Image Processing 3 + 4 Flashcards
Hamiltonian
Perpendicular to the gradient vector.
The lines are isophotes , lines of constant gray value.
Laplacian
Negative if the gradient vectors are converging
Zero if the gradient vectors are parallel
Positive if the gradient vectors are diverging
Edge detection
- Compute gradient magnitude image (Sobel Kernel)
- Threshold to only show edges with a certain edgeness
- Gaussian filter (reduces effect of noise)
- Laplacian
Discrete approximation of kernel issues
- Truncation error (far, low ends are discarded. solution: wider kernel)
- Discretization error (over and underestimation. solution: denser sampling)
Median Filtering
- Place all the grey values of a kernel around a pixel in a row, and remap it to the value in the middle of the row
- Removes noise, doesn’t shift edgess
- Min/Max filters work similarly
- These are not linear filters, but rank filters
Nearest neighbour interpolation
Value of nearest pixel
Linear interpolation
Added values of weighted value of w = 1 - distance to pixel (block kernel)
2D Linear interpolation
Added values of pixel times opposing area (pyramid kernel)
Anisotropic diffusion
Smoothing, but not equally in all directions. Is based on the heat equation.
Grey value opening
ball from below, idempotent
Grey value closing
ball from above, idempotent
Dilation/Erosion
Uses Minkowski addition/substraction, dual operators (one on f and other on f-1 makes the results also each other’s inverse)
Opening
Erosion then Dilation (details get lost, size remains) dual operators and idempotent
Closing
Dilation then Erosion (structures can fuse) dual operators and idempotent
Geodesic dilation
Apply dilation in iterations, but limiting growth with a control image. Eventually no more changes.
Residue
Difference between the results of morphological operations (is a morphological gradient)
White top hat filter
Difference between f and opening of f (used to remove background gradient when followed by tresholding)
Ultimate Residue
Erode, Reconstruct, add Residue. Iterations give the ultimate residue
Skeleton
Add ballz in image. Biggest ballz is on skeleton line.
Hit-or-Miss transform
S1 foreground pixels have to fit, S2 foreground pixels have to be absent.
morphological gradient
Difference between the erosion and dilation of an image (looks like a fucked binary laplacian with white in the middle)
Gibs ringing
Artifact van een rond, binair masker toepassen op een FT. Een geblurde versie van het masker met minder harde randen verlagen dit effect.
