Digital Image Processing (CS/ECE 545) 

Lecture 4: Filters (Part 2) 
& Edges and Contours

Prof Emmanuel Agu
Computer Science Dept.
Worcester Polytechnic Institute (WPI)
 Recall: Applying Linear Filters: 
Convolution
For each image position I(u,v): 2. Multiply all filter coefficients H(i,j)
with corresponding pixel
I(u + i, v + j)
1. Move filter matrix H over
image such that H(0,0)
coincides with current image 3. Sum up results and store
position (u,v) sum in corresponding position
in new image I’(u, v)

Stated formally:

RH is set of all pixels

Covered by filter.
For 3x3 filter, this is:
 Recall: Mathematical Properties of 
Convolution
 Applying a filter as described called linear convolution
 For discrete 2D signal, convolution defined as: 
Recall: Properties of Convolution
Same result if we convolve
 Commutativity image with filter or vice versa

If image multiplied by scalar

 Linearity Result multiplied by same scalar

If 2 images added and convolve

(notice) result with a kernel H,
Same result if we each image
is convolved individually + added
 Associativity

Order of filter application irrelevant

Any order, same result
Properties of Convolution
 Separability

 If a kernel H can be separated into multiple smaller 
kernels  Applying smaller kernels H1 H2 … HN H one by one
computationally cheaper than apply 1 large kernel H

Computationally Computationally
More expensive Cheaper
Separability in x and y
 Sometimes we can separate a kernel into “vertical” 
and “horizontal” components
 Consider the kernels
 Complexity of x/y Separable Kernels

 What is the number of operations for 3 x 5 kernel H
Ans: 15wh

 What is the number of operations for Hx followed by Hy?
Ans: 3wh + 5wh = 8wh
 Complexity of x/y Separable Kernels

 What is the number of operations for 3 x 5 kernel H
Ans: 15wh

 What is the number of operations for Hx followed by Hy?
Ans: 3wh + 5wh = 8wh

 What about M x M kernel?
O(M2) – no separability (M2wh operations, grows quadratically!)
O(M2) – with separability (2Mwh operations, grows linearly!)
Gaussian Kernel
 1D

 2D
Separability of 2D Gaussian
 2D gaussian is just product of 1D gaussians:

Separable!
Separability of 2D Gaussian
 Consequently, convolution with a gaussian is 
separable

 Where G is the 2D discrete gaussian kernel; 
 Gx is “horizontal” and Gy is “vertical” 1D discrete 
Gaussian kernels
Impulse (or Dirac) Function
 In discrete 2D case, impulse function defined as:

 Impulse function on image?
 A white pixel at origin, on black background
Impulse (or Dirac) Function
 Impulse function neutral under convolution (no effect)
 Convolving an image using impulse function as filter = image
Impulse (or Dirac) Function
 Reverse case? Apply filter H  to impulse function
 Using fact that convolution is commutative

 Result is the filter H
 Noise
 While taking picture (during capture), noise may occur
 Noise? Errors, degradations in pixel values
 Examples of causes:
 Focus blurring 
 Blurring due to camera motion
 Additive model for noise:

 Removing noise called Image Restoration
 Image restoration can be done in:
 Spatial domain, or
 Frequency domain
 Types of Noise
 Type of noise determines best types of filters for removing it!!
 Salt and pepper noise: Randomly scattered black + white pixels
 Also called impulse noise, shot noise or binary noise
 Caused by sudden sharp disturbance

Courtesy
Allasdair McAndrews
 Types of Noise
 Gaussian Noise: idealized form of white noise added to 
image, normally distributed
 Speckle Noise: pixel values multiplied by random noise

Courtesy
Allasdair McAndrews
Types of Noise
 Periodic Noise: caused by 
disturbances of a periodic 
nature

 Salt and pepper, gaussian 
and speckle noise can be 
cleaned using spatial filters

 Periodic noise can be 
cleaned using frequency 
domain filtering (later) Courtesy
Allasdair McAndrews
 Non‐Linear Filters
 Linear filters blurs all image structures points, edges and 
lines, reduction of image quality (bad!)
 Linear filters thus not used a lot for removing noise

Apply Blurred
Linear Edge
Filter Results
Sharp
edge

Blurred
Sharp Thin
Thin Line
Line Results
Using Linear Filter to Remove Noise?
 Example: Using linear filter to clean salt and pepper noise just 
causes smearing (not clean removal)
 Try non‐linear filters? Courtesy
Allasdair McAndrews
 Non‐Linear Filters
 Pixels in filter range combined by some non‐linear function
 Simplest examples of nonlinear filters: Min and Max filters

Before
filtering

After
filtering

Effect of Narrow
Step Edge Linear Ramp
Minimum Pulse
(shifted to right) (shifted to right)
filter (removed)
Non‐Linear Filters

Original Image with Minimum filter removes Maximum filter (opposite effect):
Salt-and-pepper noise bright spots (maxima) and Removes dark spots (minima) and
widens dark image structures widens bright image structures
Median Filter
 Much better at removing noise and keeping the 
structures

