Chapter 10 Image Segmentation -...
Transcript of Chapter 10 Image Segmentation -...
![Page 1: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/1.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Chapter 10 Image Segmentation
Generally, image segmentation algorithm are based on one of two basic properties of intensity values: 1. Discontinuity - Point detection - Line detection - Edge detection 2. Similarity - Thresholding - Region growing - Region spitting and merging
![Page 2: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/2.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
A General 3x3 Mask
The response of the mask with respect to its center location is
∑=
+++=
=
9
1
992211 ........
iii zw
zwzwzwR
iz is the gray level of the pixel associated with mask coefficient iw
![Page 3: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/3.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Point Detection
If ,where T is a nonnegative threshold, then a point is detected.
TR ≥
Point detector mask
max%90 ×=T
![Page 4: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/4.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Line Detection
Line Masks
![Page 5: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/5.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Illustration of Line Mask
Thresholded image Threshold =max. value in the left image
45− line detection
Absolute value of result after using
45− line detector
![Page 6: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/6.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Edge Detection
An Ideal Digital Edge v.s. A Ramp Digital Edge
The slope of the ramp is inversely proportional to the degree of blurring in the edge. Blurred edges tend to be thick and sharp edges tend to be thin.
![Page 7: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/7.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Edge Detection
The magnitude of the first derivative can be used to detect if a point is on the ramp. The sign of the second derivative can be used to determine whether an edge pixel lies on the dark or light side of an edge.
![Page 8: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/8.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Edge Detection
Two additional properties of the second derivative: 1. It produces two values for every edge in an image. 2. An imaginary straight line joining the extreme positive and negative values of the second derivative would cross zero near the midpoint of the edge. (can be used to locate the centers of thick edges)
![Page 9: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/9.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
First and Second Derivatives around a Noisy Edge
Free of noise
Corrupted by additive Gaussian noise, m=0, =0.1 σ
Corrupted by additive Gaussian noise, m=0, =1.0
σ
Corrupted by additive Gaussian noise, m=0, =10.0
σ
1st order derivative
2nd order derivative
The second order derivative is sensitive to the noise
![Page 10: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/10.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Gradient Operators
First-order derivatives in an image are computed using the gradient. The gradient of an image f(x,y) at location (x,y) is defined as:
∂∂∂∂
=
=∇
yfxf
GG
y
xf
The magnitude of this vector is:
21
22 ][)( yx GGmagf +=∇=∇ fThe direction of the gradient vector:
)(tan),( 1
x
y
GG
yx −=α
yx GGf +≈∇
![Page 11: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/11.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Gradient Operators
A 3x3 area in an image
59 zzGx −= 68 zzGy −=
)()()()(
741963
321987
zzzzzzGzzzzzzG
y
x++−++=++−++=
)2()2()2()2(
741963
321987
zzzzzzGzzzzzzG
y
x++−++=++−++=
![Page 12: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/12.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Masks for Detecting Diagonal Edges
![Page 13: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/13.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Illustrations of the Gradient and Its Components
Original image xG
yG yx GG +
![Page 14: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/14.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Illustrations of the Gradient and Its Components
Image smoothed by 5x5 averaging filter xG
yG yx GG +
To smooth the contribution made by the wall bricks
![Page 15: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/15.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
An Example of Diagonal Edge Detection
Using Diagonal Sobel Masks
![Page 16: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/16.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
The Laplacian
Second-order derivatives in an image are obtained using the Laplacian.
