I know it's important to apply Gaussian blur to an image before using Canny to detect edges. My question is: does cv2.Canny() do Gaussian blur on its own or it is necessary to apply cv2.GaussianBlur() before cv2.Canny()? The documentation isn't clear on this point.
Does cv2.Canny() perform a Gaussian blur?
Viewed 2917
2 Answers
Recall that using OpenCV in Python is in fact a wrapper to OpenCV's C++ interface. The Canny edge detection algorithm's implementation file in C++ can be found here: https://github.com/opencv/opencv/blob/master/modules/imgproc/src/canny.cpp
If you examine this source, you can see that there is no blurring done on the image whatsoever. However, thanks to the correction made by Yves Daoust (see below), the Sobel edge detection algorithm is used for finding the gradient. There is an aperture input parameter that specifies the size of the Sobel kernel. Using the default means you're using a 3 x 3 kernel and no blurring is applied. However, anything above 3 x 3 will use a Gaussian kernel. Therefore, if you want to use the default Sobel kernel, you would need to apply the blurring yourself. If you vary the size of the Sobel kernel you don't have to perform blurring yourself.
In fact, in the official OpenCV tutorial where they demonstrate the use of Canny, the image is blurred manually prior to the detection: https://docs.opencv.org/3.4/da/d5c/tutorial_canny_detector.html. However, they use a 3 x 3 box filter here prior to the algorithm as you're using a 3 x 3 Sobel kernel.
tl;dr - You need to blur the image yourself prior to using the Canny edge detection algorithm if you decide to use the default aperture / Sobel kernel size. If you go larger than 3, you don't have to blur the image.
The aperture argument of Canny controls the size of the Sobel filter (from 1 to 7 ?), which in fact applies a lowpass filter before taking the derivative. Unfortunately, the documentation is not explicit about that lowpass filter, though Gaussian is cited.
I bet that a binomial filter is in fact used (which is a close approximation to a Gaussian).
Related