Image Processing
Fundamentals
Dr. Romik Chatterjee
Vice President of Engineering
Graftek Imaging, Inc.
Agenda
• Objectives and Motivations
• Enhancing Images
• Checking for Presence
• Locating Parts
• Measuring Features
• Identifying and Verifying Components
Class Goals
• Teach the fundamental image processing tools available
in machine vision software
• Provide some background into how the algorithms work
• Accelerate your machine vision learning curve
• What not to expect
– Learn how to develop a complete vision system
– Learn specific, proprietary functions
– 3D vision, color or other advanced topics
– Discussion of the benefits of different application development
environments
Image Processing for Machine Vision
• Objective
– To extract useful information present in an image, in a limited
time
• Secondary
– To display an image for users
• Not
– Improve appearance of image in general
• Used for
– Image pre‐processing
• Minimize variations of information in the image
• Prepare the image for processing and measurement
– Application specific processing
• Use image to count, locate, and measure attributes
Image Types
Grayscale Image
• Grayscale
– 8 bit: pixel values range from 0 to 255
– 12 bit: pixel values range from 0 to 4095
– 16 bit: pixel values range from 0 to 65535
• Color
– Composed of 3 grayscale images (RGB)
• Other types
– Binary: pixel values: 0 and 1
• Commonly used to identify objects of interest
in an image
• Usually the result of image processing step
– Floating point: pixel values range from –
X.xx to X.xx
• Usually a result of a computation
Binary Image
What is an Ideal Image?
• Range of grayscale values
– Spread out between 0 and 255
– No pixels “saturated” at 255 (for most applications)
• Impossible to distinguish between saturated pixels
• Good contrast
– Between the “right” parts of the image
• Repeatable
In short, an ideal image requires the least number of image
processing steps to obtain the result.
Class Organization
Enhance Check Locate Measure Identify
• Filter • Create • Match • Detect • Read text
noise or Particles patterns edges (OCR)
unwanted • Measure • Match • Measure • Read 1D
features intensity geometry distance barcodes
• Remove • Analyze • Set‐up • Calculate • Read 2D
distortion particles coordinate geometry codes
• Calibrate systems
images
Class Organization
Enhance Check Locate Measure Identify
• Filter • Create • Match • Detect • Read text
noise or Particles patterns edges (OCR)
unwanted • Measure • Match • Measure • Read 1D
features intensity geometry distance barcodes
• Remove • Analyze • Set‐up • Calculate • Read 2D
distortion particles coordinate geometry codes
• Calibrate systems
images
Motivation
Read characters on a textured surface
Motivation
Not possible to cleanly segment characters
Motivation
Results
without
preprocessing
Motivation
Read characters on a textured surface
Motivation
Image after periodic pattern is removed
Motivation
Motivation
Results
without
preprocessing
Results
with
pre‐
processing
Objective of Image Preprocessing
• Process an image so that the resulting image is more
suitable than the original for a specific application
• A preprocessing method that works well for one
application may not be the best method for another
application
Image Preprocessing
• Pre‐processing occurs before the application specific
processing
Application
Acquisition Preprocessing Specific
Processing
• Shading • Intensity
Correction Measurements
• De‐blurring • OCR
• De‐noising • Pattern
• Contrast Matching
Enhancement • Gauging
• Feature • Barcode
Enhancement • Particle Analysis
Enhancement Techniques
• Spatial Domain: pixel‐wise operations
– Brightness, contrast and gamma
– Lookup tables
– Gray morphology
– Filtering (smoothing, median, general convolution)
• Frequency Domain
– Deblurring
– Filtering
Lookup Tables
• A lookup table is a function that maps grayscale
values in an image to new grayscale values, to create
a new result image
255 255 255
200
Result
Result
Result
0
0 255 0 255 0 255
Source Source Source
Source Result Source Result Source Result
0 255 0 200 0 240
1 254 1 200 1 245
2 253 2 200 2 246
.. .. .. .. .. ..
