Informatics Publishing Limited

Randa Khemiri ¹, Fatma Ezahra Sayadi ¹, Mohamed Atri ¹

Affiliations
1 Electronics and Microelectronics Laboratory, FSM, University of Monastir, Environment Street, Monastir – 5019, TN

Abstract

Background/Objectives: The implementation of the “Cohen-Daubechies-Feauveau9/7” lifting algorithm on a NVIDIA GPU in order to accelerate the computation with single and double-precision of some test images. Methods/Statistical Analysis: The DWT “Cohen-Daubechies-Feauveau 9/7” filter is implemented on a GPU with MatLab using the in-house parallel computation toolbox (PCT). The performance of both CPU and GPU implantations are compared with single and double precision on some test image. Findings: The investigational results show that the speedup is proportional to the image size until reaching a maximum at 40962 pixels for single-precision and 20482 pixels for double-precision; beyond these values the speedup decreases. The performance with GPU is enhanced, compared to that with CPU, by a factor above 2 for a single-precision of 40962 pixels image size and by a factor above 3 for a double-precision of 20482 pixels image size. By computing the Peak Signal-to-Noise Ratio (PSNR) at critical points after compression, we concluded that with single-precision we can compress larger image sizes without altering the image quality. Applications/Improvements: An efficient parallel implementation of the algorithm on GPU may lead to higher performances. The proposed idea in this work could be also extended to provide high efficiency video coding.

Keywords

Cohen-Daubechies-Feauveau 9/7, DWT (Discrete Wavelet Transformation), JPEG2000, MatLab-GPU, Single and Double Precision

Full Text