high performance computing on graphics processing units: hgpu.org

The discrete wavelet transform (DWT) has been extensively used for image compression and denoising in the areas of image processing and computer vision. However, the intensive computation of DWT due to its inherent multilevel data decomposition and reconstruction operations brings a bottleneck that drastically reduces its performance and implementations for real-time applications when facing large size digital images and/or high-definition videos. Although various software-based acceleration solutions, such as the lifting scheme, have been devised and achieved a higher performance in general, the pure software accelerated DWT still struggle to cope with the demands from real-time and interactive applications. With the growing capacity and popularity of graphics hardware, personal computers (PCs) nowadays are often equipped with programmable graphics processing units (GPUs) for graphics acceleration. The GPU offers a cost-effective parallel data processing mechanism for operations on large amount of data, even for applications beyond graphics. This practice is commonly referred as general-purpose computing on GPU (GPGPU). This paper presented a GPGPU framework with the corresponding parallel computing solution for wavelet-based image denoising by using off-the-shelf consumer-grade programmable GPUs. This framework can be readily incorporated with different forms of DWT by customizing the parameter of the wavelet kernel. Experiment results show that the framework gains applicability in data parallelism and satisfaction performance in accelerating computations for wavelet-based denoising.