There is also some intuition for being able to identify these filters as those that extract gradients. Let's consider the horizontal gradient extraction filter. If we applied this filter to a section of an image, we are effectively interested in only the pixel columns to the left and right of our pixel of interest. If these columns are the same, then our resulting convolution should return a 0 value for our pixel of interest, thus resulting in a black pixel in our resulting image. If they differ, then it is the case that our convolution result is non 0, resulting in a non-black pixel in our image.
I don't believe that the values for the columns of this filter necessarily have to be -1 -2 -1 and 1 2 1 respectively. Instead, I believe they could have had different values so long as the magnitudes of the columns are the same and the signs are opposite each other
antony-zhao
It seems these filters are also the ones that are commonly used for edge detection (https://sbme-tutorials.github.io/2018/cv/notes/4_week4.html#edge-detection-kernels), although wikipedia listed some different filters as well.
It also seems Brandon's intuition about the values on the filters are also correct, as there are Prewitt filters listed which use -1 and +1 rather than -1, -2, +1, +2,
Edge7481
Seems like the term gradient in this context is the same as the one used in calculus. the magnitude of the gradient tells us how quickly the value of the pixel changes, which typically occur around edges, whereas the direction of the gradient tells us the orientation of the edge
There is also some intuition for being able to identify these filters as those that extract gradients. Let's consider the horizontal gradient extraction filter. If we applied this filter to a section of an image, we are effectively interested in only the pixel columns to the left and right of our pixel of interest. If these columns are the same, then our resulting convolution should return a 0 value for our pixel of interest, thus resulting in a black pixel in our resulting image. If they differ, then it is the case that our convolution result is non 0, resulting in a non-black pixel in our image.
I don't believe that the values for the columns of this filter necessarily have to be -1 -2 -1 and 1 2 1 respectively. Instead, I believe they could have had different values so long as the magnitudes of the columns are the same and the signs are opposite each other
It seems these filters are also the ones that are commonly used for edge detection (https://sbme-tutorials.github.io/2018/cv/notes/4_week4.html#edge-detection-kernels), although wikipedia listed some different filters as well.
It also seems Brandon's intuition about the values on the filters are also correct, as there are Prewitt filters listed which use -1 and +1 rather than -1, -2, +1, +2,
Seems like the term gradient in this context is the same as the one used in calculus. the magnitude of the gradient tells us how quickly the value of the pixel changes, which typically occur around edges, whereas the direction of the gradient tells us the orientation of the edge