kuliah 10 - (new) image...
TRANSCRIPT
-
10/16/2009
1
Lecture 9Program Studi S1 Reguler DTE FTUISlides 2009
IMAGE SEGMENTATIONIMAGE SEGMENTATION
PreviewPreviewSubdivide an image into its constituent regions or objectsProcess stop when the objects of interest have been isolatedAccuracy -- determine the eventual success or failure of computerized analysis procedure
2 Image SegmentationImage Segmentation
Based on one of the two properties of intensity values --discontinuity (edge-based) and similarity (region-based) gray level discontinuities (abrupt changes): point , line
and edges assemble edges into region boundary (edge linking)
region-oriented segmentation: watershed segmentation Similarity (according to a set of criteria): region
growing, split and merge
Region-based, edge based and boundary based or threshold-based
Detection of discontinuities For 33 mask, the response of the mask at any point
in the image below Point detection of isolated points is determined if
|R|T then a point has been detected Def. of isolated point: a point whose gray level is
significantly different from its background
3 Image SegmentationImage Segmentation
significantly different from its background
Line detection Move a specific mask around an image Preferred direction of each mask is weighted with a
large coefficient than other possible direction Use the mask associated with one direction and
thresholds its output |Ri| > |Rj|Horizontal, vertical, and diagonal line
General mask
4
A general 3x3 mask
Point detection maskPoint detection mask
(a) Point detection mask
(b) X-ray image of
5 Image SegmentationImage Segmentation
a turbine blade with a porosity
(c) Result of point detection
(d) Result of using eq. (10.1-2)
Line masks (with angle)Line masks (with angle)
6 Image SegmentationImage Segmentation
-
10/16/2009
2
ExampleExampleIllustration of line detection(a) Binary wire-
bond mask(b) Absolute value
7 Image SegmentationImage Segmentation
(b) Absolute value of result after processing with -45 line detector
(c) Result of thresholding image (b)
Edge detectionthe most common approach for detecting meaningful discontinuities in gray level Basic formulation The difference between an edge (local) and
boundary (global)Edge: ability to measure gray-level transitions
8 Image SegmentationImage Segmentation
an ideal edge: a set of connected pixels Models of step edge, ramp edge, roof
edgeRamp edge: Edge blurring problem cause by: optics, sampling, and other image acquisition ramp-like profileThe thickness of the ramp edge determined by the length of the ramp
Ideal digital & ramp edgeIdeal digital & ramp edge(a) Model of an
ideal digital edge
(b) Model of a ramp edge
9 Image SegmentationImage Segmentation
ramp edge. The slope of the ramp is proportional to the degree of blurring in the edge
ExampleExample(a) Two regions
separated by a vertical edge
(b) Detail near the edge
10 Image SegmentationImage Segmentation
the edge, showing a gray-level profile, and the first and second derivatives of the profile
Random Gaussian noise:Mean = 0
(a) =0
(b) =0.1
11 Image SegmentationImage Segmentation11 Image SegmentationImage Segmentation
(c) =1.0
(d) =10.0
the magnitude of the first derivative can be used to detect the presence of an edge at a point in an image the sign of the second derivative can be used to
determine whether an edge pixel lies on the dark or light side of an image
two additional properties: (1) double edges (2) zero-crossing properties: useful for locating the centers of thick edges
Gradient operators First derivative operator based on approximation of 2-D
gradient operator
12 Image SegmentationImage Segmentation
gradient operator The gradient of an image f(x,y) at location (x,y) is defined
as
The magnitude of this vector is denoted as The direction of this is perpendicular to the direction of
an edge direction
=
=
yfxf
GG
fy
x
2/122 ][)( yx GGfmagf +==
=
x
y
GG
yx 1tan),(
-
10/16/2009
3
Robert cross-gradient operator First order derivative
Prewitt operator Sobel operator Use a weight of 2 in the center coefficient to achieve smoothing
by giving more important to the center point
Have superior noise suppression to Prewitt operator Harder to implement than Prewitt operator yx GGf +
13 Image SegmentationImage Segmentation
Approximate the gradient operator by absolute vales The approximation will not be isotropic (invariant to rotation)
The Laplacian operator Is a second-order derivative defined as
For a 33 region A digital approximation including the diagonal neighbor is given
by
yf
xff 2
2
2
22
+
=
)(4 864252 zzzzzf +++=
)......(8 9152 zzzf ++=
Edge detector masksEdge detector masks
14 Image SegmentationImage Segmentation14 Image SegmentationImage Segmentation
Prewitt and Sobel masks for Prewitt and Sobel masks for diagonal edgesdiagonal edges
15 Image SegmentationImage Segmentation15 Image SegmentationImage Segmentation
Gradient operatorsGradient operators
16 Image SegmentationImage Segmentation16 Image SegmentationImage Segmentation
With smoothing (5With smoothing (55) averaging5) averaging
