Nonlinear Filtering Flashcards
What is salt and pepper noise, and how it is produced?
Salt and pepper noise is a type of noise that leaves little ‘flakes’ all over the image, looks like pepper and salt has been sprinkled on top of the image
Caused by sensors either failing to respond, or saturating in error
False saturation gives a white spot, whereas a failed response gives a black spot
What is the median filter, and how does it work?
Takes a small window of pixels
Then, gets the median value i.e. the middle of the pixels, and sets each pixel to the median value in the local window.
Noise pixels are considered outliers, and noise would have to affect > 1/2 the pixels to appear in the output.
What is the advantage of using median filtering, and what is a disadvantage?
Good given small regions of speckle noise
Less good if edges are important
What is diffusion, and what filters can be seen as diffusion processes?
Diffusion is spreading out
Gaussian and mean filters can be seen as diffusion processes
What is anisotropic diffusion and how does it work?
Anisotropic - not the same in all directions
Basic idea - Make each pixel more like those neighbours that it is already similar to
What does the similarity function mean in anisotropic diffusion?
s(p,q) - has values in the range between 0 to 1
If the pixels p and q are similar, then s(p,q) is close to 1
If the pixels p and q are different, then s(p,q) is close to 0
You use s(p,q) to compute a weighted average of pixel values - the new value at a pixel, p, is based on all its neighbours, q
The smoothing function needs to be found.
If d is the difference between p and q, and D is the maximum possible difference we can use, then:
(D-d)/D
What does a higher value of K do within anisotropic filtering?
A higher value of K gives greater smoothing, but edges are still quite sharp
What happens when you repeatedly apply an anisotropic filter?
It gives greater smoothing, but starts to lose edges quite rapidly with higher iterations.
What does an anisotropic filter become when its similarity function is equal to 1?
It devolves into a mean filter
What is bilateral filtering, and how does it work?
Bilateral filters modify Gaussian smoothing in a similar way:
One Gaussian weights pixels that are near the source
Another Gaussian weights pixels that have similar intensity to the source pixel
What is Bilateral filtering made up of?
A normalisation factor
Space weight
Range weight
What is different between Gaussian filtering and Bilateral filtering?
Bilateral filtering has a different kernel shape depending on the image context, whereas Gaussian has the same kernel everywhere.
What is a binary image?
A binary image is up of pixels that are either equal to 0 or 1, no other values
What is thresholding and how does it work?
Choose a threshold value (T)
Consider each pixel in turn:
If the brightness at a pixel is less than T, that pixel is an object (i.e. 1)
Otherwise, it is part of the background (i.e. 0)
In summary, thresholding is used to highlight objects or areas of interest, whilst removing all other noise. Meant to bring out the parts of the image you want, whilst ignoring the rest of it.
What happens when the threshold value is either too high or too low?
If it is too high, background pixels will be classified as foreground objects
If it is too low, then object pixels will be considered background
