-dimensional images is a critical step in automated surveillance; and identification of anomalies in network traffic is a crucial step in detecting various attacks. Each of these applications is illustrative of an underlying theme, where subsets of data or image sub-regions are deemed interesting based on their relationship with (or, more precisely, their deviation from) typical behavior exhibited by the bulk of the data. In this sense, saliency can be understood as a natural generalization of the notion of sparsity, but one that is intrinsic to the data itself.

Recent developments in compressive and adaptive sensing have demonstrated that tremendous improvements in sensing resource efficiency can be realized when inferring high-dimensional data or objects that possess simple, low-dimensional representations. This research develops new theory and methods extending resource-efficient compressive and adaptive sensing techniques, which exploit sparsity as a model for data parsimony, to procedures that exploit saliency as a low-dimensional model for certain high-dimensional data. In particular, this effort (1) advances the current state-of-the-art theory and methods in compressive and adaptive sensing by developing a novel set of efficient saliency-based sensing methods, (2) demonstrates the robustness of these methods to uncertainties and noise, (3) integrates these new developments into the curriculum at the University of Minnesota, and (4) leverages emerging mobile device technologies as a novel vehicle for demonstrating and broadly disseminating the results of this effort to potentially new and diverse audiences.