17 Image SegmentationImage Segmentation17 Image SegmentationImage Segmentation
Diagonal edge detectionDiagonal edge detection
18 Image SegmentationImage Segmentation18 Image SegmentationImage Segmentation
-
10/16/2009
4
Is not used in its original form for edge detectionIs unacceptably sensitive to noiseProduces double edges, complicate segmentation Is unable to detect edge directionConsists of: use its zero crossing for edge location use it for the complementary purpose of establishing whether a pixel is
on the dark or light region of an edge
The Laplacian operator is combined with smoothing
19 Image SegmentationImage Segmentation
The Laplacian operator is combined with smoothing as precursor to finding edge via zero crossingConsider the function (Laplacian Gaussian) This approximation is not unique Capture the essential shape of Smooth the image, and provide an image with zero
crossing Sometime it is called the Mexican hat functions
2
2
2
2
24
2222 ][)( )(
rr
er
rrhthenerh ==
Zero crossing image are thinner than the gradient edges
Drawbacks:Form numerous closed loops (Spaghettis effect in Fig. 10.15.g)The computation
Advantages: noise reduction and strong/tough performance
20 Image SegmentationImage Segmentation
noise reduction and strong/tough performanceEdge-finding based on gradient still are used
LoG LoG
21 Image SegmentationImage Segmentation21 Image SegmentationImage Segmentation
ExampleExample
(a) Original image(b) Sobel gradient
(shown for compariton)S i l G i
22 Image SegmentationImage Segmentation
(c) Spatial Gaussian smoothing function
(d)Lapacian mask(e) LoG(f) Thresholded LoG(g) Zero crossings
Edge linking and boundary detectionCharacterize an edge completely because of noise, break in the edge from non-uniform illumination, and other effects that introduce spurious intensity discontinuities Edge detection operators are followed by linking procedures to
assemble pixels into meaningful edges or contours
Local processing Analyze the characteristics of pixels in a small neighborhood about
23 Image SegmentationImage Segmentation
every point (x,y) in a small neighborhood All points that are similar according to a set of criteria: (1) the
strength of the response of the gradient operator; (2) the direction of the gradient operator
Similar in magnitude to the pixel at (x,y) if The predefined neighbor of (x,y) is similar in edge direction to the
pixel at (x,y) if A point in the neighborhood of (x,y) is linked to the pixel (x,y) if
both magnitude and direction are satisfied
Eyxfyxf ),(),( 00
0 0( , ) ( , )x y x y A
=TyxfifTyxfif
yxg),( 0),( 1
),(
ExampleExample
39 Image SegmentationImage Segmentation39 Image SegmentationImage Segmentation
1. Object and background occupy comparable areas, then T is the average gray value of the image
2. Object are small comparable to background, then the average value is not a good estimate, T is a midway between the maximum and minimum gray levels.
the distorted histogram f(x,y)=i(x,y) r(x,y)
40 Image SegmentationImage Segmentation
then z(x,y)=ln f(x,y)=ln i(x,y)+ln r(x,y) the histogram of z(x,y) is given by the convolution
of i(x,y) and r(x,y) if i(x,y) were constant, i(x,y) would be constant
also, its histogram would be a simple spike (impulse) if i(x,y) had a broader histogram (resulting from
non-uniform illumination), the convolution process would smear the histogram of r(x,y)
Accessing to illumination source is availablea sol. to compensate for non-uniformity
Project the illumination pattern onto a constant, white reflective surface (compensate for non-uniformity), g(x,y)=k i(x,y)H(x,y)=f(x,y)/g(x,y)=r(x,y)/k
Basic global thresholding
41 Image SegmentationImage Segmentation
The simplest of thresholding-partition the image histogram by using a global thresholdSegmentation is accomplished by Scanning the image pixel Labeling each pixel depending on whether the gray
level of that pixel is greater or less than T
ExampleExample
42 Image SegmentationImage Segmentation42 Image SegmentationImage Segmentation
-
10/16/2009
8
threshold by using a heuristic approach, based on the visual inspection of the histogram Select an initial estimate for T Segment the image using T(G1, G2) Compute the average gray level values Compute a new threshold: T=1/2(1+2) Repeat steps 2 and 4 until the difference in
T i ll h d fi d T
43 Image SegmentationImage Segmentation
T is smaller than a predefined T0(convergence)In general, a good initial value for T is the average gray level of the imageWhen objects are small compared to the area occupied by the background, average value is not a good initial
ExampleExample
44 Image SegmentationImage Segmentation44 Image SegmentationImage Segmentation
Basic adaptive thresholdinguneven illumination causes a histogram that cannot be partitioned by a global threshold Solution: (1) divide the original image into sub-
images; (2) utilize a different threshold to segment each sub-image
45 Image SegmentationImage Segmentation
g g How to divide an image? and How to estimate the threshold?
