FPGA implementations for parallel multidimensional filtering algorithms

Abstract

One and multi dimensional raw data collections introduce noise and artifacts, which need to be recovered from degradations by an automated filtering system before, further machine analysis. The need for automating wide-ranged filtering applications necessitates the design of generic filtering architectures, together with the development of multidimensional and extensive convolution operators. Consequently, the aim of this thesis is to investigate the problem of automated construction of a generic parallel filtering system. Serving this goal, performance-efficient FPGA implementation architectures are developed to realize parallel one/multi-dimensional filtering algorithms. The proposed generic architectures provide a mechanism for fast FPGA prototyping of high performance computations to obtain efficiently implemented performance indices of area, speed, dynamic power, throughput and computation rates, as a complete package. These parallel filtering algorithms and their automated generic architectures tackle the major bottlenecks and limitations of existing multiprocessor systems in wordlength, input data segmentation, boundary conditions as well as inter-processor communications, in order to support high data throughput real-time applications of low-power architectures using a Xilinx Virtex-6 FPGA board. For one-dimensional raw signal filtering case, mathematical model and architectural development of the generalized parallel 1-D filtering algorithms are presented using the 1-D block filtering method. Five generic architectures are implemented on a Virtex-6 ML605 board, evaluated and compared. A complete set of results on area, speed, power, throughput and computation rates are obtained and discussed as performance indices for the 1-D convolution architectures. A successful application of parallel 1-D cross-correlation is demonstrated. For two dimensional greyscale/colour image processing cases, new parallel 2-D/3-D filtering algorithms are presented and mathematically modelled using input decimation and output image reconstruction by interpolation. Ten generic architectures are implemented on the Virtex-6 ML605 board, evaluated and compared. Key results on area, speed, power, throughput and computation rate are obtained and discussed as performance indices for the 2-D convolution architectures. 2-D image reconfigurable processors are developed and implemented using single, dual and quad MAC FIR units. 3-D Colour image processors are devised to act as 3-D colour filtering engines. A 2-D cross-correlator parallel engine is successfully developed as a parallel 2-D matched filtering algorithm for locating any MRI slice within a MRI data stack library. Twelve 3-D MRI filtering operators are plugged in and adapted to be suitable for biomedical imaging, including 3-D edge operators and 3-D noise smoothing operators. Since three dimensional greyscale/colour volumetric image applications are computationally intensive, a new parallel 3-D/4-D filtering algorithm is presented and mathematically modelled using volumetric data image segmentation by decimation and output reconstruction by interpolation, after simultaneously and independently performing 3-D filtering. Eight generic architectures are developed and implemented on the Virtex-6 board, including 3-D spatial and FFT convolution architectures. Fourteen 3-D MRI filtering operators are plugged and adapted for this particular biomedical imaging application, including 3-D edge operators and 3-D noise smoothing operators. Three successful applications are presented in 4-D colour MRI (fMRI) filtering processors, k-space MRI volume data filter and 3-D cross-correlator.