The Laplacian of f(x,y) is defined as:
2
2
2
22
yf
xff
∂∂
+∂∂
=∇
![Page 17: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/17.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Laplacian Masks
)(4 864252 zzzzzf +++−=∇
)(8 98765432152 zzzzzzzzzzf ++++++++−=∇
![Page 18: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/18.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
The Role of the Laplacian in Segmentation
1. Using its zero-crossing property for edge location.
2. Using it for the complementary purpose of establishing whether a pixel is on the dark or light side of an edge.
![Page 19: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/19.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Edge Finding by Zero Crossing
Gaussian function:
2
2
2)( σr
erh−
−=
222 yxr +=where, and is the standard deviation σ
The Laplacian of Gaussian (LoG) is:
2
2
24
222 )( σ
σσ
r
errh−
−−=∇
The purpose of Gaussian is to smooth the image. The purpose of the Laplacian operator is to provide an image with zero crossing for establishing the location of edges.
![Page 20: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/20.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Laplacian of Gaussian
3-D plot
Zero crossing
5x5 approximation mask
![Page 21: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/21.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Edge Finding by Zero Crossing
Original image Sobel gradient (for comparison)
Gaussian smoothing function Laplacian mask
LoG
Thresholded LoG
Zero crossing
![Page 22: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/22.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Edge Linking and Boundary Detection
The edge detection algorithms typically are followed by linking procedures to assemble edge pixels into meaningful edges. There are several approaches for this purpose: -Local processing -Global processing via the Hough Transform -Global processing via Graph-theoretic techniques
![Page 23: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/23.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Local Processing
An edge pixel with coordinates (xo,yo) in a predefined neighborhood of (x,y), is similar in magnitude to the pixel at (x,y) if:
Eyxfyxf ≤∇−∇ ),(),( 00where E is a nonnegative threshold
An edge pixel at (xo,yo) in the predefined neighborhood of (x,y) has an angle similar to the pixel at (x,y) if
Ayxyx <− ),(),( 00ααwhere A is a nonnegative threshold The direction of the edge at (x,y) is perpendicular to the direction of the gradient vector at that point.
All points that are similar according to these predefined criteria are linked, forming an edge of pixels that share the criteria.
![Page 24: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/24.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Edge –Point Linking Based on Local Processing
Original image xG
yG Result of edge linking
![Page 25: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/25.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Global Processing via the Hough Transform
xy-plane ab-plane (parameter plane)
a’ is the slope and b’ the intercept of the line containing both (xi,yi) and (xj,yj) in xy-plane.
All points on this line have lines in parameter space that intersect at (a’,b’)
![Page 26: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/26.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Global Processing via the Hough Transform
Accumulator cells
(amax,amin) and (bmax,bmin) are the expected ranges of slope and intercept values.
The cell at coordinates (i,j), with accumulator value A(i,j), corresponds to the square associated with parameter space coordinates (ai,bi). Initially, A(i,j)=0. 1. For every point (xk,yk) in the image, let a equal each of the allowed subdivision values on the a-axis and solve b using b=axk + yk, then round off b to its nearest value in b-axis. 2. If ap results in bq, then A(p,q)=A(p,q)+1. 3. A value of Q in A(i,j) corresponds to Q points in the xy-plane lying on the line y= aix+bj.
![Page 27: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/27.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Global Processing via the Hough Transform
The normal representation of a line is:
ρθθ =+ sincos yx
Subdivision into cells
sinusoidal curves in the -plane instead of straight lines in the ab-plane.
ρθ
![Page 28: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/28.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Illustration of the Hough Transform
Image with 5 labeled points
5 points mapped onto the -plane ρθ
“A”: points 1,3 and 5 lie on a straight line passing through =0 and =-45o
“B”: points 2,3 and 4 lie on a straight line passing through =1/2 diagonal distance, = 45o
ρ θ
ρθ
Hough transform has a reflective adjacency relationship at the right and left edge of the parameter space.
![Page 29: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/29.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Using the Hough Transform for Edge Linking
1. Compute the gradient of an image and threshold it to obtain a binary image. 2. Specify subdivisions in the -plane. 3. Examine the counts of the accumulator cells for high pixel concentrations. 4. Examine the relationship (principally for continuity) between pixels in a chosen cell. Continuity: computing the distance between disconnected pixels identified during traversal of the set of pixels corresponding to a given accumulator cell. A gap at any point is significant if the distance between that point and its closest neighbor exceeds a certain threshold.
