Fig. 9: Computational performance comparison of Quasar-based and alternative available implementations. | Nature Communications

Fig. 9: Computational performance comparison of Quasar-based and alternative available implementations.

From: An interactive ImageJ plugin for semi-automated image denoising in electron microscopy

Fig. 9

Absolute computational performance (in milliseconds) of (a) bilateral filtering41, (b) anisotropic diffusion38, (c) BLS-GSM14, and (d) nonlocal means denoising16 for different input sizes. A comparison is made between our proposed GPU-based Quasar framework and alternative implementations that are available to the scientific community: bilateral filter (ImageJ52 and MATLAB53 based), anisotropic diffusion (ImageJ54 and MATLAB55 based), BLS-GSM (MATLAB56) and nonlocal means denoising (ImageJ57 and MATLAB58 based). For each algorithm, we consider inputs of 2562, 5122, 10242, 20482, and 40962 pixels. In general, our Quasar implementation performs 10 to 100 times faster compared with the existing software packages. Notice that the obtained speedup increases as the input image size increases due to the fact that GPUs are able to process more pixels in parallel.

Back to article page