Feature Detection Flashcards
Edge Tracking
Edge Strength: magnitude of intensity gradient
Edge orientation: direction of intensity gradient
edge location: zero crossings of Laplace operator
Canny Edge Operator
problem: seperate edge from noise
assumptions: 1. edge gradient stronger than noise gradient
2. edge direction varies slowly
3. edge strength varies slowly
thresholds: 1. t1 voxel is edge
t2 voxel is edge if neighbor is edge
+ detects parts of edges
- sensitive to noise
Hough Transform
idea. find instances of specific edge model viat voting scheme
hough space represents all parameter combinations
+ robust against occlusion, signal drops, noise
structure tensor
summarizes the predominant directions of the gradient neighborhood
Harris Corner Detector
value depends on average intensity variations in arbitrary directions
high value = point of interest
+ detects corners and edges
- computational expensive
SUSAN corner detector
intensity of poi differs from intensities in neighborhood
intensity difference is higher than threshold
poi is maximum of difference
+ detects corners and single points
+ faster computation
- sensitive to noise
Blob detection
Blob: circular structures in image
use: laplacian of gaussian, difference of gaussions, determinant of hessian
Feature detection SIFT
generates and uses features to detect and identify objects in image
- key point generation
- key point reduction
- feature computation
Feature detection MSER
seperation of image into local homogeneous regions with maximum contrast
intensity based feature detection
Segmentation
part image to connected regions which can be classified
Segmentation - Thresholding
image voxels into labeled voxels and background voxels based on intensity thresholds
- interactive thresholding
- automatic thresholding
Segmentation - Otsu
find optimal threshold
- characteristics of foreground and background are nearly equal
Segmentation - homogeneity based
seperate image into smallest number of segments that all fulfills homogeneity criterion
Region Growing
start with seed point in object region
homogeneity criterion determines inclusion of adjacent voxels
neighborhood definition: bigger (robust against noise), smaller (robust against leaking into other regions)
Region Merging
- each voxel is own region
- if two adjacent voxel fulfill homogeneity criterion, merge to one region
relational structures
information between semantically important image parts (inside, neighboors)
Region Splitting and Merging
- start: image is one region
- split until homogeneity criterion is met
- Splitting result in RAG
- merging is speed up
Watershed Transform (WST)
- treat as image as landscape
- local minimum is sink
- fill sinks with water, when water meets include watershed
Live Wire
find minimum cost path between user-defined markers and create contour as series of path segments
Morphological Operations
- study of shape
- enhancements of object structure (skeletonization, thinning, thickening)
Opening and Closing
- opening: erosion followed by delation (remove small isolated regions)
- closing: dilation followed by erosion (close holes)
- remove leakage / close gaps
Skeletonization
- skelaton: minimal structure discribing topology of object
- analysis of branching objects / tree structures
- erosing from vessel surface
- hierachry position of brances (lowest levels can be irrelevant)
Measurement of object extend
- identify of size of object
- use axis-oriented bounding box
- maybe need to rotate first
