Approximate compressed sensing for hardware-efficient image compression

Abstract

Recently, compressive sensing has attracted a lot of research interest due to its potential for realizing lightweight image compression solutions. Approximate or inexact computing on the other hand has been successfully applied to lower the complexity of hardware architectures for applications where a certain amount of performance degradation is acceptable (e.g. lossy image compression). In our work, we present a novel method for compressive sensing using approximate computing paradigm, in order to realize a hardware-efficient image compression architecture. We adopt Gaussian Random matrix based compression in our work. Library based pruning is used to realize the approximate compression architecture. Further we present a multi-objective optimization method to fine tune our pruning and increase performance of architecture. When compared to the baseline architecture that uses regular multipliers on 65-nm CMOS technology, our proposed image compression architecture achieves 43% area and 54% power savings with minimal PSNR degradation.