A gausian blur is basically a convolution operation between an input image and a gaussian filter kernel. As such, it can be implemented in two ways. Either the convolution is computed in real space, by nested loops. Or you do a Fast Fourier Transform (FFT) to go to Fourier space, then a complex multiplication and an inverse FFT to go back to real space. Which one is faster depends on the size of the Gausian kernel. For larger kernels, the Fourier space implementation is faster. For very small kernels below a certain threshold, the overhead for the FFT does not pay off and the realspace implementation is faster. The exact threshold depends on the hardware: on CPUs it is typically lower, for GPUs and HPC platforms it tends to be higher.

A gausian blur is basically a convolution operation between an input image and a gaussian filter kernel. As such, it can be implemented in two ways. Either the convolution is computed in real space, by nested loops. Or you do a Fast Fourier Transform (FFT) to go to Fourier space, then a complex multiplication and an inverse FFT to go back to real space. Which one is faster depends on the size of the Gausian kernel. For larger kernels, the Fourier space implementation is faster. For very small kernels below a certain threshold, the overhead for the FFT does not pay off and the realspace implementation is faster. The exact threshold depends on the hardware: on CPUs it is typically lower, for GPUs and HPC platforms it tends to be higher.

Hi,

Python by itself is not fast but you can make good callbacks to the native C code. Numpy broadcasting is much faster than native Python. You can check some good examples from this link. The author uses both core-oriented methods and GPU to perform "Gaussian Blur":

http://nbviewer.ipython.org/github/mroberts3000/GpuComputing/blob/master/IPython/GaussianBlur.ipynb

Got some amazing speeds also.

Yes, without numpy or cython, you're never going to speed this code up enough.  I'd recommend just using an image processing library in python that's already established like mahotas or scikit-image instead of reinventing the wheel.

Wht don't you use 'gaussian_filter' in Scipy

http://docs.scipy.org/doc/scipy/reference/generated/scipy.ndimage.filters.gaussian_filter.html