Segmentation and Superpixels Flashcards
What three things make for optimal edge detection?
- Good detection (marks all and only the edges)
- Good localisation (points mark should be as close to the edge as possible)
- Minimal response (each edge should only be reported once)
What is non-maximal supression?
Checks if pixel is a local maximum along gradient direction to check is line is actually likely to be part of edge (no weak lines)
What kind of thresholding did Canny use?
Thresholding with Hysterisis
What kind of equation did Canny use to create his edge detection method?
Optimisation of the 3 factors (localisation, detection, response)
How does thresholding with hysteresis work?
High values over high thresholding are considered part of the edge, and low values below the threshold are only considered if they connect to the main edge
What method did Canny show was the best way to detect step edges in noise?
1st derivative of a Gaussian
What is segmentation?
Dividing an image into meaningful regions for analysis
How does hysteresis prevent noise in thresholding?
By using two thresholds
What are some properties we can segment an image by?
- Intensity or colour
- Textures or patterns
- Spectral profiles that provide multidimensional image data
What are the three main approaches to segmentation?
Clustering, Region-based, edge based
What is clustering as a segmentation approach?
Finding similar pixels with no regard for where they are
What is region-based as a segmentation approach?
Finding physically connected pixels (e.g. region growing, split and merge)
What is edge-based as a segmentation approach?
Emphasis on the boundaries between regions e.g. watersheds
What is region growing used for?
Region based segmentation
How does region growing work?
Start with a seed pixel / group of pixels then expand by adding similar pixels using statistics calculated for the region
What’s an example of a statistic used for region growing?
Grey average calculated then if nearby pixels have similar averages they are added to the region
Where should the seed be placed for region growing?
In a uniform region
What is split & merge used for?
Region based segmentation?
How does split and merge work for region based segmentation?
Splitting divides two inconsistent regions and merging combines adjacent regions that are consistent
How does split and merge start?
The whole image is taken to be one region then a measure of internal similarity is decided on and computed
What is done in split and merge from the point of the whole image being a region?
If there is two much variety (based on the computed measure) the image is divided (normally into quarters) and this is repeated until each region is similar enough or a minimal size is reached
How does edge based segmentation work?
Finds edges and then defines a region as an area bounded by that edge
What are catchment basins and watersheds?
On a gradient image, catchment basins are low areas (regions of the image) and watersheds are the lines dividing them
Do we have to use intensity gradient for finding watersheds?
No, could use hue etc
What are watersheds used for?
Edge based segmentation
What is a good way to visualise the gradient of an image?
As a 3D terrain
How does the watershed algorithm work after sorting pixels low to high?
Each pixel is given is:
- given a new label if its neighbours are unlabelled
- if its neighbours have a single label it gets that
- if it has neighbours with two or more neighbours it is a watershed
What are two drawbacks of the basic watershed process?
- it can give thick watersheds rather than fine lines
- it is sensitive to noise so can generate lots of small regions
Why might bin sort be needed for watershed?
To sort the pixels in numerical order
How does bin sort work?
Makes a bin for each possible value and places values in their relevant bins
What might be done to gradient before watershed?
Quantisation, reducing range of gradient to reduce amount of segments and effect of noise
What can be done to stop loads of mini regions being created in watershed process?
Having a threshold value where no more watersheds are created
