Here is a nice entry, as you all know I have a soft spot for hardware implementation. A recent one is a chip embedding the AMP reconstruction solver. Today, we have something on the acquisition side, woohoo!
It all started with Zainul Charbiwala sending me the following:
Hi Igor,I wanted to point to something new we've been working on. Its a low power CS hardware implementation using some analog tricks. We just submitted the camera ready version and I'd love to hear comments about it from the community at large.The schematics and code are all open for anyone else wanting to try it out.Thanks,Zainul.--Zainul CharbiwalaResearcher, IBM Research, India
Sure Zainul and here is the paper: CapMux: A Scalable Analog Front End for Low Power Compressed Sensing by Zainul M Charbiwala, Paul D Martin, Mani B Srivastava . The abstract reads:
Although many real-world signals are known to follow standard models, these signals are usually first sampled, rather wastefully, at the Nyquist rate and only then parametrized and compressed for efficient transport and analysis. Compressed sensing (CS) is a new technique that promises to directly produce a compressed version of a signal by projecting it to a lower dimensional but information preserving domain before the sampling process. Designing hardware to accomplish this projection, however, has remained problematic and while some hardware architectures do exist, they are either limited in signal model or scale poorly for low power implementations. In this paper, we design, implement and evaluate CapMux, a scalable hardware architecture for a compressed sensing analog front end. CapMux is low power and can handle arbitrary sparse and compressible signals, i.e. it is universal. The key idea behind CapMux’s scalability is time multiplexed access to a single shared signal processing chain that projects the signal onto a set of pseudo-random sparse binary basis functions. We demonstrate the performance of a proof-of-concept 16-channel CapMux implementation for signals sparse in the time, frequency and wavelet domains. This circuit consumes 20µA on average while providing over 30dB SNR recovery in most instances.
The attendant schematic is here.
Which makes me think I need to update the Compressive Sensing hardware page.
Liked this entry ? subscribe to Nuit Blanche's feed, there's more where that came from. You can also subscribe to Nuit Blanche by Email, explore the Big Picture in Compressive Sensing or the Matrix Factorization Jungle and join the conversations on compressive sensing, advanced matrix factorization and calibration issues on Linkedin.