Image Denoising Accuracy MATLAB vs. CUDA

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi all!

I’m doing a gauss filtering on image.

The steps of my goal are:

  1. read data from an image

  2. create a kernel

  3. applying FFT to image and kernel data

  4. pointwise multiplication

  5. applying IFFT to 4. results.

If I do the same operation in MATLAB I have best accuracy.

I have read in other topics, that I must normalize the IFFT results with 1/size_fft,

or use only Z2Z/C2C cufft type because matlab use double precision data type.

I have done it, but I haven’t improvements.

The only doubt come from the C2C use, my data are float and I convert them in Complex before call cufftExecC2C:

//for

image_complex[i].x = image[i].x;

image_complex[i].y = 0;

//end for

cufftExecC2C(plan, (cufftComplex *)image_complex, (cufftComplex *)image_complex, CUFFT_FORWARD);

and on IFFT I need unsigned char data from complex:

cufftExecC2C(plan, (cufftComplex *)image_complex, (cufftComplex *)image_complex, CUFFT_INVERSE);

//for

filtered[i]=(char)image_complex[i].x;

//end for

In the end I use filtered data for bitmap visualization.

thank you for support.

2

I’m curious if you’ve tried Jacket. The Jacket GPU version in MATLAB should be identical to the CPU version. Free 15-day trial is available here (to purchase is only a few hundred bucks): http://www.accelereyes.com

-John

3

I need to do Image Denoising in CUDA and compare results in execution time and accuracy with Matlab that use CPU.

So, I don’t know if Jacket is useful for my goal.

4

External Image

5

There are many people using a combination of Jacket’s core functions as well as Jacket’s Image Processing Library (http://wiki.accelereyes.com/wiki/index.php/Image_Processing_Library) for these kinds of applications. I think it would do what you need it to. Feel free to email me (see my signature below) if I can help any further.

6

I report this to my tutor, but it says that matlab/freemat and cudaFFT using FFTW both, so the results can’t be too different.

7

I have solved!

The problem was the padding!

Is it normal that for hight equal value of kernel radius and gauss delta,

is accuracy better in Matlab than CUDA ?