255 0 255 200 255 171
Brightness
• Adds a constant
grayscale to all of the
image
• Can improve
appearance but not
useful for most image
processing steps Result
Result
Source Source
Contrast
• Used to increase or decrease
contrast
• Normal = 45 (degree line)
– High = Higher than 45
– Low = Lower than 45
• Typical use is to improve edge
detection
• Sacrifices one range of values to
improve another Result
Result
Source Source
Gamma
• Nonlinear contrast
enhancement
• Higher gamma improves
contrast for larger grayscale
values
• Does not cause saturation
Result
Result
Source Source
Histogram Equalization
• Alters grayscale values of pixels so that
they become evenly distributed across
the full grayscale range
• The function associates an equal
number of pixels per constant grayscale
interval
• Takes full advantage of the available
shades of gray
• Enhances contrast of the image without
modifying the structure Equalized
Histogram Equalization
Cumulative Histogram
Bright
Improved contrast
Dark between characters
and background
Low
Contrast
High
Contrast
Other Lookup Tables
Original
Reverse Square Power 1/x
Gray Morphology
Morphology Function
Erosion Min(Neighbors) Erosion
Dilation Max(Neighbors) Dilation
Open Dilation(Erosion(I)) Open
Close Erosion(Dilation(I)) Close
Spatial Filtering
Filter Type Filters Gaussian
• Smoothing
• Gaussian
Lowpass
• Median
• Nth Order Median
• Gradient
• Laplacian
Highpass • Roberts
• Sobel
Gradient
• Prewitt
Sobel
Frequency Domain Filtering
• Standard filtering can be done in frequency domain
– Low Pass, High Pass, Band Pass, Band Stop, etc.
• Compute the Fourier transform of the image
• Multiply with the transfer function of the filter
• Take the inverse Fourier transform to get the filtered image
Input Output
FFT H(u,v) IFFT
Image Image
I(x,y) F(u,v) F(u,v).H(u,v) R(x,y)
Periodic noise Bandstop filtered
Low Pass Filter Examples
• Low Pass with Frequency Domain Filter
• Low Pass with Gaussian Filter
High Pass Filtering Examples
• Detect edges
• Sharpen image
ENHANCE IMAGES:
IMAGE CALIBRATION
Spatial Calibration
• Corrects for lens and perspective distortion
• Allows the user to take real‐world measurements from
image based on pixel locations.
lens distortion
perspective
known orientation offset
Calibrating Your Image Setup
• Acquire image of a calibration grid with known real‐
world distances between the dots
• Learn the calibration (mapping information) from its
perspective and distortion
• Apply this mapping information to subsequent
images and measurements
Image Correction
• Use calibration to adjust image geometry so features
are represented properly.
Types of Calibration
• 2D spatial calibration
– Applied only to a plane
– Corrects for lens and perspective distortion
– Does not improve resolution of a measurement
– Cannot compensate for poor lighting or unstable conditions
• 3D spatial calibration: x, y, z
• Flat field calibration
– Corrects for lighting variations
• Color
Class Organization
Enhance Check Locate Measure Identify
• Filter noise • Create • Match • Detect • Read text
or Particles patterns edges (OCR)
unwanted • Measure • Match • Measure • Read 1D
features intensity geometry distance barcodes
• Remove • Analyze • Set-up • Calculate • Read 2D
distortion particles coordinate geometry codes
• Calibrate systems
images
Histograms and Thresholding
• Histogram
– Provides a quantitative distribution of pixels in an image
– Indicates the number of pixels at each gray level
• Thresholding
– Converts each pixel value in an image to 0 or 1 according to
the value of the original pixel
– Helps extract significant structures in an image
Histogram and Thresholding
Original Image
Histogram of Image
Binary Image
Finding Gray Objects
• Can also set upper and lower limits for pixel values
• Pixels inside the bounds of the limit (blue region) are
set to 1, and those outside the limit are set to 0
Automatic Thresholding
Manual Auto
Threshold Threshold
Local Thresholding
Global Local
Threshold Threshold
B C
Particles and Connectivity
• Thresholding creates binary image:
– Pixels are either ‘0’ or ‘1’
• A particle in a binary image is a group of
connected ‘1’ pixels
• Defining connectivity
– Connectivity‐4: two pixels are considered part
of the same particle if they are horizontally or
vertically adjacent
– Connectivity‐8: two pixels are considered part
of the same particle if they are horizontally,
vertically, or diagonally adjacent.
How Connectivity Affects a Particle
– How many particles with
connectivity‐4?
How Connectivity Affects a Particle
– How many particles with
1
connectivity‐4?
1 1 1
3 1 2
2 2 2
2 2 2 2
3 2 2 2
2
How Connectivity Affects a Particle
– How many particles with
connectivity‐4?
– How many particles with
connectivity‐8?
How Connectivity Affects a Particle
– How many particles with
1
connectivity‐4?
