Compressed Sensing Hardware Implementations (part deux)

[Update Nov. '08: Please find the Compressed Sensing Hardware page here ]

In a previous entry, I mentioned the most recent Compressed Sensing Hardware implementations I knew about. According to the recent presentation by Justin Romberg at the Short Course on Sparse Representations and High Dimensional Geometry at IPAM, it looks as though I missed two other implementations as described by Romberg in his talk:

* a Hyperspectral imager at Yale where algorithms of Ronald Coifman and Mauro Maggioni are being used against data gathered from a Texas Instrument DMD and a tuned laser light. You can see in the abstract of the article where a feature recognition is doing better when using random projection for hyperspectral imagery:

This leads us to ask how discriminatory spectral features should be selected. The features in previous work on cancer spectroscopy have been chosen according to heuristics. We use the “best basis” algorithm to select a Haar wavelet packet basis which is optimal for the discrimination task at hand. These provide interpretable spectral features consisting of contiguous wavelength bands. However they are outperformed by features which use information from all parts of the spectrum, combined linearly at random.

Hyperspectral imagery has a lot of potential because most of the data gathered is redundant in the first place. As far I can tell, in this example, the large dataset is compressed using random projections and they are then using machine learning techniques (diffusion methods/maps) in order to figure out the neighborhood of a sample given labeled training sets. More information can be found in the presentation of Coifman at IMA.

* an Analog imager at Georgia Tech but I could not find a trace of it for the moment.

The analog route is a fascinating one for two reasons:

• one can use random projections but one can also use other bases. It used to be that you could find analog sensor that were decomposing data directly in their fourier spectrum.
  
  
• analog used to be the way to go before the digital revolution and so there is a tremendous knowledge in acquiring analog signals. Compressed sensing opens the door to using these systems again.