1
Image Filtering (Ch. 4)
Images courtesy of Forsyth & Ponce.
It is the process of emphasizing/de-emphasizing certain spatial
properties of an image.
Linear filtering: When the intensity at a pixel in the output
is a linear combination of intensities of pixels in a small
neighborhood in the input.
H
This is mathematically defined by
H is assumed to be (2k+1)(2k+1). This operation is known as
correlation.
Convolution: The same as above but after flipping the filter
kernel. Note that when the filter kernel is symmetric, correlation and convolution become the same. Most filter kernels are
symmetric.
CEG-7550 Computer Vision
Fall 2014
Arthur Goshtasby
Convolution:
Example: When sum of the entries in H is 1, H is called a
smoothing filter. In addition, if all entries of H are the same, H
is called a mean filter.
Smoothing by a mean filter
When filter kernel is a 2-D Gaussian, the filter is called a
Gaussian filter.
CEG-7550 Computer Vision
Fall 2014
Arthur Goshtasby
A Gaussian filter
Smoothing by a Gaussian filter
Note that a 2-D Gaussian can be split into two 1-D Gaussians:
so that
Therefore, when a Gaussian is used, smoothing is performed
in 1-D, first row-by-row, and then column-by-column.
CEG-7550 Computer Vision
Fall 2014
Arthur Goshtasby
If a 2-D m xn filter can be separated into two 1-D filters, the
number of multiplications at each pixel can be reduced from
mn to m+n.
As the size of a filter kernel increases, more smoothing is
performed but at a proportionately higher computational
cost.
Difference operator: When the sum of entries of a filter
kernel is 0, the filter detects intensity changes in an image.
Example: When H = [1 0 -1], the filter responses to changes
in intensity as the image is scanned horizontally.
An image and its gradients in horizontal direction. Zero values are shown by 128 (gray).
This difference filter can be considered the first derivative of
mean filter [1 1]/2.
CEG-7550 Computer Vision
Fall 2014
Arthur Goshtasby
When the smoothing filter is a Gaussian, since
We have,
implying that we can convolve the image with the first
derivative of a 1-D Gaussian row-by-row, and convolve the
result with a 1-D Gaussian column-by-column.
Note: Convolution with a Gaussian de-emphasizes or blurs
intensity changes, while convolution with the first derivative
of a Gaussian enhances intensity changes. Also, note that
intensity changes can be dues to noise or image details.
Intensity derivative/changes horizontally of an image and its noisy versions.
Detecting intensity changes in arbitrary directions: Find
changes horizontally and vertically. Then, find direction and
magnitude of change from them.
Horizontal
change:
CEG-7550 Computer Vision
Fall 2014
Arthur Goshtasby
Vertical
change:
Gradient
magnitude:
Gradient
direction:
Iu and I v can be computed via convolution.
1. Convolve G u(u)G(v) with image I: Iu
2. Convolve G v(v)G(u) with image I: Iv
3. Calculate gradient magnitude and direction pixel-bypixel.
Therefore, we can calculate gradient magnitude and
direction by filtering.
CEG-7550 Computer Vision
Fall 2014
Arthur Goshtasby
Gaussian pyramid: An image pyramid where the bottom
image is the original image and images above it are smoothed
versions of the original image with an increasingly larger
Gaussian and resampling at a proportionately lower rate.
Image pyramids can be used to locate a pattern in an image
efficiently. A coarse but quick search is performed at the
lowest resolution to approximately locate the pattern. Then,
the localization accuracy is gradually improved by increasing
the resolution and searching locally.
CEG-7550 Computer Vision
Fall 2014
Arthur Goshtasby
