Image Denoise Filters using Python, PyCuda and Cuda

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

for those who want to apply denoise filter to opencv images (using cv2.ImRead and therefore classical image format), i wrote some routines which could interest you.

The base routines are from Nvidia Cuda SDK example (many thanks Nvidia). I changed code to not use Tex2D and just use 3D arrays from cv2.imread (colour image) or 1D array for mono images.

There are 4 files :

  • KNN denoise filter for colour image
  • KNN denoise filter for monochrome image
  • Fast NLM denoise filter for colour image
  • Fast NLM denoise filter for monochrome image

For each program, i compare OpenCV routine (cv2.fastNlMeansDenoisingColored or cv2.fastNlMeansDenoising) to PyCuda/Cuda routine.

You can find the files here :

https://github.com/AlainPaillou/PyCuda_Denoise_Filters

I made some speed tests comparisons between OpenCV and PyCuda for colour images using a Jetson Nano :

1544*1040 pixels image :

  • OpenCV FastNLMDenoise : 0.92 seconds
  • fast NLM PyCuda : 0.17 seconds
  • KNN PyCuda : 0.068 seconds

3096*2080 pixels image :

  • OpenCV FastNLMDenoise : 2.6 seconds
  • fast NLM PyCuda : 0.67 seconds
  • KNN PyCuda : 0.25 seconds

That’s a really good improvement.

I would like to thanks again Nvidia for their work, their support and for having provided me for free a Jetson Nano to support my project (related to astronomy imaging).

If you find some improvements to those routines, i am very interested in.

Alain

2

Hello,

I have modified the KNN & NLM2 routines to avoid border effect error (array index out of range).

GitHub is up to date with those modifications.

I also tried those filters with a PC laptop i7-6700HQ CPU and GTX960M GPU under Windows 10 64b.

Here are the speed tests results :

1544*1040 pixels image :
Laptop PC :

  • OpenCV FastNLMDenoise : 0.32 seconds
  • fast NLM PyCuda : 0.033 seconds
  • KNN PyCuda : 0.025 seconds

Jetson Nano :

  • OpenCV FastNLMDenoise : 0.92 seconds
  • fast NLM PyCuda : 0.17 seconds
  • KNN PyCuda : 0.068 seconds

3096*2080 pixels image :
Laptop PC :

  • OpenCV FastNLMDenoise : 0.82 seconds
  • fast NLM PyCuda : 0.11 seconds
  • KNN PyCuda : 0.07 seconds

Jetson Nano :

  • OpenCV FastNLMDenoise : 2.6 seconds
  • fast NLM PyCuda : 0.67 seconds
  • KNN PyCuda : 0.25 seconds

Alain

3

hii can you please provide new link to code this one is not working

4

Looking at Alain Paillou’s GitHub account, it appears that he is still actively developing code, or at least he was as of March 2024, but at present all his repositories are private, none are public.

You could try to reach out to Mr. Paillou via GitHub.

5

Hello,

My GitHub repository in public now. You can access it.

Alain

1 Like
7

Thanks Alain