Enabling inference in the compressive domain
Abstract
Compressive sensing (CS), a recent signal acquisition paradigm, attempts to combat the data deluge problem by sensing only the relevant information. CS suggests that signals that exhibit redundant structures, such as sparse transform-domain coefficients, have significantly fewer degrees of freedom as compared to their dimensionality, and that it is possible to sense the signal with a number of measurements proportional only to their degrees of freedom. This has immense benefits in reducing the complexity of the sensor in that a high-dimensional scene can often be sensed from a suitably modified low-dimensional sensor. The main results of compressive sensing are directed towards providing novel sampling theorems that determine the feasibility of signal reconstruction from an under-determined set of linear measurements. However, reconstruction is often not the eventual goal in most applications which range from detection, tracking, recognition and classification. While all of this can be done post-reconstruction (on the output of a reconstruction procedure), there are important benefits to be gained by performing them directly on the compressive domain. First, tasks like detection, tracking, and recognition are inherently simpler than reconstruction --- and hence, there is hope that we can perform them with fewer measurements. Second, CS reconstruction is intrinsically tied to the signal models used for the unknown signal --- and these signal models prioritize features that deal with visual perception which often is not the most relevant for the subsequent processing tasks. Third, reconstruction algorithms associated with CS have high computational complexity --- and hence, avoiding a reconstruction step in the overall processing pipeline can be beneficial. In this proposal, a foundational theory of compressive inference is proposed by enabling local feature extraction from compressive measurements. The key proposition is the design of measurement matrices that enable local feature extraction by preserving correlations with filters of interest. The research objectives here are three-fold: first, the design of measurement operators that enable local feature extraction; second, the design of measurement operators that promote inference-specific objective criteria as opposed to reconstruction; and third, measurement operators that have specific structures for ease of optical implementation.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 30, 2018
- Source ID
- W911NF1510126
Entities
People
- Aswin C. Sankaranarayanan
Organizations
- Army Contracting Command
- Massachusetts Institute of Technology
- United States Army