POOKA is a Reconfigurable Computer based system suitable for solving pixel-level classification problems in multi-spectral image data sets. The system is first optimized for a particular feature of interest and particular data set. It can then be applied to large image databases with two orders of magnitude greater throughput compared to software. The system is most suitable for solving large scale broad area search problems, to characterize land cover, terrain type etc.

The system has been implemented with the Firebird Reconfigurable Computer from Annapolis Microsystems. This is a 64-bit PCI card that contains a Virtex 2000E FPGA and a total of 40 Mbytes of on-board memory distributed in 5 independent banks.

The Reconfigurable Computer implements a massively parallel-shared weight network like the one below:

Each layer is parameterized and can programmed to execute many different image processing operations. The picture below shows some example configurations for edge detection and texture measures.

To solve classification problems, we optimize the network parameters, or program the POOKA machine using Evolutionary Algorithms. The FPGA serves two purposes:

1). To accelerate computationally intensive Evolutionary Algorithms. The combination of Evolutionary Algorithms, and FPGA acceleration, means we can solve more difficult problems than is traditionally attempted. For example, we can include the performance constraints required in a deployed platform.

2). After optimization, to provide orders of magnitude greater throughput in application.

A more detailed description of the POOKA system can be found in publications.

Check out our first experiments with video.

Here's how it all fits in your desktop!