Canny-GPU

CUDA implementation of Canny edge detector in C/C++.
You can use cmake to compile the files. I have made a CMakeLists available for compilation.

Run the code

I have made available a main file that executes the code.
In particular, these are the parameters to be given on the command line:

./main argv[1] argv[2] argv[3] argv[4] argv[5] argv[6]

where :

• argv[1] : input image path
• argv[2] : kernel size of Sobel
• argv[3] : low threshold for Hysteresis step
• argv[4] : high threshold for Hysteresis step
• argv[5] : L2 norm -> 0 activated 1 deactivated (uses approximation with abs)
• argv[6] : modes -> [0] CPU , [1] GPU custom (my implementation) , [2] Runs all modes. With [0] run OpenCV Canny CPU while with [1] run Opencv GPU. At last, with [2] run both.

During the execution of the algorithm, the execution times are also calculated, expressed in ms.

Results example

Examples of image output of my Canny GPU version.

Original	Canny GPU Output

Original	Canny GPU Output

Original	Canny GPU Output

N.B: obviously, the results may vary according to the value chosen for the thresholds in the hysteresis step.

Kernel config

I tried several kernel configurations but the one that gave the best results was the one where I used a thread block size of 16x16.

Kernel Configuration

Kernel time esecution

This is the pie chart showing the execution times of the various kernel device function and data transfer memcpy routines on 720p image resolution.

Kernel time esec

CPU v.s. GPU

This is the comparison analysis between the OpenCV CPU version and my parallel version on GPU.

CPU v.s. GPU

As you can see from the graph, with a low resolution image the results of the two versions are similar. As the image resolution increases, the parallel version gets significantly better results.