ρθ
![Page 30: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/30.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Using the Hough Transform for Edge Linking
Original image Thresholded gradient image
Hough transform of thresholded image
Result of edge linking
The criteria for linking pixels: 1. The pixels belonged to one of the three accumulator cells with the highest count. 2. No gaps were longer than five pixels.
![Page 31: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/31.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Global Processing via Graph-Theoretic Techniques
This method is based on representing edge segments in the form of a graph and searching the graph for low-cost paths that correspond to significant edges. It performs well in the presence of noise. Definitions: 1. Graph G=(N,U): a finite,nonempty set of nodes N, together with a set U of unordered pairs of distinct elements of N. 2. Arc: each pair of U. A cost is associated with it. 3. Directed graph: a graph in which the arcs are directed. 4. Successor, parent: if an arc is directed from node to node , then is said to be a successor of the parent node . 5. Path: a sequence of nodes , with each node being a successor of node , is called a path from to . 6. The cost of the path:
),( ji nn
in jninjn
knnn ,....,, 21 in1−in 1n kn
),(2
1∑==
−k
iii nncc
),( ji nnc
![Page 32: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/32.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Global Processing via Graph-Theoretic Techniques
Edge element: boundary between two pixels p and q, such that p and q are 4-neighbors.
The edge element here is defined by the pairs ),)(,( qqpp yxyx
![Page 33: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/33.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Global Processing via Graph-Theoretic Techniques
Gray level value
The cost for each edge element defined by pixels p and q: )]()([),( qfpfHqpc −−= H: the highest gray-level value in the image
f(p), f(q): gray level values of p and q.
Cost
Edge: the lowest-cost path
![Page 34: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/34.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Global Processing via Graph-Theoretic Techniques
Graph for Fig.10.23
The lowest-cost path is shown dashed.
![Page 35: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/35.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
The Heuristic Graph Search Algorithm
Let r(n) be an estimate of the cost of a minimum-cost path from the start node s to a goal node, where the path is constrained to go through n. r(n)=g(n)+h(n), g(n) is the lowest-cost path from s to n, h(n) is obtained by using any available heuristic information. Step1. Mark the start node OPEN and get g(s)=0. Step 2. If no node is OPEN exit with failure; otherwise, continue. Step 3. Mark CLOSED the OPEN node n whose estimate r(n) computed is smallest. Step 4. If n is a goal node, exit with the solution path obtained by tracing back through the pointers; otherwise, continue.
![Page 36: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/36.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
The Heuristic Graph Search Algorithm
Step 5. Expand node n, generating all of its successors. (If there are no successors go to Step 2.) Step 6. If a successor ni is not marked, set mark it OPEN, and direct pointers from it back to n. Step 7. If a successor ni is marked CLOSED or OPEN, update its value by letting
),()()( ii nncngnr +=
)],()(),(min[)(' iii nncngngng +=
mark OPEN those CLOSED successors whose g’ values were thus lowered and redirect to n the pointers from all nodes whose g’ were lowered. Go to step 2. This algorithm does not guarantee a minimum-cost path, but it has higher speed.
![Page 37: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/37.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Edge Finding by Graph Search
The heuristic used at any point on the graph was to determine and use the optimum path for five levels down from that point.
![Page 38: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/38.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Thresholding
(a) Single thresholding if , then (x,y) is called an object point. (b) Multilevel thresholding if , then (x,y) belongs to one object. if , then (x,y) belongs to another object. if , then (x,y) belongs to the background.
