Skip to content

Latest commit

 

History

History
39 lines (29 loc) · 2.2 KB

README.md

File metadata and controls

39 lines (29 loc) · 2.2 KB
Raw

Canny-Edge-Implementation

(My own implementation on Canny Edge runs on both CPU & GPU: CUDA)

In Edge.h, there are two class member functions to do canny edge detection: cannyEdge() & cannyEdge2()

cannyEdge2() is an optimized version of cannyEdge() in 2D convolution process, it separates the sobel kernel into two smaller kernels for faster convolution process and reduces one for-loop.

Update 6/30

  • Provided a docker image: docker pull wchern/dev:opencv
  • Added Makefile (make clean supported) for ubuntu. Just make to compile.
  • Added argument support for selecting benchmark parameteres (smalle or large image, number of iterations, instructions below)
  • Edited define marco for gcc (version 7.4.0) compiler & microsoft's compiler (SIMD vectorization available on g++)
  • Factorized a lot of codes
  • Removed atan() & sqrt(gx^2 + gy^2) for speed boost
  • Added thread control function to select number of threads according to image size
  • vectorized some for-loops
  • optimized nonMaxSuppression() & hysteresis_threshold() to reduce if statements

Update on 7/1

  • Vectorized a for-loop using omp inbranch for flow control

Update on 7/9

  • Revised hysteresis threshold's algorithm, using recursive DFS now with same performance
Input size Time (ms) (Avg. of 1000 runs) OpenMP Enable? Env
637 x 371 0.593 ms (before: 3.647 ms) Yes, 4 threads gcc 7.4.0, ubuntu 18.04 (docker), -O2 optimization, omp simd
3840 x 2160 20.65 ms (before: 106.775 ms) Yes, 8 threads gcc 7.4.0, ubuntu 18.04 (docker), -O2 optimization, omp simd

Parameter Usage: ./app_name -firstPara -secPara thirPara

  • -firstPara = valid inputs are -small or -large
  • -secPara = valid input is -iter
  • -thirPara = valid input is any positve integer number for iteration

1

CUDA implementation added 7/22 (See Edge_Cuda.cuh, tested on Windows with Cuda 10.1)

Can be further optimized and factorized tho!