Mathematical Morphology
• Erosion
• Dilation
• Opening
• Closing
1
Erosion
• Used to reduces objects in the image
• Definition, binary images:
– Does the structure element fit the set?
B X ={x∣B x ⊆X }
where Bx means B translated with x,
X is the image, and
B is the structure element
2
Erosion
• Example, binary image
3
Erosion
• Definition, grayscale images
We remember the definition for binary images:
B X ={x∣B x ⊆X }
Can be rewritten into the intersections of the translated sets X-b:
B X = ∩b∈B X −b
Which can be extended to also include grayscale images:
B f = ∧b∈B f −b
4
Erosion
• Based on
B f = ∧b∈B f −b
we can define an algorithm to find the erosion of image f:
[ B f ] x=min b∈B f xb
5
Erosion
• Example, grayscale image
f f+1
6 f-1 εΒ(f)
Erosion example
Erosion using
a SE matching
the inner
texture
7
Another erosion example
• Branches are important in a finger print comparison
SE used to identify branches
8
Original image
Another erosion example
Thresholded image Erosion result
9
Another erosion example
10
Identified finger print branches
Dilation
• Used to increase objects in the image
• Definition, binary images:
– Does the structure element hit the set?
B X ={x∣B x ∩X ≠∅}
where Bx means B translated with x,
X is the image, and
B is the structure element
11
Dilation
• Example, binary image
12
Dilation
• Definition, grayscale images
We remember the definition for binary images:
B X ={x∣B x ∩X ≠∅}
Can be rewritten into the unions of the translated sets X-b:
B X = ∪b∈ B X −b
Which can be extended to also include grayscale images:
B f = ∨b∈B f −b
13
Dilation
• Based on
B f = ∨b∈B f −b
we can define an algorithm to find the dilation of image f:
[dilation B f ] x =max b∈B f xb
14
Dilation
• Example, grayscale image
f f+1
15 f-1 δΒ(f)
Dilation
• Example 2, grayscale image
16
Dilation and erosion example
17
Beucher gradient
• Dilation and erosion can be used to extract edge information from
images
– Example: Beucher gradient
B = B −B
18
Properties of dilation and erosion
• Dilation and erosion:
– are dual operators (C=complement):
B =C B C
– are translation invariant operators
– do not change the ordering relations between two images
– are invariant with respect to threshold decomposition
– Do not necessarily keep the homotopy of an image
• There exist no inverse transform to restore the original image
19
Dilation and erosion composition
B B =
2 1 B 2 B1
B B =
2 1 B 2
B1
• For instance can a dilation with a 5x5 SE instead be computed using two
3x3 SEs:
20
Improving calculation efficiency using
parallel computing
• Dilation and erosion are algorithms that can be efficiently parallelised
• Solutions:
– Large multiprocessor computer
– Computer cluster
– Cellular neural network (implemented in hardware)
– The graphics processing units (GPU) on a modern graphics card
– More++?
21
GPU
• The GPU on modern graphics cards has evolved into an extremely
flexible and powerful processor
– Cheap
– Programmable
– Precise
– Fast
• 3 Ghz Pentium 4 theoretical: 6 GFLOPS
• GeForceFX 5900 observed: 20 GFLOPS
– Improving rapidly (faster than Moore's law)
• CPUs: annual growth ≈1.5x
• GPUs: annual growth > 2.0x
22 Courtesy of Kurt Akeley et al., GPU Gems
GPU Performance
23
Graphs courtesy of Ian Buck, Stanford University
GPU shortcomings
• Shortcomings
– No better than 32 bit floating point throughout the pipeline
– GPU programming currently requires GPU-specific knowledge
– Not all image processing tasks can be implemented on the GPU
– If branches are not efficient (can sometimes be replaced by more efficient
max or min operators)
24
GPU example uses
• Example previous uses in image processing
– Image filtering (for instance blurring, edge enhancement, noise removal)
– Level set solver
– Registration
– Motion estimation
Example image blurring,
25 courtesy of Robert Strzodka, IEEE Visualization 2004 Tutorial
GPU and mathematical morphology
• Both dilation and erosion can be implemented efficiently on the GPU
• Instead of using if branches, built-in max and min operators can be used
• Example Cg code for implementing dilation on the GPU:
float4 dilate(half2 coords : TEX0, uniform sampler2D image) : COLOR {
static const half offset = 1.0 / 256.0;
float4 p=tex2D(image, coords);
float4 p1=tex2D(image, coords+half2(offset, 0));
float4 p2=tex2D(image, coords+half2(-offset, 0));
float4 p3=tex2D(image, coords+half2(0, offset));
float4 p4=tex2D(image, coords+half2(0, -offset));
return max(p, max(p1, max(p2, max(p3, p4))));
}
26
GPU and mathematical morphology
• Problem: GPUs do not have 3D frame buffers (cannot handle 3D data
volumes throughout the pipeline)
• Solution: Flat 3D arrays
27
Images courtesy of Aaron Lefohn, University of California
Morphological opening
• Used to remove unwanted structures in the image (e.g. noise)
• Morphological opening is simply an erosion followed by a dilation:
B f = B [ B f ]
• Binary example:
28
Morphological opening
• Grayscale example
29
Morphological closing
• Is used to merge or fill structures in an image
• Morphological closing is dilation followed by erosion:
B f = B [ B f ]
• Binary example:
30
Another example
• How to segment the text from the uneven illumination in the image?
31
Another example
• How to segment the text from the uneven illumination in the image?
32
Another example
• How to segment the text from the uneven illumination in the image?
33
Properties of opening and closing
• Opening and closing are duel operators
• Closing is an extensive transform
• Opening is an anti-extensive transform
• Opening and closing keep the ordering relation between two images
• Opening and closing are idempotent transforms
34