Tyxf >),(
21 ),( TyxfT ≤<2),( Tyxf >
1),( Tyxf ≤
![Page 39: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/39.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Thresholding
Thresholding may be viewed as an operation that involves tests against a function T of the form:
)],(),,(,,[ yxfyxpyxTT =
f(x,y): gray level of point (x,y) p(x,y): a local property of (x,y) The thresholded image:
=01
),( yxgif Tyxf >),(if Tyxf ≤),(
Global thresholding: T depends only on f(x,y) Local thresholding: T depends on both f(x,y) and p(x,y) Dynamic/Adaptive thresholding: T depends on x and y
![Page 40: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/40.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Example Showing the Role of Illumination on Segment
The image resulting from poor (e.g. nonuniform) illumination could be quite difficult to segment.
![Page 41: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/41.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Basic Global Thresholding
Original image
Thresholded image T is the midway between the max. and min. gray levels.
![Page 42: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/42.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Algorithm to Obtain T Automatically
1. Select an initial estimate for T. 2. Segment the image using T. - G1 consists of all pixels with gray level values>T - G2 consists of all pixels with gray level values T 3. Compute a new threshold value:
≤
)(21
21 µµ +=T
where and are the average gray level values for the pixels in regions G1 and G2 respectively.
1µ 2µ
4. Repeat 2 to 3 until the difference in T in successive iterations is smaller than a predefined parameter
T
![Page 43: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/43.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Image Segmentation Using an Estimated Global Thresholding
Original image
Thresholded image 3 iterations T=125
![Page 44: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/44.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Basic Adaptive Thresholding Original image with uneven illumination Result of global threshold
Image subdivided into individual subimages Result of adaptive thresholding
![Page 45: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/45.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Basic Adaptive Thresholding
Properly segmented subimage
Improperly segmented subimage
Subdivided the above sub- image into smaller subimages
Result of adaptively segmenting the left image
![Page 46: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/46.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Optimal Global and Adaptive Thresholding
p1(z), p2(z): probability density function (PDF) of the objects gray levels and background gray levels.
The PDF of the overall gray level variation in the image is: )()()( 2211 zpPzpPzp +=
: the probability that a random pixel with value z is an object pixel. : the probability that a random pixel with value z is a background pixel.
1P2P
121 =+ PP
Estimating thresholds that produce the minimum average segmentation error
![Page 47: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/47.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Optimal Global and Adaptive Thresholding
Let T be the threshold. The probability of erroneously classifying a background point as an object point is: The probability of erroneously classifying an object point as background is:
∫= ∞−T dzzpTE )()( 21
∫=∞T dzzpTE )()( 12
The overall probability of error is: )()()( 2112 TEPTEPTE +=
The threshold value for which this error is minimal: )()( 2211 TpPTpP =
![Page 48: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/48.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Optimal Global and Adaptive Thresholding
If we use Gaussian density, then
2
2
2
2
2
1
2
1
2)(
2
22)(
1
122
)( σµ
σµ
σπσπ
−−
−−
+=
zz
ePePzp
The solution for the threshold T:
02 =++ CBTATwhere
)ln(2
)(2
21
1222
21
21
22
22
21
212
221
22
21
PPC
B
A
σσσσµσµσ
σµσµ
σσ
+−=
−=
−=
![Page 49: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/49.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Optimal Global and Adaptive Thresholding
If 22
21
2 σσσ ==
then
−
++
=1
2
21
221 ln
2 PPT
µµσµµ
![Page 50: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/50.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Use of Optimum Thresholding for Image Segmentation
cardioangiogram cardioangiogram before preprocessing after preprocessing
3 preprocessing steps: 1. Log function: counter exponential effects caused by radioactive absorption. 2. (image captured after the medium was injected) – (image captured before the medium was injected): remove the spinal column present in both images. 3. Several images were summed: reduce random noise.
![Page 51: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/51.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Use of Optimum Thresholding for Image Segmentation
black dots: histogram of region A in Fig.10.33(b). “o’s” and “x’s”: two fits for the histogram by bimodal Gaussian density curves. Then the optimum thresholds were obtained by the equations for the Gaussian curves.
histogram of region B in Fig.10.33(b)
![Page 52: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/52.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Use of Optimum Thresholding for Image Segmentation
Boundaries superimposed on the original image
Boundaries were obtained by: 1. Obtaining the binary picture.