1 1 1
1 1
1 1 1
– How many particles with
connectivity‐8? 1 1 1 1
2
2 1 1 1
1
Binary Morphology
• Binary morphological functions extract and alter the
structure of particles in a binary image
• Morphological functions remove unwanted information
caused by the thresholding process:
– Noise particles
– Particles touching the border of an image
– Particles touching each other
– Particles with uneven borders
Erosion
• Decreases the size of objects in an image
– Removes a layer of pixels along the boundary of the particle
– Eliminates small isolated particles in the background and removes
narrow peninsulas
• Use Erode to:
– Separate particles for counting
Erosion
Dilation
• Increases the size of objects in an image
– Adds a layer of pixels around the boundary of an object (including the inside
boundary for objects with holes)
– Eliminates tiny holes in objects
– Removes gaps or bays of insufficient width
• Use Dilate to:
– Make particles touch
Dilation
Erosion vs. Dilation
Erosion
Dilation
Open
• An erosion followed by a dilation
– Remove small particles and smooth boundaries
– Does not significantly alter the size or shape of particles
– Borders removed by the erosion process are replaced by the dilation process
• Use Open To:
– Eliminate small particles that constitute noise
Open
Close
• A dilation followed by an erosion
– Fills holes and creates smooth boundaries
– Does not alter the size or shape of particles
– Particles that do not connect after the dilation are not changed
• Use Close To:
– Eliminate small holes that constitute noise
Close
Advanced Morphology
• Advanced morphological functions are combinations
of operations, each of which is responsible for a
single operation
• These functions execute the following tasks:
– Separate particles
– Remove small or large particles
– Keep or remove particles identified by morphological
parameters
– Fill holes
– Remove particles from an image border
– Segmenting the image
Particle Filtering
• Keeps or removes particles based on geometric features
– Area, length, aspect ratio and other features are
commonly used to filter
• Typically used on binary images
• Cleans up noisy images
Threshold Particle Filter
Application: Analyze Particles
Original
Threshold
Particle Filter
Class Organization
Enhance Check Locate Measure Identify
• Filter • Create • Match • Detect • Read text
noise or Particles patterns edges (OCR)
unwanted • Measure • Match • Measure • Read 1D
features intensity geometry distance barcodes
• Remove • Analyze • Set‐up • Calculate • Read 2D
distortion particles coordinate geometry codes
• Calibrate systems
images
LOCATING PARTS:
PATTERN MATCHING
Introduction to Matching
• Locates regions of a grayscale image that match a
predefined template
– Calculate a score for each matching region
– Score indicates quality of match
• Returns the XY coordinates, rotation angle and scale
for each match
Applications
• Presence Detection
• Counting
• Alignment
• Inspection
How It Works
• Two step process:
– Step 1: Learn Template
• Extract information useful for uniquely characterizing the pattern
• Organize information to facilitate faster search of the pattern in
the image
– Step 2: Match
• Use the information present in the template to locate regions in
the target image
• Emphasis is on search methods that quickly locate matched
regions
Pattern Matching Methods
• Different ways to perform pattern matching based
on the information extracted from the template
• Two common methods:
– Correlation Pattern Matching
• Relies on the grayscale information in the image for
matching
– Geometric Pattern Matching
• Relies on edges and geometric features in the image
for matching
Correlation Pattern Matching
• Grayscale information present
in the image Template (Pattern)
• Directly uses the underlying
grayscale distribution in the
image for matching
• Grayscale values in the
pattern are matched to
regions in the image using
normalized cross‐correlation
• Score ranges from 0‐1000
• Used to allow imperfect match
Correlation Pattern Matching
• When to use:
– Template primarily characterized by grayscale information
– Matching under uniform light changes
– Little occlusion and scale changes in image
– Good for the general case
Good template Bad template
Correlation Pattern Matching
• When NOT to use:
– Non‐uniform lighting
– Occlusion more than 10%
– Scale changes
Geometric Pattern Matching
• Matching tool you can use to locate parts that
contain distinct edge information
• Not useful when template is predominantly defined
by texture
GPM is Tolerant to…
Scale
Changes
Occlusion
Non‐uniform Background
Lighting Changes
GPM – Feature‐based
Target Image
Template
Image
Extract Curves
Extract Features
circles
Match Features
parallel lines
Feature Comparison
Feature CPM GPM
Template contains texture‐like information Yes
Template contains geometric information Yes Yes
Find multiple match locations Yes Yes
Rotation Yes Yes
