digital image processing - sharif university of...
TRANSCRIPT
Outline
Point operations
Histogram modeling
Spatial operations
Transform operations
Multispectral image enhancement
2
Point operations
Zero memory operations
Given gray-level is mapped to a gray-level according
to a transformation
3
𝒖 ∈ [𝟎, 𝑳] 𝒗 ∈ [𝟎, 𝑳]
𝒗 = 𝒇(𝒖)
Point operations
Contrast stretching
Noise clipping and thresholding
Gray scale reversal
Gray-level window slicing
Bit extraction
Range compression
Contrast stretching 4
Original image Enhanced image
𝒗 = 𝒇(𝒖)
Image histogram
𝑢
𝑣
Transformation function
Thresholding 7
Original image Image after thresholding 𝒗 = 𝒇(𝒖)
Image histogram
𝑢
𝑣
Transformation function
Thresholding 8
Original image Image after thresholding 𝒗 = 𝒇(𝒖)
Image histogram
𝑢
𝑣
Transformation function
Gray-level window slicing 11
Original image Image after window slicing
Transformation function
Transformation function
With background
Gray-level window slicing 12
Original image Image after window slicing
Transformation function
Without background
Bit extraction 13 Example:
intensity = 132
8 7 6 5 4 3 2 1
1 0 0 0 0 1 0 0
Original
image 1 2
3 4 5
6 7 8
Bit extraction 14
Can be used for image compression
Reconstruction using bits number 6, 7, and 8 Reconstruction using bits number 7 and 8
Range compression 15
image 2-D Fourier transform Log of 2-D Fourier transform
Dynamic range very large
Can be compressed via the logarithmic transformation to be more visible
Enhances the low pixel values at the expense of loss of information in the high
pixel values
Range compression 16
image 2-D Fourier transform Log of 2-D Fourier transform
What if the compressing image has important high valued pixels?
Range compression 17
image 2-D Fourier transform
What if the compressing image has important high valued pixels?
? Desired output must be
something like this
Scale down the Fourier image before
applying the logarithmic transform
Histogram modeling 18
Histogram-modeling techniques modify an image so that its
histogram has a desired shape
Useful in stretching the low-contrast levels of images with narrow
histograms
Histogram modeling
Histogram equalization
Histogram modification
Histogram specification
Histogram equalization 19
Desired output image histogram: Uniform histogram
is the estimated probability distribution of discrete random variable ,
(indicating image pixel value) is determined using the histogram of the image
is the cumulative probability distribution of which is uniformly
distributed over
Corresponding output pixel value of each pixel value is , which is
followed by a normalization bellow to lie in
is the smallest value of between all obtained values
Histogram equalization 20
Original image Equalized image
Original image Equalized image
histogram Approximation of
uniform histogram
histogram Approximation of
uniform histogram
Histogram modification 21
Instead of cumulative probability distribution we can use other
distributions, such as
And apply a normalization
This approach is called histogram modification
Histogram specification 23
Desired output histogram: Any arbitrary histogram!
: Probability distribution of input image pixel value, approximated from it’s
histogram
: Probability distribution of output image pixel value, approximated from
desired output histogram
If and :
to obtain output value corresponding to input value :
First: Find such that for the smallest value of
Then:
Histogram specification 24
Original image Specified image
histogram Specified histogram
Desired output image histogram: Any arbitrary histogram!
Desired histogram
Spatial operations 25
Spatial operations
Spatial averaging and spatial low-pass filtering
Directional smoothing
Median filtering
Other smoothing techniques
Unsharp masking and crispening
Spatial low-pass, high-pass, and band-pass filtering
Inverse contrast ratio mapping and statistical scaling
Magnification and interpolation (Zooming)
Spatial averaging and spatial low-pass filtering 26
Each pixel is replaced by a weighted average of its neighborhood pixel
A common class has all equal weights
Why is it low-pass filtering?
Original image 3 * 3 average filter Image with Guassian noise
Directional smoothing 27
Directional averaging filter can be used to protect the edges from blurring
is selected such that the difference between input pixel value and the
average value of the pixels in the neighborhood window corresponding to
. is minimum
Median filter 28
Input pixel is replaced by the median of the pixels contained in a window
around the pixel
Original image 3 * 3 median filter Image with binary noise
Other smoothing techniques 29
Can we use averaging filter to remove binary noise?
Yes, but we need a threshold to determine whether we should replace the
pixel value with the averaged value or not
For additive Gaussian noise more sophisticated smoothing algorithms
are possible
What if image noise is multiplicative?
Unsharp masking and crispening 30
Unsharp masking is used commonly in printing industry for crispening the edges
An unsharp or low-pass filtered signal of the image subtracted form the image,
or equally high-pass signal or gradient of the image is obtained
The result is added to the original signal with a factor to crispen the edges
A commonly used gradient function is the discrete Laplacian:
2-D signal:
1-D signal:
Unsharp masking and crispening 31
Example of 1-D signal
Result of crispening
Original signal
Discrete Laplacian
Spatial low-pass, high-pass, and band-pass filtering 32
As mentioned earlier, spatial averaging is low-pass filter, because it somoothes the image
edges
Spatial averaging,
Low-pass filter
High-pass filter
Original signal
Band-pass filter
Spatial low-pass, high-pass, and band-pass filtering 33
Original image Low-pass filter,
k1 = 10
High-pass filter,
k2 = 10
Band-pass filter,
k1 = 10, k2 = 20
Inverse contrast ratio mapping and statistical scaling 34
Inverse contrast ratio transformation
Where and are the local mean and standard deviation of
. measured over a neighborhood window
Generally high contrast ratio for an object results in more
detectability of it. Therefor this transformation enhances week edges.
A special case for this transformation is also called
statistical scaling
35
Original image Inverse contrast ratio transformation,
with k = 3, and
followed by a log transformation
Inverse contrast ratio mapping and statistical scaling
Magnification and interpolation (Zooming) 36
Replication
Each pixel in each row is repeated one, then each resulting row is repeated
Equally: Interlace the image by rows and columns of zeros and then convolve
the result with
Linear interpolation
Straight line is first fitted in between pixels along a row. Then pixels along each
column are interpolated along a straight line
Equally: Interlace the image by rows and columns of zeros and then convolve
the result with
Transform operations 39
Transform operations
Generalized linear filtering
Root filtering
Generalized Cepstrum and homomorphic filtering
In the transform operation enhancement techniques, zero-memory
operations are performed on a transformed image, followed by the inverse
transformation
Generalized linear filtering 40
In this case the zero-memory transform domain operation is a pixel-by-pixel
multiplication
is called a zonal mask
Example:
DFT zonal mask
Spatial low-pass, high-pass, and band-pass filtering 41
Original image Low-pass filter,
a = 25
High-pass filter,
b = 40
Band-pass filter,
a = 25, b = 40
DFT transformation
Root filtering 42
The transform coefficient can be written as
In root filtering the zero-memory operation is
Multispectral image enhancement 44
Multispectral image
enhancement
Intensity ratios
Log-ratios
Principal components
In multispectral imaging given a sequence of images, it is desired to
combine these images to generate a single or a few display images that
are representative of their features.
Intensity ratios 45
Define the ratios
This method gives combinations for the ratios
So, the result is a multispectral image with spectral bands
Log-ratios 46
Taking the log of both sides of pervious equation
Log-ratio gives a better display when the dynamic range of is
very large
Principal components 47
We define
Matrix is determined from the autocorrelation matrix of the ensemble of
vectors . The rows of , which are eigenvectors of
the autocorrelation matrix are arranged in decreasing order of their
associated eigenvalues
is resulting multispectral image