april 5 notes
links
key ideas
- convolution combines two functions (or images) by putting one (scaled)
at every position of the other.
- for image processing, one function is the image, the other is the "kernel",
typically much smaller
- gaussian blur is a primary example; the gaussian is the "kernel"
- by using kernels like [-1,0,1] or [1,-2,1] we can take first or 2nd derivatives
- in horizontal, vertical, or diagonal directions
- which can give a gradient (direction and magnitude of pixel change)
at each point
- the length of that gradient is a typical "edge detection"
- inverse of a matrix is typically done by LU decomposition
(gaussian elimination)
- sparse matrices are once where most of the entries are zero;
there are special storage methods and algorithms for dealing with them,
depending on where the zeros are.
- rotations and projections are 3D/2D transformations
of objects described by (x,y,z) or (x,y) coordinates in space.
Both are accomplished with matrix multiplications using specific
rotation or projection matrices.