The Wavelet Chip Project
The first set of papers relates to the Wavechip project, an attempt
to implement the continuous wavelet transform (CWT) in silicon, using analog
VLSI technology. The CWT is implemented as a filter bank having a gaussian
kernel on each channel, and center frequencies spaced logarithmically in
frequency space. The output of each channel is sampled in proportion to
the bandwidth of the channel, making this a multiresolution scheme that
breaks incoming signals into a compact description in both time and frequency.
For time-varying signals, this description is more accurate than an FFT
decomposition, and like the discrete wavelet transform (DWT) it should be
applicable to data compression.
Three papers (having a great deal of overlap) detail the creation of the
chip through several chip implementations and also the ICNN '93 conference
in San Francisco (Int'l Conference on Neural Networks). These papers are
listed below:
- WaveChip version 1 describes the first chip
implementation to be fabricated.
- WaveChip version 2 describes the second chip
implementation to be fabricated, including on-board sampling.
- WaveChip version 3 describes the third chip
implementation to be fabricated, with a number of improvements over
the previous designs but without the on-board sampling.
The wavechip project continued with a fourth iteration based on mixed
analog and digital technology. One paper was presented at the NIPS '95
conference (in Colorado, in November) (available as
wave6nips.ps),
and another was published in
Express Letters of IEEE Circuits and Systems, August, 1996:
``Oversampling Architecture for Analog Harmonic
Modulation'' by R. Timothy Edwards and Gert Cauwenberghs. This
chip has been designed and fabricated, and the ingenious modulation
scheme works beautifully.
Discrete Wavelet Transform Processor
This project, completed during the summer of 1991, is an implementation
of the 2 dimensional discrete wavelet transform (DWT) on a commercially
available DSP. The transform is computed in parallel and scheduling is
performed which yields the minimum latency between the input and the
transformed output.
Results of this project encouraged Dr. Michael Godfrey and me to start
designing a continuous wavelet transform implementation.
The paper is found here: Wavelet paper.
Back to my home page. . .
email:
|
|
Last updated: April 9, 2012 at 2:21pm