Sort pixel values

within filter region Replace filter “hot spot” pixel
with median of sorted values
Illustration: Effects of Median Filter

Isolated pixels Thin lines

are eliminated are eliminated

A step edge is A corner is

unchanged rounded off
Effects of Median Filter

Original Image with Linear filter removes some of Median filter salt-and-pepper noise
Salt-and-pepper noise the noise, but not completely. and keeps image structures largely
Smears noise intact. But also creates small spots
of flat intensity, that affect sharpness
Median Filter ImageJ Plugin

Get Image width + height,

and Make copy of image

Array to store pixels to be filtered. Good

data structure in which to find median

Copy pixels within filter

region into array

Sort pixels within filter using

java utility Arrays.sort( )

Middle (k) element of sorted array

assumed to be middle. Return as median
Weighted Median Filter
 Color assigned by median filter determined by colors of 
“the majority” of pixels within the filter region
 Considered robust since single high or low value cannot 
influence result (unlike linear average)
 Median filter assigns weights (number of “votes”) to 
filter positions

 To compute result, each pixel value within filter region is 
inserted W(i,j) times to create extended pixel vector
 Extended pixel vector then sorted and median returned
Weighted Median Filter
Sort extended pixel
Insert each pixel within filter
Pixels within vector and return median
region W(I,j) times into
filter region extended pixel vector

Weight
matrix

Note: assigning weight to center pixel larger than sum of all other pixel
weights inhibits any filter effect (center pixel always carries majority)!!
Weighted Median Filter
 More formally, extended pixel vector defined as

 For example, following weight matrix yields extended 
pixel vector of length 15 (sum of weights)

 Weighting can be applied to non‐rectangular filters
 Example: cross‐shaped median filter may have weights
An Outlier Method of Filtering
 Algorithm by Pratt, Ref: Alasdair McAndrew, Page 116
 Median filter does sorting per pixel (computationally expensive)
 Alternate method for removing salt‐and‐pepper noise
 Define noisy pixels as outliers (different from neighboring pixels by an 
amount > D)
 Algorithm:
 Choose threshold value D
 For given pixel, compare its value p  to mean m of 8 neighboring pixels
 If |p – m| > D, classifiy pixel as noise, otherwise not
 If pixel is noise, replace its value with m; Otherwise leave its value 
unchanged
 Method not automatic. Generate multiple images with 
different values of D, choose the best looking one
Outlier Method Example
 Effects of choosing different values of D

Courtesy
Allasdair McAndrews

D value too large: removes

D value too small: removes
noise from light regions
noise from dark regions

 D value of 0.3 performs best 
 Overall outlier method not as good as median filter
Other Non‐Linear Filters

 Any filter operation that is not linear (summation), is 
considered linear
 Min, max and median are simple examples
 More examples later:
 Morphological filters (Chapter 10)
 Corner detection filters (Chapter 8)
 Also, filtering shall be discussed in frequency domain
 Extending Image Along Borders

Extend: pixels outside

Pad: Set pixels outside
border take on value of
border to a constant
closest border pixel

Mirror: pixels around Wrap: repeat pixels

image border periodically along
coordinate axes
Filter Operations in ImageJ
 Linear filters implemented by ImageJ plugin class 
ij.plugin.filter.Convolver
 Has several methods in addition to run( )

Define filter matrix

Create new instance of

Convolver class

Apply filter (Modifies Image I destructively)

 Gaussian Filters
 ij.plugin.filter.GaussianBlur implements 
gaussian filter with radius (σ)
 Uses separable 1d gaussians

Create new instance of

GaussianBlur class

Blur image ip with

gaussian filter of
radius r
Non‐Linear Filters
 A few non‐linear filters (minimum, maximum and 
median filters implemented in  
ij.plugin.filter.RankFilters
 Filter region is approximately circular with variable 
radius
 Example usage:
Recall: Linear Filters: Convolution
Convolution as a Dot Product
 Applying a filter at a given pixel is done by taking  
dot‐product between the image and some vector
 Convolving an image with a filter equal to:
 Filter         each image window (moves through image)

Dot product
Digital Image Processing (CS/ECE 545) 
Lecture 4: Filters (Part 2) 
& Edges and Contours

Prof Emmanuel Agu
Computer Science Dept.
Worcester Polytechnic Institute (WPI)
What is an Edge?
 Edge? sharp change in brightness (discontinuities)
 Where do edges occur?
 Actual edges: Boundaries between objects 
 Sharp change in brightness can also occur within object
 Reflectance changes
 Change in surface orientation
 Illumination changes. E.g. Cast shadow boundary
Edge Detection
 Image processing task that finds edges and contours in 
images
 Edges so important that human vision can reconstruct 
edge lines
Characteristics of an Edge
 Edge: A sharp change in brightness
 Ideal edge is a step function in some direction
Characteristics of an Edge
 Real (non‐ideal) edge is a slightly blurred step function
 Edges can be characterized by high value first derivative

