ImageJ
Java image processing tool from NIH
[Link]
Reads / writes a large variety of images
Many image processing operations are implemented
Good rapid prototyping / testing tool
Includes the ability to write your own plugins
Convenient way to implement your own algorithms, using
their image class and i/o methods
1 Ellen L. Walker
Point Processes
Pixel by pixel transformation
Output pixel depends only on corresponding input pixel
Examples:
Out(r,c) = In(r,c) * 1.25
Out(r,c) = In(r,c) + 25
Out(r,c) = (In(r,c))2
2 Ellen L. Walker
Linear Transformations
Out(r,c) = In(r,c) * gain + bias
Gain controls contrast
Bias controls brightness
Location dependent:
Out(r,c) = In(r,c) * gain(r,c) + bias(r,c)
E.g. “sky darkening filter”
Linear Blend
Out(r,c) = (lambda) * In1(r,c) + (1-lambda) * In2(r,c)
If In1 and In2 are images, a sequence of these from
lambda = 0 to lambda=1 is an image dissolve
3 Ellen L. Walker
Histogram
An image histogram counts the number of pixels at each brightness.
Color images have 3 histograms (red, green, blue)
4
Series2
3
0
0 1 2 3
4 Ellen L. Walker
Information in Histogram
Contrast (variation in brightness)
Are bars spread over the whole range?
Foreground vs. background color
Are there two separate “peaks” with a valley between?
5 Ellen L. Walker
Applications of Histogram
Thresholding
Find a value that separates foreground / background
values
Look for a “valley” between two peaks (may or may not be
what you need)
Contrast enhancement
Histogram equalization – spread the data as evenly
through the histogram as possible
Goal: wide, flat histogram
6 Ellen L. Walker
Example: Histogram Equalization
Original
Modified
7 Ellen L. Walker
Algorithm: Histogram Equalization
Find cumulative distribution
For each intensity I, c(I) = # pixels <= I
I
1 1
c(I) = ∑ h(i) = c(I −1) + h(I)
N i= 0 N
Code:
€ C[0]=Hist[0];
For(greylevel =1; greylevel < max; greylevel++){
C[greylevel] = Hist[greylevel] + C[greylevel-1];
C[greylevel] = C[greylevel] / (double)(rows*cols);
}
8 Ellen L. Walker
Algorithm: Histogram Equalization
Use C(I) as a lookup table to determine the final value of
each pixel.
Since C(I) ranges from 0 to 1 (why?), multiply C(I) by the
max pixel value to get the output value
Code:
For(r=0;r<rows;r++)
for(c=0;c<cols;c++)
out[r][c] = C[(in[r][c])]*MAX;
9 Ellen L. Walker
Locally Adaptive Histogram Equalization
Instead of doing histogram equalization using the whole
image, compute the distribution for a moving window
around the given pixel.
Avoids effect of bright light at one corner washing out
everything in the image.
10 Ellen L. Walker
Image Neighborhoods
Neighborhoods can be defined for each pixel
The two most common neighborhoods
4-neighborhood
N
W E
S
8-neighborhood
11 Ellen L. Walker
Applying a Mask
Mask is a set of relative pixel positions. One is
designated the origin (0,0) - usually at center
Each mask element is weighted
To apply the mask, put the origin pixel over the image
pixel and multiply weights by the pixels under them, then
add up all the values.
Usually this is repeated for every pixel in the image .
Assumptions must be made for pixels near the edge of
the image.
12 Ellen L. Walker
Mask application example
Mask = 1 1 1
Apply to every pixel Result is
in image: 00011
00001 00122
00011 01232
00111 12332
01111 23332
11111 Boundary pixels are gray
13 Ellen L. Walker
Mathematical Representation of Mask Operations
Equation from Chapter 3
g(i, j) = ∑ f (i + k, j + l)h(k,l)
k,l
g is the output image
f is the input image
€ h is the mask (also called kernel)
Short form (convolution operator)
g= f ⊗h
14 Ellen L. Walker
Masks that "blur"
"Box mask" - every pixel gets the average of its
neighborhood
111 After computing, divide by 9 (mask sum)
111 to keep image from getting too bright
111
"Weighted average" - divide by 16 after application
121
242
121
15 Ellen L. Walker
Why blur?
Avoid effects of small random noise (“salt and pepper”)
Remove small features to emphasize larger ones
Bigger masks blur more / remove larger features
Sequence of masks generates sequence of increasingly
blurred images (useful for some matching algorithms)
First step in sharpening the image (!)
Sharp(x,y) =
orig(x,y) + gamma (orig(x,y) – (blur * orig(x,y)))
16 Ellen L. Walker
Boundary Effects (padding)
Figure 3.12
17 Ellen L. Walker
Common Masks
Figure 3.13
18 Ellen L. Walker
Median Filtering
Example of a non-linear filter
Replace the central pixel of a window with the median
pixel of the window
Compare to box filter, which replaces with average value
in the window
Tends to preserve edges (why?)
19 Ellen L. Walker