Convolutions Flashcards
How might we normalise a kernel?
By dividing each element in the kernel by the number of elements in the kernel.
For example dividing each element in a 3x3 kernel by 9.
Is the convolution operation associative?
Yes
When is a kernel separable?
A kernel A is separable if there exists kernels B and C, such that
B * C = A
What are the benefits of using 2 separate kernels, rather than one large kernel?
Efficiency Gain
Consider a MxN image:
Given a 3x3 kernel we must perform 9MN calculations when convolving.
However if we split this up into 3x1 and 1x3 kernels, we only perform 23MN operations, 6MN.
What is a box filter?
A smoothing or averaging kernel,
where each pixel is set to the average of all of its neighbours.
What number do all of the values in our kernel usually add up to?
We try to make the values in our kernel add up to 1.
This is so that after performing the convolution, we do not go over the range of 255.
What is noise?
Noise is the deviation of a value from its expected value
What is salt and pepper noise?
This is noise where some values are set to the maximum or minimum possible value randomly.
How can you smooth using multiple images?
Take multiple images and add them up, because the noise is zero mean, the sum of the noise approaches zero.
We can then take an average giving us a noiseless image.
This process is called image stacking.
What is adaptive smoothing?
It is where we apply smoothing, but only use the smoothed value if it is not too different from the original value
What are the benefits of a Gaussian filter?
It reduces ringing
It also applies weighted smoothing, values further away from the current pixel are less relevant
What is a good way to remove salt and pepper noise?
Using a median filter