=01
),( yxf xyTyxf ≥),(
otherwise2. Taking the gradient of the binary picture.
![Page 53: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/53.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Use of Boundary Characteristics for Histogram Improvement and Local Thresholding
Image is coded by the following equation:
−+=0
),( yxsif
if if
Tf <∇
Tf ≥∇ and 02 ≥∇ f02 <∇ f
For a dark image in a light background, it results in: 1. All pixels that are not on an edge are labeled “0”. 2. All pixels on the dark side of an edge are labeled “+”. 3. All pixels on the light side of an edge are labeled “-”. The “+” and “-” will be reversed if a light object in on a dark background.
Tf ≥∇ and
![Page 54: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/54.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Image Segmentation by Local Thresholding
Original image
Image segmented by local thresholding (T at or near the midpoint of the valley shown in Fig.10.38)
Histogram of pixels with gradients greater than 5.
![Page 55: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/55.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Thresholds Based on Several Variables
![Page 56: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/56.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Region-Based Segmentation
Let R represent the entire image region. We may view segmentation as a process that partitions R into n subregions, R1,R2,……,Rn, such that:
.1
RRn
ii =
=
(a)
(b) iR is a connected region, i=1,2,….,n.
(c) φ=∩ ji RR for all i and j, i j. ≠
(
(d) TRUERP i =)( for i=1,2,…,n.
(e) FAULSERRP ji =∪ )( for i j. ≠
Here, )( iRP is a logical predicate defined over the points in set and iRφ is the null set.
![Page 57: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/57.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Region Growing
Original image Seed points: gray level=255
Result of region growing Boundaries of segmented defective welds
Criteria for a pixel to be annexed to a region: (1) The absolute gray- level difference between any pixel and the seed < 65. (2) The pixel had to be 8-connected to at least one pixel in that region.
![Page 58: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/58.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Region Growing
Histogram of Fig. 10.40(a)
![Page 59: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/59.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Region Splitting and Merging
1. Split into four disjoint quadrants any region Ri for which P(Ri)=FALSE. 2. Merge any adjacent regions Rj and Rk for which P(Rj Rk)=TRUE. 3. Stop when no further merging or splitting is possible.
∪
![Page 60: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/60.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Region Splitting and Merging
(a) original image (b)result of splitting (c)result of and merging thresholding (a)
Here, define P(Ri)=TRUE if at least 80% of the pixels in Ri have the property iij mz σ2≤− , where is the gray level of the jth pixel in Ri, is the mean
gray level of that region, and is the standard deviation of the gray levels in Ri. If P(Ri)=TRUE under this condition, the values of all the pixels in Ri were set equal to mi.
jz imiσ
![Page 61: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/61.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Segmentation by Morphological Watersheds
Watershed is based on visualizing an image in three dimensions - two spatial coordinates versus gray levels. (topographic view) We consider three types of points: 1. Points belonging to a regional minimum. 2. Points at which a drop of water, if placed at the location of any of those points, would fall with certainty to a single minimum. - catchment basin or watershed of that minimum. 3. Points at which water would be equally likely to fall to more than one such minimum. - divide lines or watershed lines. The principal objectives of segmentation algorithms based on these concepts is to find the watershed lines.