Threshold depends on the location of the pixel in terms of sub-images
ExampleExample
do not contain a boundary between object and background: variances < 75 75 100 using the threshold discussed in the
previous section the initial T was selected at the point
midway between the minimum and the maximum
ExampleExample
47 Image SegmentationImage Segmentation47 Image SegmentationImage Segmentation
ExampleExample
48 Image SegmentationImage Segmentation48 Image SegmentationImage Segmentation
-
10/16/2009
9
Optimal global and adaptive Optimal global and adaptive thresholdingthresholding
A method for estimating thresholds that produce the minimum average segmentation error histogram may be considered as probability density
function (PDF), p(z) the overall density function is the sum or mixture of
49 Image SegmentationImage Segmentation
the overall density function is the sum or mixture of two densities If the form of the densities is known, it is possible to
determine an optimal threshold the mixture probability density function is
p(z)=P1p1(z)+ P2p2(z) an image is segmented by classifying as background all
pixels with gray level is greater than a threshold T Objective: select the value of T that minimize the
average error
Determine a given pixel belongs to an object or to the background
the extreme case: background points are known never to occur (P2=0)differentiate T with respect to T (using Leibneiz rule) to find minimum error; and the optimal threshold
Three level image based on gradient For a black object on a light background
50 Image SegmentationImage Segmentation
For a black object on a light backgroundEstimating these densities is not feasible sol: employ densities whose parameters are reasonably
simple to obtain One of the principal densities is Gaussian density
two threshold values may be required to obtain the optimal solutionif the variances are equal, a single threshold is sufficient (10-3.14)If P1=P2, the optimal threshold is the average of the means
Using measures based on gradient and Laplacianto deep the valley between histogram peak the optimal threshold may be accomplished for
other densities if the Rayleigh and log-normal densities
Mean square error may be used to estimate a composite gray-level PDF of an image from the
51 Image SegmentationImage Segmentation
image histogram For Example:
the mean square error p between p(z) and the image histogram
determine analytically that minimize this mean square error is not a simple matter
=
=n
iiirms zhzpn
e1
)]()([1
GL PDF (2 regions)GL PDF (2 regions)
52 Image SegmentationImage Segmentation52 Image SegmentationImage Segmentation
Use of boundary characteristics for histogram improvement and local thresholding good threshold are enhanced considerably if the histogram peaks are tall, narrow, and separated by deep valleysImprove the shape of histogram by considering only those pixels that lie on or near the edges
If h f h l h d b
53 Image SegmentationImage Segmentation
If histogram of the pixel on or near the edge between objects and background were used: improve the symmetry of the histogram peak The histograms have peaks of approximately the same
height Equal probability: improve the symmetry of the
histogram peak2
2
0( , ) and 0
and 0
if f Ts x y if f T f
if f T f
-
10/16/2009
10
ExamplesExamples
55 Image SegmentationImage Segmentation55 Image SegmentationImage Segmentation
ExampleExample
56 Image SegmentationImage Segmentation56 Image SegmentationImage Segmentation
Threshold based on several Threshold based on several variablesvariables
Multi-spectral thresholding 3-D histogram
Find clusters of points in 3-D space
ShortcomingCluster seeking becomes an increasingly complex
57 Image SegmentationImage Segmentation
Cluster seeking becomes an increasingly complex
ExampleExample
58 Image SegmentationImage Segmentation58 Image SegmentationImage Segmentation
Region-based segmentationBasic formulation
segmentation must be complete;every pixel must be in a regionpoints in a region must be connectedthe region must be disjointedthe pixels in a segmented region (P(Ri)=TRUE) if all pixel in Ri have the same gray levelP(Ri Rj)= False --Ri an Rj is different
Region growing
59 Image SegmentationImage Segmentation
a procedure that groups pixel or sub-region into larger regions based on predefined start with a set of seed criteria-similar to the seed (specific gray-level or color) points and from these grow regions by appending each seed priori information is not available--compute the same set of
