Abstract

The discrete wavelet transform (DWT) has been widely used in many fields, such as image compression, speech analysis and pattern recognition, because of its capability of decomposing a signal at multiple resolution levels. Due to the intensive computations involved with this transform, the design of efficient VLSI architectures for a fast computation of the transforms have become essential, especially for real-time applications and those requiring processing of high-speed data. The objective of this thesis is to develop a scheme for the design of hardware resource-efficient high-speed pipeline architectures for the computation of the DWT. The goal of high speed is achieved by maximizing the operating frequency and minimizing the number of clock cycles required for the DWT computation with little or no overhead on the hardware resources. In this thesis, an attempt is made to reach this goal by enhancing the inter-stage and intra-stage parallelisms through a systematic exploitation of the characteristics inherent in discrete wavelet transforms.
In order to enhance the inter-stage parallelism, a study is undertaken for determining the number of pipeline stages required for the DWT computation so as to synchronize their operations and utilize their hardware resources efficiently. This is achieved by optimally distributing the computational load associated with the various resolution levels to an optimum number of stages of the pipeline. This study has determined that employment of two pipeline stages with the first one performing the task of the first resolution level and the second one that of all the other resolution levels of the 1-D DWT computation, and employment of three pipeline stages with the first and second ones performing the tasks of the first and second resolution levels and the third one performing that of the remaining resolution levels of the 2-D DWT computation, are the optimum choices for the development of 1-D and 2-D pipeline architectures, respectively. The enhancement of the intra-stage parallelism is based on two main ideas. The first idea, which stems from the fact that in each consecutive resolution level the input data are decimated by a factor of two along each dimension, is to decompose the filtering operation into subtasks that can be performed in parallel by operating on even- and odd-numbered samples along each dimension of the data. It is shown that each subtask, which is essentially a set of multiply-accumulate operations, can be performed by employing a MAC-cell network consisting of a two-dimensional array of bit-wise adders. The second idea in enhancing the intra-stage parallelism is to maximally extend the bit-wise addition operations of this network horizontally through a suitable arrangement of bit-wise adders so as to minimize the delay of its critical path.
In order to validate the proposed scheme, design and implementation of two specific examples of pipeline architectures for the 1-D and 2-D DWT computations are considered. The simulation results show that the pipeline architectures designed using the proposed scheme are able to operate at high clock frequencies, and their performances, in terms of the processing speed and area-time product, are superior to those of the architectures designed based on other schemes and utilizing similar or higher amount of hardware resources. Finally, the two pipeline architectures designed using the proposed scheme are implemented in FPGA. The test results of the FPGA implementations validate the feasibility and effectiveness of the proposed scheme for designing DWT pipeline architectures.