Midlevel,Hough+morphologica,Segmentation,Otsu+cluster Flashcards
BIlateral Filter
- Very commonly used
- A spatial filtering!
-Gaussian + Edge Preseve
-Kernel depends on image content
-Weight depends on distance and also Grey Intensity level
Derivatives
1st Order
Large Values: On all the edge
Edge type: Single
Edge thickness: Thick
Noise Sensitivity:moderate
2nd order:
Start and end of edge
Double
Thin
High
They aplify noise
First edge detector
Steps of a possible approach:
1. Low-pass filter for noise removal
2. Gradient calculation
3. Gradient thresholding |𝛻𝑓| > 𝑇
1-2: Smoothing (low pass) can be used prior calculation reduces noise effect
3-Selects only strongest edges
Canny Edge Detector: Targets + Steps
Targets:
-Low error rate
-Edge points are well localized
-Single edge point response
Canny algorithm
1. Smoothing with a Gaussian filter
2. Gradient computation (magnitude and phase)
3. Quantize the gradient angles
4. Non-maxima suppression
5. Hysteresis thresholding
Canny steps
1-Smoothing: Noise Reduction
2-Gradinet Computation: Edges are calculated using
vertical, horizontal and
diagonal masks
Edge direction is also
calculated
3-Edge quantization: Group orientations on bins of 45 degrees.
4-Non Maxima Suppression:
Reduces thickness
Gradient->1 edge
Accurate location of edge point
This process crosses an edge and selects the strongest point
5-Hysteresis thresholding: Keep strong edges
Keep week ones connected to strong edges
Reject isolated ones
Strength = Gradient Magnitude
Used thressholdin + masks
Small Sigma - finer details
Resampling
Change Image dimension
Interpolation->enlarge
Decimation->minimize image
Resampling Algorithms
1-Nearest neighbor interploation
Pixel color taken from closest pixel.
Fast but poor performance
2-Bilinear Interpolation
Weighted average of the closest 4 samples
Coefficients depend on the distance to the
samples
Balanced
3-Bicubic interpolation
Closest 16 samples
3degree polynomials
slow but better
Hough
Works in parameter space
line->intersection
Few cells: Handle not perfectly aligned pixels.
Poor line localization
Many cells:
Require Precise Alignment
Accurate
Finding lines:
Compute fi/ro values, crossed cells ++
Morphologica operators
a)Erosion
-thinning
b)Diltation
-thickening and merging
c)Opening
Erosion + diltation
-contur smoothing and removes protursions
d)Closing
Diltation + erosion
-fuse narrow breaks without increasing size
Segmentation
Subdivide image in regions based on similarity and discontinuity
Segmentation: Region Growing
Group pixels or subregions into larger regions
Based on the concept of connectivity
Makes use of predefined merging criteria
Starts from seed points, to be defined
– E.g., every pixel, pixels in a neighborhood
satisfying merging criteria
Stopping rule
Segmentation: Watershed
Features:
a)Connected segmentation boundaries
b)Connected components
Topographic surface related to intensity
Goal of the algorithm: find
watershed lines where a point would drop in 2+ minimas
Algorithm
a)Fill Dam
-Flood suface from minima where each flooded region is a basin
-Prevend merging of water->build dams taller then others-> watershed lines (diltation)
Using it on gradient:
Useful when small gradients -> thick edges
Correspond to homogenous greylevel regions
Watershed:oversegmentation
Direct application because of noise->oversegmentation.
Use markers to:
-drive segmentation
Internal:Object of interest
External:Background
Selection:
Preprocessing + criteria definition
Segmentation By thressholding
Threshold selecting 2 resulting segments.
Right threshold is tricky:
-noisy images
-illumination changes
-small regions
-distance between peaks
Otsu makes this process automatic
Segmentation:Otsu method
Global thresholding method based on histogram.
-Maximise interclass variance
-Minimize intraclass variance
Otsu algorithm
- Compute normalized histogram and set threshold
- Compute probability to be below thresshold and above
- Compute global mean (avvg image intensity) and cumulative mean up to k.
4.Compute pixel mean intensity value in classes.
5.Calculate global variance and inter class variance, quality is raport of them.
6.Find optimal threshold maximizin that raport
Smoothing helps it.
problem-noise and size
Extending otsu
Can be combined with edge detection.
Can be generalized to nonglobal thresgolding and multicategories.
Edge detection:
compute histogram and threshold on the edge of image.
Apply it in original image
Segmentation Clustering
each pixel a feature vector.
types:
1. Dvisive: from dataset to clusters.
2. Agglomerative: from pixels to clustrs
Segmentation clustering kmeans
Init:
1.Forgy-random in entire dataset
2.Random partition: build k clusters randomly then compute centroids
- Pros
– Light and simple
– Computational complexity can be reduced using euristics
– Fast convergence - Cons
– Optimality is not guaranteed
– Solution found depends on initialization
– The number of clusters, k, needs to be known in advance
– Forces spherical symmetry of clusters (in the Ndimensional space) - K-means clustering can be done based on
– The histogram (AKA gray levels, faster)
– Pixel vectors (better results, tunable) - Possible distance measures
– Intensity level difference (grayscale)
– Color channel difference (color image, depends on
color space)
– Combinations of position, color, texture
descriptor
Clusters
* Some segments shown – not necessarily connected
* Some clusters associated with objects
– Similar objects in the same cluster
* Some clusters are meaningless
* Problems with textured objects (e.g., the cabbage)
