Linearity and Convolution
Linearity
- Additivity
- Multiplicative scaling
Shift Invariance
An operator behaves the same everywhere. In other words, the output depends on the pattern of the image neighborhood, rather than the position of the neighborhood. An operator must give the same result on a pixel regardless of where the pixel (and its neighbors) is located to maintain shift invariance.
Convolution
Computational Complexity
If an image is and a filter is , the time complexity is , which can be fairly large
Separability
If the filter is separable, meaning you can get the kernel by convolving a single column vector by a single row vector, as in the example: Then we can use the associative property to remove a lot of multiplications.