Inexact Implementation of Wavelet Transform and Its Performance Evaluation Through Bit Width Reduction

Abstract

Resource and energy optimization in computing is gaining a lot of importance due to the increasing demand of smart and portable devices. These devices have a stiff budget in terms of resource and energy. Most of the applications running in these devices are media intensive and hence special efforts are needed to minimize the resource and energy requirements for the various computational tasks involved in media processing. Discrete wavelet transform (DWT) is an important transform, which is utilized in various forms of image and video processing applications. It is a complex transform and hence demands a direct hardware implementation instead of software execution in many application scenarios, to increase the overall system throughput. Inexact computing sacrifices the precision of computing accuracy by rejecting one or few bits of data storage. The inexactness in computing does not hamper those applications whose quality is not much compromised due to such inaccuracy. In this paper, we propose a low-resource and energy-aware hardware design for DWT through dynamic bit width adaptation, thus performing the computation in an inexact way. We have performed field programmable gate array (FPGA) based prototype hardware implementation of the proposed design. To the best of our knowledge this is a first of its kind of modeling of DWT involving inexact computing.