properties Criteriagray-level, texture, color the selection of similarity
For Ex: land-use satellite imagery depends on color
when the images are monochrome, carried out with set of descriptor based on gray levels and spatial properties (such as moments or textures)
The stopping rule when on more pixels satisfy the criteria additional criteria
utilize size, likeness between a candidate pixel and the pixels grown so far, and the shape of the region being grown (history of the growth)
60 Image SegmentationImage Segmentation
seed points and seed regions (Fig. 10.40)
If connectivity or adjacency information is not used, descriptor alone can yield misleading result
Region splitting and mergingSplitting subdivide the entire region successively into smaller and small
quadrant regions if P(R)=FALSE quad tree -- the root of the tree corresponds to the entire image
and each node corresponds to a subdivision
-
10/16/2009
11
only splitting causes the final partition would contain adjacent regions with identical properties
can be solved by mergingMerging Is limited to groups of four blocks the splitting and merging steps
split into four disjoint quadrants
61 Image SegmentationImage Segmentation
split into four disjoint quadrants merge adjacent adjacent regionsstop when no further merging or splitting
Variations of the schemeSplit the image into a set of blocks
The advantage: use the same quad-tree for splitting and merging Texture segmentation
EXAMPLEEXAMPLE(a) Image showing
defective welds(b) Seed points(c) Result of region
(a) (b)
(c) Result of region growing
(d) Boundaries of segmented defective welds (in black)
(e) Histogram of (a)
(c) (d)
(e)
QUADTREEQUADTREE(a) Partitioned image(b) Corresponding quadtree
(a) (b)
Split & MergeSplit & Merge
64 Image SegmentationImage Segmentation64 Image SegmentationImage Segmentation
(a) Original image; (b) Result of split and merge procedure; (c) Result of thresholding (a)
No further merging The procedure is terminated by one final merging of regions
satisfying step2 The merged regions may be of different size Texture segmentation is based on the measures of texture
Segmentation by morphological watershedsAdvantage of global thresholding: speeddisadvantages of traditional segmentation: (1) the need to
65 Image SegmentationImage Segmentation
disadvantages of traditional segmentation: (1) the need to post-processing; (2) need edge linking for discontinuity The watershed segmentation often produces stable segmentation results (continuous segmentation boundaries)
Basic conceptsBased on visualizing an image in three-dimension:two spatial coordinates versus gray levelsTopographic interpretation (three types of points):
(a) Points belonging to a regional minimum (b) Points at which of a drop of water (catchment
basin or watershed) (c) Points at which water would be equally likely to
fall to more than one such minimum (crest line on the topographic surface; divide line or watershed lines)
66 Image SegmentationImage Segmentation
The principal objective is to find the watershed linesSuppose that a hole is punched in each regional minimum and the entire topography is flooded Water rise through the holes Rising water in distinct catchment basin is about to
merge A dam is built to prevent the merging Reach when the tops of the dam are visible
-
10/16/2009
12
Dam constructionDam constructionBased on binary imagesThe simplest way Construct dams separating sets of binary points based on
morphological operationsUse dilationDef. of catchment basin and the sets of points in two regional
i i t t 1 C (M ) d C (M )
67 Image SegmentationImage Segmentation
minimum at stage n-1: Cn-1(M2) and Cn-1(M2)The two individual components extracted from q by performing the simple AND operation qC[n-1] becomes one connected componentEach of the connected components is dilated by the structuring element, subject to two conditions: The dilation has to be constrained to q The dilation cannot be performed on points that would cause
the sets being dilated to merge
FloodingFlooding
68 Image SegmentationImage Segmentation68 Image SegmentationImage Segmentation
(a) Original image(b) Topographic view(c) - (d) Two slopes of flooding
e) Result of further flooding
f) Beginning of merging of water from two catchment basins (a short dam was built between them)
69
Flooding (continued )Flooding (continued )
g) Longer damsh) Final watershed
(segmentation)
Flooding (contd)Flooding (contd)a) Two partially flooded
catchment basins at stage n 1 of flooding,
b) Flooding at stage n, showing that water has
ill d b b i
70