Download Image Segmentation: Techniques and Algorithms and more Study notes Digital Signal Processing in PDF only on Docsity!
ECE 6258 Russell M. Mersereau
ECE6258 Lecture 39
Image SegmentationSource: Images taken from Gonzalez and Woods
ECE 6258 Russell M. Mersereau
Image Segmentation ^
Segmentation subdivides an image into itsconstituent regions or objects ^
Definition of regions is application dependent
^
Approaches ^
Edge linking and boundary detection
^
Thresholding
^
Region-based segmentation
^
Morphological watersheds
ECE 6258 Russell M. Mersereau
Edge Detection ^
One of the most intuitive ways to segment objects is to findtheir boundaries using an edge detector.
^
Edge detectors try to estimate the directional derivative. ^
This operation is inherently noisy ^ Different types of results depending upon the edge type. ^ Edges are frequently broken. ^ Many “edges” do not correspond to boundaries.
ECE 6258 Russell M. Mersereau
Effect of Noise
ECE 6258 Russell M. Mersereau
Prewitt and Sobel Gradient Operators ^ Each tries to smooth in the direction along the edge.
ECE 6258 Russell M. Mersereau Results
Still requires alot of work to sortout the boundaries.
ECE 6258 Russell M. Mersereau
With initial 5x5 Pre-smoothing
ECE 6258 Russell M. Mersereau Edge Linking ^ Edge detection algorithms are typically followed by linkingprocedures to assemble edge pixels into meaningful edges. ^ One approach ^ Edge points that share one or more criteria are linked ^ Edge strength in response to edge operator. ^ Direction of the gradient vector
^
Other approaches ^ Hough transforms ^ Graph theoretic approaches
^
Edges are more commonly used in conjunction withother methods.
ECE 6258 Russell M. Mersereau
Region-Based Segmentation ^ The objective is to partition an image into disjointhomogeneous regions. ^ Let R represent the entire image and R the subregions. i These are defined by for some logical predicate P (.)
ECE 6258 Russell M. Mersereau Region growing Region growing groups pixels or subregionsinto larger regions. Approach starts with a set of “seed” pixelsand appends neighboring pixels that havesimilar properties. 11/26/ ECE 6258 Russell M. Mersereau
Result
Histogram of original
image
image
pixel intensity=
grown regionsmin intensity=
boundariessuperimposed
ECE 6258 Russell M. Mersereau Split-and-Merge Algorithms ^ An alternative to region growing is to subdivide an image initially intoa set of arbitrary disjoint regions and then split or merge the regionsin an attempt to make them more homogeneous. ^ The most common implementation subdivides square regions intoquadrants (quadtrees).
ECE 6258 Russell M. Mersereau
Algorithm 1. Split into four disjoint regions and region for which 2. Merge any adjacent regions R and j R k^ for which 3. Stop when no further splitting or merging is possible.
ECE 6258 Russell M. Mersereau Split-and-Merge Result original thresholding result Split-and-Merge result 11/26/ ECE 6258 Russell M. Mersereau
Segmentation by Morphological Watersheds ^ Think of the image as a 3-D map, with thecoordinates corresponding to longitude and latitudeand the gray-value corresponding to altitude. ^ We can distinguish three types of points: ^ (a) local minima ^ (b) points where, a drop of water placed at that point wouldflow to a unique local minimum. These points form the watershed of the minimum. ^ (c) points at which the water would be equally likely to flowto two or more minima. ( divide lines or watershed lines .) ^ These segmentation algorithms try to find thewatershed lines.
ECE 6258 Russell M. Mersereau Approach 1. Punch a hole in the surface at each regionalminimum.
Allow water to flow into the hole at a uniformrate.
When the rising water in different catchmentbasins is about to merge, construct a damto prevent the merging.
When only the tops of the dams are visibleabove the watersheds, stop. These are thewatershed lines.
ECE 6258 Russell M. Mersereau
Oversegmentation image oversegmentation
ECE 6258 Russell M. Mersereau Using markers to limit oversegmentation^ ^ Define a marker as a region surrounded by “higher points”, that is aconnected component, in which all points in the component have thesame value.
Watershed linesAnd markers
Segmentation
result
Computed usingwatershedson gradient ofeach region.
ECE 6258 Russell M. Mersereau
Region labeling and counting ^ How many bacteria are there in this picture? (region counting) ^ Which pixels belong to the same object (region labeling)? ^ How large is each object (region statistics)? Original Bacteria image After thresholding
ECE 6258 Russell M. Mersereau ^ A region is a set of pixels, where each pixel can be reached fromany other pixel in the region by a finite number of steps, where eachstep starts at a pixel and ends in the neighborhood of the pixel. ^ These typically lead to the same regions, except when a region isonly connected across diagonally adjacent pixels. Neighborhoods 4-neighborhood 8-neighborhood
ECE 6258 Russell M. Mersereau
Region labeling algorithm (4-neighborhood) ^ Loop through all pixels x [ n , n 1 ], left to right, top to bottom. 2 ^ If^ x [ n^1 , n ]=0, do nothing. 2 ^ If^ x [ n^1 , n ]=1, distinguish four cases 2 ^ Generatenew regionlabel Second pass through image to consolidate equivalent labels.
Copy labelfrom above
Copy labelfrom the left
Copy label from theleft. If labels aboveand to the leftare different, storeequivalence
ECE 6258 Russell M. Mersereau Example: region labeling 11/26/ ECE 6258 Russell M. Mersereau
Region statistics ^ Measures the size of each region of a labeledimage. ^ Initialize counter ( label )=0 for all label ^ Loop through all pixels x [ n , 1 n ], left to right, top to 2 bottom ^ If^ x [ n^1 , n ]=0, do nothing. 2 ^ If^ x [ n^1 , n ]= 2 1 , increment counter ( label [ n , n 1 ]). 2
ECE 6258 Russell M. Mersereau Small region removal ^ Loop through all pixels x [ n , 1 n^2 ], left to right, top to bottom ^ If^ x [ n^1 , n ]=0, do nothing. 2 ^ If^ x [ n^1 , n ]=1 and counter(label( 2 x [ n^1 , n ]))<S, then set 2 x [ n , n 1 ]=0. 2 ^ Removes all regions smaller than S pixels .
