Red0 - Yellow60 - Green120 - Cyan180 - Blue240-Magenta300 Thinning & Skeleton
Hue: The “true color” attribute. Saturation: The amount by The end points of the skeleton extend all the way to the edges
which the color as been diluted with white. Value: The degree of the input object. Medial axis skeletonization:
of brightness: a well-lit color has high intensity; a dark color Segmentation [[TN, FP], [FN, TP]] P: Predictive Postive
has low intensity Accuracy: (TP + TN) / all; Precision = TP / (TP + FP); Recall
Aliasing occurs when a signal is sampled at regular time =TP/(TP + FN); F = 2 x pr/(p+r); Jaccard Index: IoU, overlap /
intervals at slightly less than the period of the original signal. union
linear filters: a linear combination of the intensity values of the Segmentation principles: Discontinuity: To partition an image
center pixel and all neighboring pixels. based on abrupt changes in intensity (Point, Line
Laplacian [.-.] – sharpen / Gaussian / Averaging filters and Edge Detection); Similarity: To partition an image into
Median Filter - non linear filter similar regions (thresholding, region growing)
The basic difference between convolution and correlation is Canny Edge Detector: Noise reduction; Gradient calculation
that the convolution process rotates the kernel by 180 degrees. (Directional change in intensity in an image; Change in the
Edge Opreator [ - , 0, +] intensity in both the horizontal and vertical directions); Non-
You should normalize your image (scale between 0 and 1) for maximum suppression (finds the pixels with the maximum
Log & Gamma transformations value in the edge directions);Double thresholding (Strong –
High frequency components -large changes in grey values over most likely an edge; Weak – possibly an edge; Non-relevant –
small distances; (edges and noise) not an edge); Connectivity analysis (Connects weak pixels to
Low frequency components -by little change in the gray values. strong ones, if and only if at least one of the pixels around the
(backgrounds, skin textures) one being processed is a strong one); sigma : large detects
High pass filters (Sharpening) passes over the high frequency large scale edges; small detects fine features
components and reduces or eliminates low frequency Morphological Gradient: Beucher gradient - difference
components. Low pass filters (Smoothing) between the dilation and the erosion (D - E) by the SE
Band pass filters passes frequencies within a certain range and Morphological Watersheds:1.Initially, the set of pixels with
rejects (attenuates) frequencies outside that range. minimum gray level are 1, others 0. 2. In each subsequent step,
notch filter and butterworth filter are band-stop filter with a we flood the 3D topography from below and the pixels covered
narrow stopband. by the rising water are 1s and others 0s.
Smoothing Filters: Spatial Domain: Gaussian, Averaging - linear The watershed transform finds "catchment basins" and
filters, Median - good for impulse noise "watershed ridge lines” in an image by treating it as a surface
Frequency domain: Gaussian LPF, Ideal - Block all frequencies where - light pixels are high - dark pixels are low. (Dam
higher than the cut-off frequency ;Ringing (ripple effect) when Construction)
the cut-off too high,Butterworth: motion Distance transform of a binary image is defined by the distance
Sharpening Filter Spatial Domain: Laplacian (can be converted from every pixel to the nearest non-zero valued pixel
to frequency space) Frequency domain: HPF Ideal, Butterworth Detect Lines: Hough Transform (max votes in Hough space ->
and Gaussian image space); Each edge pixel votes (accumulator) in
XOR, A + B - AB parameter space for each possible line through it; Overlap of
Dilation (Overlap+) 1. If there is no overlap, the input pixel is circles can cause spurious centers
left at the background value. 2.If at least one pixel in the SE Markers: Internal markers are used to limit the number of
overlaps with a foreground pixel in the image underneath, the regions by specifying the objects of interest; External markers
input pixel is set to the foreground value. are those pixels we are confident to belong to the background
Erosion (Overlap-)1.If at least one pixel in the structuring The markers should be the local minima values; The further
element overlaps with a background pixel in the image away these pixels are from the markers, the higher its value.
underneath, the input pixel is set at the background value. 2.If Types of Edges: step edge: ideal edge
all pixels in the structuring element overlap with a foreground
pixel in the image underneath, the input pixel is left at the
foreground value.
opening (Erosion + Dilation) retaining the original object size;
clear an image of noise;may distort the shape size of the Superpixels create: Felzenszwalb's algorithm (a graph-based
object; Opening can remove small bright spots (i.e. “salt”) and approach)
connect small dark cracks. D(C1, C2) = true: Dif(C1, C2) > MInt(C1, C2) -> no NOT merge
closing (Dilation + Erosion): retaining the original object size; Else: Merge; MInt(C1, C2) = min( Int(C1) + t, Int(C2) + t);
fill holes in a region; Closing can remove small dark spots (i.e. Int(C) = max(inter distance)
“pepper”) and connect small light cracks
opening + closing remove both bright and dark artifacts of
noise.
