wk5 Flashcards
Name some types of images and their computational representations
GrayScale: 2D arrays
Colour: 3D Arrays
EM Spectrum: Multidimensional arrays
What are the 4 ways to characterise images
- Spatial Resolution
- Intensity Resolution
- Time Resolution
- Spectral Resolution
How can we analyse and or visualise intensity resolution
histograms
they can be also represented as probability distributions (of which histograms may be a form)
How is signal to noise ratio represented
ratio of mean to std
what are the steps to characterising images
Segmentation,
- Start by segmenting the image according to some characteristics
- Then separate the images according to divisions within that characteristic
The methods themselves are case specific
Name the 2 types of segmentation
Regional Segmentation
Boundary Based Segmentation
What are the 3 approaches to conducting segmentation. Give their advantages and disadvantages as well as examples
Non-Automatic (by hand):
- likely to be correct for difficult and holistic problems
-takes ages for complex problems
Semi-Automated:
-Thresholding:
-Active Contours:
Automated:
-Model Based
What is the thresholding problem and what are the potential solutions to it
Thresholding Problem:
- could ordinarily do this no problem by manually selecting a threshold and inspecting
- If the image is noisy then it will not be possible to differentiate between noise and ground truth by simply thresholding
- this can be visualised by overlapping histograms characterised by a large std
It could be remedied by noise filtering, adaptive thresholding via variance minimisation (Otsu Method)
What is the Otsu method and how does the algorithm work
The Otsu method minimizes the intra-class variance / maximizes inter-class variance:
1) compute histograms and pixel probabilities (that they belong to the same group) per intensity level
2) set up initial weights and means for each intensity levels
3) for each possible threshold 1:N
3.1) update weights and means per calss
3.2) compute standard deviations
4) the desired threshold will be smallest std
std is equal to the sum of the std per class, each class is multiplied by the corresponding weights
what are the issues with manually thresholding
-time consuming
-subjective
-operator error
Aside from noise filtering, how can we deal with the thresholding problem, describe this method
Mathematical Morphology:
Operations are defined by two sets:
1) actual image, 2) structured shapes / images
A structuring element can be thought of as a brush with which an image is overpainted with them.
3) Find all pixels in an image that correspond to the structure and “paint over” the image and link the pixels together. (call dilation)
4) Remove the layer for each time a dilation is applied (this is called erosion)
What are the two principles of Mathematical Morphology
Dilation:
-Adding a luaer of pixels to the periphery of objects
-Object wil lgrow larger, closed objects will be merged and holes in structures will aslo be closed
Erosion:
-removes a layer of pixels around an object
-Object will get thinner, if it already is thin itll break into several sections
Num of dilations must be equal to the number of erosions and vice versa
Briefly what is snake segmentation
A sum of 3 weighted integrals which degine a segmentation between an object and backgrounds
-First term corresponds to sensitivity of the stretch of the snake
-Second term correpsonds to sensitivity according to some image properties
-Third term corresponds to some prior domain knowledge
What is watershed segmentation
Segments images according to a histogram:
- local minima
- areas that would fall into local minima
- watersheds, local maxima
Give an example of an application of image registrations
Consider many brain scan images, some healthy, some not
- we want to segment them into healthy and not healthy
- we will take all images and create a “standard” or common brain map and compare signals in an area from healthy and non healthy brains
- the standard brain image created is called a map or atlas
This becomes an alignment problem
which two properties are taken into account for registrations
Geometric and Photometric alignment
How do we find geometric and photometric alignments between images
Between two types of images e.g. MRI and CT scans, we find a function to map one image to the other
Whar are the two types of registration
Individual Registration
Group Representations
What is the minimisation problem associated with image registrations
Minimise difference between reference image and target image when reference image is transformed by some function to type of the target image
What should we consider when setting up a transformation function T
(aka Registration Components)
- what entities do we match
- what class of transform
- what similarity criterion to use
- what search algorithm to find the minimum T
- what interpolation method to use
What other features can we consider for matching between different images in image registrations
- Image values
- Edges Contours
- Salient features
- corners
- centres
- points of high curvature
- line intersections
Example of landmarks to match in image
What are some examples of similarity measures
Mean Square Error
Information similarity
why is mean square error often not good enough for image refistration
there may be a functional dependence between intensities as there often is in medical images
what similarity measures are used for aligning images via histograms
-suppose some image has N pixels. The number of image intensities per intensity value divided by the number of pixels will give the probability that some pixel will have that intensity
-Given some transformation, the concept is extended to that of a joint histogram
- a joint histogram can quantify alignment. a tighter histogram means better alignment