![Page 62: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/62.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Segmentation by Morphological Watersheds Original image Topographic view
![Page 63: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/63.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Segmentation by Morphological Watersheds
Watershed lines - continuous boundaries
![Page 64: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/64.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Dam Construction
![Page 65: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/65.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Illustration of the Watershed Segmentation Algorithm
![Page 66: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/66.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Illustration of Oversegmentation
![Page 67: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/67.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
The Use of Markers
![Page 68: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/68.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
The Use of Motion in Segmentation - Spatial Techniques
ADI: Accumulative Difference Image – formed by comparing the reference image with every subsequent image in a sequence of image frames. A counter for each pixel location in the accumulative image is incremented every time a difference occurs at that pixel location between the reference and an image in the sequence. Three types of accumulative difference images: -Absolute -Positive -Negative
![Page 69: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/69.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
The Use of Motion in Segmentation - Spatial Techniques
Let f(x,y,k) denote the image at time tk , R(x,y)=f(x,y,1) denote the reference image. The values of the ADIs are counts. Assume that the gray-level values of the moving objects are larger than the background. Define:
+
=−
−
),(1),(
),(1
1
yxAyxA
yxAk
kk
TkyxfyxRif >− ),,(),(
otherwise
+
=−
−
),(1),(
),(1
1
yxPyxP
yxPk
kk
TkyxfyxRif >− )],,(),([otherwise
+
=−
−
),(1),(
),(1
1
yxNyxN
yxNk
kk
TkyxfyxRif −<− )],,(),([
otherwise
Absolute ADI
Positive ADI
Negative ADI
![Page 70: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/70.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
An Example of ADIs
ADIs of a rectangular object in a southeasterly direction
Absolute ADI Positive ADI Negative ADI
![Page 71: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/71.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
Building a Reference Image
![Page 72: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/72.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
The Use of Motion in Segmentation - Frequency Domain Techniques
For a sequence of K digital images of size M N, the sum of the weighted projections onto the x axis at any integer instant of time is:
×
∑ ∑=−
=
−
=
∆1
0
1
0
21
1),,(),(M
x
N
y
txajx etyxfatg π
The 1-D Fourier transform is:
KtujK
txx eatg
KauG /21
0111
1),(1),( π−−
=∑=
The sum of the weighted projections onto the y axis is:
∑ ∑=−
=
−
=
∆1
0
1
0
22
2),,(),(N
y
M
x
tyajy etyxfatg π
And the 1-D Fourier transform is: KtujK
tyy eatg
KauG /21
0222
2),(1),( π−−
=∑=
t = 0,1,…,K-1
t = 0,1,…,K-1
u1= 0,1,….,K-1
u2= 0,1,….,K-1
![Page 73: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/73.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
The Use of Motion in Segmentation - Frequency Domain Techniques
The frequency-velocity relationship is:
111 vau =
The sign of the x-component of the velocity is obtained by computing:
ntx
x dtatgdS == 2
12
1)],(Re[
222 vau =
ntx
x dtatgdS == 2
12
2)],(Im[
- If the velocity component v1 is positive, then S1x and S2x will have the same sign at an arbitrary point in time n. - If v1 is negative, then S1x and S2x will have the opposite sign. - If either S1x or S2x is zero, we consider the next closest point in time tnt ∆±=Similar comments apply to computing the sign of v2.
![Page 74: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/74.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
An Example of Detection of a Small Moving Object via the Frequency Domain
One of 32-frame sequence of LANDSAT images generated by adding white noise to a reference image.
![Page 75: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/75.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
An Example of Detection of a Small Moving Object via the Frequency Domain
Intensity plot of the previous image
![Page 76: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/76.jpg)
© 2002 R. C. Gonzalez & R. E. Woods u1=3 yields v1=0.5
a1= 6
An Example of Detection of a Small Moving Object via the Frequency Domain
![Page 77: Chapter 10 Image Segmentation - users.encs.concordia.causers.encs.concordia.ca/~weiping/ELEC6641-W14/Chapter 10.pdf · 2014. 4. 7. · to significant edges. It performs well in the](https://reader034.fdocuments.us/reader034/viewer/2022051822/5fecd72ee48fe13bad5f03d4/html5/thumbnails/77.jpg)
© 2002 R. C. Gonzalez & R. E. Woods
An Example of Detection of a Small Moving Object via the Frequency Domain
a2= 4
u2=4 yields v2=1.0