Performance and error analysis of recursive edge-aware Gaussian filters on GPUs

  • Hermes H. Ferreira UFRGS
  • Eduardo S. L. Gastal UFRGS
  • Lucas Schnorr UFRGS
  • Philippe Navaux UFRGS

Resumo


We present a schematic for image edge-aware Gaussian GPU filtering which has linear complexity on the number of pixels of the image. It allows us to reduce the execution time as we increase the number of Streaming Multiprocessors (SMs) on the GPU. We make use of a domain transformation and use a complex-valued recursive formulation of the Gaussian filter. The algorithm partitions the image in disjoint regions, where we compute in parallel the filtering operations, avoiding communication between regions. Our implementation leads to a real-time solution using a modern GPU. With the RTX 2080 Ti, we achieved an execution time of less than 10 milliseconds for 2 filtering iterations on high-resolution RGB images of dimensions 2048x2048.
Palavras-chave: edge aware filtering, GPU processing, high performance computing, recursive filtering, image processing
Publicado
07/11/2020
FERREIRA, Hermes H.; GASTAL, Eduardo S. L.; SCHNORR, Lucas; NAVAUX, Philippe. Performance and error analysis of recursive edge-aware Gaussian filters on GPUs. In: CONFERENCE ON GRAPHICS, PATTERNS AND IMAGES (SIBGRAPI), 33. , 2020, Evento Online. Anais [...]. Porto Alegre: Sociedade Brasileira de Computação, 2020 . p. 118-125.