Canny Edge Detection on GPU using CUDA

Edge detection is a crucial step in many of today's computer vision applications. Canny edge detection in particular involves several steps to achieve real-time results. Many systems currently deployed leverage the compute capability that a graphics processing unit (GPU) can achieve. This paper covers the implementation and testing of a Canny edge detection algorithm using CUDA C. The results cover a comparison of the naive implementation in sequential C, a parallelized implementation using OneAPI Threading Building Blocks (TBB), and a tiled, shared memory approach using CUDA C. A comparison between the NVIDIA GTX 1060 and NVIDIA RTX 3090 are also performed. The CUDA C implementation shows an improvement of up to 100 times that over the naive sequential implementation for an RGB image at 4k resolution, and an improvement of 10 times when compared to the TBB approach. Additionally, the RTX 3090 showed roughly a speed up of 1.5 times that of the GTX 1060, demonstrating the advances made between the generations of GPUs. These results overall show the benefits of using a GPU accelerated approach to edge detection, with further improvements left to achieve.