grayscale dilation (Max) and grayscale erosion (Min)
White Top-hat: Original image minus its opening; Returns the
bright spots of the image that are smaller than the structuring
element
Black Top-hat: Closing minus the original image; Returns the
dark spots of the image that are smaller than the structuring Int(red) = 18 - 15 (or 15 – 12) = 3 (Note: there is no edge
element between 18 and 12.); Int(gray) = 29 – 23 = 6; T = 3; MInt(red,
gray) = min(3 + 3, 6 + 3) = 6; Dif(red, gray) = min(23 – 15, 26 -
Hue: The “true color” attribute. Saturation: The amount by The end points of the skeleton extend all the way to the edges
which the color as been diluted with white. Value: The degree of the input object. Medial axis skeletonization:
of brightness: a well-lit color has high intensity; a dark color Segmentation [[TN, FP], [FN, TP]] P: Predictive Postive
has low intensity Accuracy: (TP + TN) / all; Precision = TP / (TP + FP); Recall
Aliasing occurs when a signal is sampled at regular time =TP/(TP + FN); F = 2 x pr/(p+r); Jaccard Index: IoU, overlap /
intervals at slightly less than the period of the original signal. union
linear filters: a linear combination of the intensity values of the Segmentation principles: Discontinuity: To partition an image
center pixel and all neighboring pixels. based on abrupt changes in intensity (Point, Line
Laplacian [.-.] – sharpen / Gaussian / Averaging filters and Edge Detection); Similarity: To partition an image into
Median Filter - non linear filter similar regions (thresholding, region growing)
The basic difference between convolution and correlation is Canny Edge Detector: Noise reduction; Gradient calculation
that the convolution process rotates the kernel by 180 degrees. (Directional change in intensity in an image; Change in the
Edge Opreator [ - , 0, +] intensity in both the horizontal and vertical directions); Non-
You should normalize your image (scale between 0 and 1) for maximum suppression (finds the pixels with the maximum
Log & Gamma transformations value in the edge directions);Double thresholding (Strong –
High frequency components -large changes in grey values over most likely an edge; Weak – possibly an edge; Non-relevant –
small distances; (edges and noise) not an edge); Connectivity analysis (Connects weak pixels to
Low frequency components -by little change in the gray values. strong ones, if and only if at least one of the pixels around the
(backgrounds, skin textures) one being processed is a strong one); sigma : large detects
High pass filters (Sharpening) passes over the high frequency large scale edges; small detects fine features
components and reduces or eliminates low frequency Morphological Gradient: Beucher gradient - difference
components. Low pass filters (Smoothing) between the dilation and the erosion (D - E) by the SE
Band pass filters passes frequencies within a certain range and Morphological Watersheds:1.Initially, the set of pixels with
rejects (attenuates) frequencies outside that range. minimum gray level are 1, others 0. 2. In each subsequent step,
notch filter and butterworth filter are band-stop filter with a we flood the 3D topography from below and the pixels covered
narrow stopband. by the rising water are 1s and others 0s.
Smoothing Filters: Spatial Domain: Gaussian, Averaging - linear The watershed transform finds "catchment basins" and
filters, Median - good for impulse noise "watershed ridge lines” in an image by treating it as a surface
Frequency domain: Gaussian LPF, Ideal - Block all frequencies where - light pixels are high - dark pixels are low. (Dam
higher than the cut-off frequency ;Ringing (ripple effect) when Construction)
the cut-off too high,Butterworth: motion Distance transform of a binary image is defined by the distance
Sharpening Filter Spatial Domain: Laplacian (can be converted from every pixel to the nearest non-zero valued pixel
to frequency space) Frequency domain: HPF Ideal, Butterworth Detect Lines: Hough Transform (max votes in Hough space ->
and Gaussian image space); Each edge pixel votes (accumulator) in
XOR, A + B - AB parameter space for each possible line through it; Overlap of
Dilation (Overlap+) 1. If there is no overlap, the input pixel is circles can cause spurious centers
left at the background value. 2.If at least one pixel in the SE Markers: Internal markers are used to limit the number of
overlaps with a foreground pixel in the image underneath, the regions by specifying the objects of interest; External markers
input pixel is set to the foreground value. are those pixels we are confident to belong to the background
Erosion (Overlap-)1.If at least one pixel in the structuring The markers should be the local minima values; The further
element overlaps with a background pixel in the image away these pixels are from the markers, the higher its value.
underneath, the input pixel is set at the background value. 2.If Types of Edges: step edge: ideal edge
all pixels in the structuring element overlap with a foreground
pixel in the image underneath, the input pixel is left at the
foreground value.
opening (Erosion + Dilation) retaining the original object size;
clear an image of noise;may distort the shape size of the Superpixels create: Felzenszwalb's algorithm (a graph-based
object; Opening can remove small bright spots (i.e. “salt”) and approach)
connect small dark cracks. D(C1, C2) = true: Dif(C1, C2) > MInt(C1, C2) -> no NOT merge
closing (Dilation + Erosion): retaining the original object size; Else: Merge; MInt(C1, C2) = min( Int(C1) + t, Int(C2) + t);
fill holes in a region; Closing can remove small dark spots (i.e. Int(C) = max(inter distance)
“pepper”) and connect small light cracks
opening + closing remove both bright and dark artifacts of
noise.
grayscale dilation (Max) and grayscale erosion (Min)
White Top-hat: Original image minus its opening; Returns the
bright spots of the image that are smaller than the structuring
element
Black Top-hat: Closing minus the original image; Returns the
dark spots of the image that are smaller than the structuring Int(red) = 18 - 15 (or 15 – 12) = 3 (Note: there is no edge
element between 18 and 12.); Int(gray) = 29 – 23 = 6; T = 3; MInt(red,
gray) = min(3 + 3, 6 + 3) = 6; Dif(red, gray) = min(23 – 15, 26 -