Rising slope causes positive

+ high value first derivative Falling slope causes negative
+ high value first derivative
 Characteristics of an Edge
 Ideal edge is a step function in certain direction.
 First derivative of I(x) has a peak at the edge
 Second derivative of I(x) has a zero crossing at edge

Real edge

Ideal edge

First derivative
shows peak

Second derivative
shows zero
crossing
Slopes of Discrete Functions
 Left and right slope may not be same
 Solution? Take average of left and right slope
Computing Derivative of Discrete Function

Actual slope
(solid line) Estimated slope
(dashed line)
Finite Differences
 Forward difference (right slope)

 Backward difference (left slope)

 Central Difference (average slope)
Definition: Function Gradient
 Let f(x,y) be a 2D function
 Gradient: Vector whose direction is in direction of maximum 
rate of change of f  and whose magnitude is maximum rate of 
change of f
 Gradient is perpendicular to edge contour
Image Gradient
 Image is 2D discrete function
 Image derivatives in horizontal and vertical directions

 Image gradient at location (u,v)

 Gradient magnitude

 Magnitude is invariant under image
rotation, used in edge detection
Derivative Filters
 Recall that we can compute derivative of discrete function as

 Can we make linear filter that computes central differences
Finite Differences as Convolutions
 Forward difference

 Take a convolution kernel
Finite Differences as Convolutions
 Central difference

 Convolution kernel is: 

 Notice: Derivative kernels sum to zero
x‐Derivative of Image using Central Difference
y‐Derivative of Image using Central Difference
Derivative Filters Gradient slope in
horizontal direction

A synthetic
image

Magnitude of
gradient

Gradient slope in
vertical direction
Edge Operators
 Approximating local gradients in image is basis of many 
classical edge‐detection operators
 Main differences?
 Type of filter used to estimate gradient components
 How gradient components are combined
 We are typically interested in
 Local edge direction
 Local edge magnitude
Partial Image Derivatives
 Partial derivatives of images replaced by finite differences

 Alternatives are:
Prewitt

 Robert’s gradient
Sobel
Using Averaging with Derivatives
 Finite difference operator is sensitive to noise
 Derivates more robust if derivative computations are 
averaged in a neighborhood
 Prewitt operator: derivative in x, then average in y

Derivative in x direction

Average in y direction Note: Filter kernel is

flipped in convolution

 y‐derivative kernel,        defined similarly
Sobel Operator
 Similar to Prewitt, but averaging kernel is higher in middle

Average in x direction
Note: Filter kernel is
Derivative in y direction flipped in convolution
Prewitt and Sobel Edge Operators
 Prewitt Operator

Written in
separable form

 Sobel Operator
Improved Sobel Filter
 Original Sobel filter relatively inaccurate
 Improved versions proposed by Jahne
Prewitt and Sobel Edge Operators
Scaling Edge Components
 Estimates of local gradient components obtained from 
filter results by appropriate scaling

Scaling factor for

Prewitt operator

Scaling factor for

Sobel operator
Gradient‐Based Edge Detection
 Compute image derivatives by convolution

Scaled Filter results

 Compute edge gradient magnitude

 Compute edge gradient direction

Typical process of
Gradient based
edge detection
Gradient‐Based Edge Detection
 After computing gradient magnitude and orientation 
then what?
 Mark points where gradient magnitude is large wrt 
neighbors
Non‐Maxima Suppression
 Retain a point as an edge point if:
 Its gradient magnitude is higher than a threshold
 Its gradient magnitude is a local maxima in gradient direction

Simple thresholding will

compute thick edges
Non‐Maxima Suppression
 A maxima occurs at q, if its magnitude is larger than 
those at p and r
Roberts Edge Operators
 Estimates directional gradient along 2 image diagonals
 Edge strength E(u,v): length of vector obtained by adding 2 
orthogonal gradient components D1(u,v) and D2(u,v)

 Filters for edge components
Roberts Edge Operators
 Diagonal gradient components produced by 2 Robert filters
Compass Operators
 Linear edge filters involve trade‐off

Sensitivity to Sensitivity to
Edge magnitude orientation

 Example: Prewitt and Sobel operators detect edge magnitudes 
but use only 2 directions (insensitive to orientation)
 Solution? Use many filters, each sensitive to narrow range of 
orientations (compass operators)
Compass Operators
 Edge operators proposed by Kirsh uses 8 filters with 
orientations spaced at 45 degrees

Need only to compute 4 filters

Since H4 = - H0, etc
Compass Operators
 Edge strength EK at position(u,v) is max of the 8 filters

 Strongest‐responding filter also determines edge orientation 
at a position(u,v) 
Edge operators in ImageJ
 ImageJ implements Sobel operator
 Can be invoked via menu Process ‐> Find Edges
 Also available through method void findEdges( )
for objects of type ImageProcessor
 References
 Wilhelm Burger and Mark J. Burge, Digital Image 
Processing, Springer, 2008
 University of Utah, CS 4640: Image Processing Basics, 
Spring 2012
 Rutgers University, CS 334, Introduction to Imaging 
and Multimedia,  Fall 2012