Scale Yes
Occlusion Yes
Matching under non‐uniform lighting Yes
Sub‐pixel match locations Yes Yes
LOCATING PARTS:
COORDINATE SYSTEMS
Region of Interest
• Region of Interest
– A portion of the image upon which an image processing step may be
performed
– Can be defined statically (fixed)
– Or dynamically: based on features located in the image
• Used to
– Process only pixels that are interesting
Coordinate Systems
• Defined by a reference point (origin) and angle within the image, or by the
lines that make up its axes
• Allows you to define search areas that can move around the image with
the object you are inspecting
• Usually based on a characteristic feature of the object under inspection
– Use pattern match to locate features
– Use features to establish coordinate system
Coordinate Systems – Set Up
– Define an origin
• Locate an easy‐to‐find
feature in your reference
image
• Set a coordinate system
based on its location and
orientation
– Set up measurement ROIs in
reference to the new origin
– Acquire a new image
– Locate reference point
– Adjust measurement ROIs
Class Organization
Enhance Check Locate Measure Identify
• Filter • Create • Match • Detect • Read text
noise or Particles patterns edges (OCR)
unwanted • Measure • Match • Measure • Read 1D
features intensity geometry distance barcodes
• Remove • Analyze • Set‐up • Calculate • Read 2D
distortion particles coordinate geometry codes
• Calibrate systems
images
Edge Detection Overview
• Process of detecting transitions in an image
• One of the most commonly used machine vision tools
• Attractive because:
– Simple to understand and use
– Localized processing – fast
– Applicable to many applications
Different Illuminations
Different Edges
1D Edge Detection
• Detect edge points along a line
• Basic operation:
– Get pixel values along the
line
– Compute gradient
Pixel Value
information
– Peaks and valleys represent
edge locations Position Along Line (pixels)
– Get first, first & last, best, or
Gradient Value
all edges
Position Along Line (pixels)
Subpixel Accuracy
• Subpixel location of edge can be computed using
parabolic interpolation
Parabolic Fit
Gradient Value
Position Along Line (pixels) Subpixel Location
Edge Detector Tools
• High level tools based on the edge
detectors
• Rake:
– Used to find multiple edges and fit a
shape through them
• Configurable search directions, sub‐
sampling ratios, and display settings
Straight Edge (Line) Detection
• Detect straight lines in an image
– Extension of 1D edge detection
• Straight edge detection options:
– Rake‐based
– Projection‐based
Edge Detection Applications
• Detect
Features
• Alignment
• Gauging
• Inspection
Application: Locating Parts
Failed due
to intensity
Application: Dimension Verification
• Dimensional measurements, such as lengths, distance, and diameter
– Inline gauging inspections are used to verify assembly and
packaging routines
– Offline gauging is used to judge product quality according to a
sample
Class Organization
Enhance Check Locate Measure Identify
• Filter • Create • Match • Detect • Read text
noise or Particles patterns edges (OCR)
unwanted • Measure • Match • Measure • Read 1D
features intensity geometry distance barcodes
• Remove • Analyze • Set‐up • Calculate • Read 2D
distortion particles coordinate geometry codes
• Calibrate systems
images
Identify
• 1D and 2D Codes
• Marking methods
• Reading
• Examples
1D Codes
• Applications using 1D bar codes have been around for
over 35 years
• Barcode data is an index into a large central data
storage
• Code is easily read by laser scanners
• Low data capacity in large footprint
Code 3 of 9 Code 128 EAN 13
2D Codes
• Usually not an index into a database
• Camera‐based vision systems are preferred reading
method
• High data capacity in small footprint
Data Matrix PDF 417 QR Code
1D vs 2D
1D Codes 2D Codes
Low data capacity High data capacity
Index into large database Self contained data
Large footprint Small footprint
Redundancy in Y dimension Error correction capability
Readable by laser scanner Requires camera based reader
Requires as much as 80% contrast Can be read in low contrast
Optical Character Recognition
Particle‐based OCR
TRAIN READ
Typical steps:
• Region of
interest around
line of text
• Threshold
• Character
segmentation
• Compare to
library
• Character is
learned or
recognized
Class Organization
Enhance Check Locate Measure Identify
• Filter • Create • Match • Detect • Read text
noise or Particles patterns edges (OCR)
unwanted • Measure • Match • Measure • Read 1D
features intensity geometry distance barcodes
• Remove • Analyze • Set‐up • Calculate • Read 2D
distortion particles coordinate geometry codes
• Calibrate systems
images
Contact Information
Dr. Romik Chatterjee
Vice President of Engineering
Graftek Imaging Inc.
8900 Shoal Creek Boulevard
Building 300, Suite B
Austin, Texas 78757
USA
Phone: +1 512-416-1099
Email: romik@graftek.com
www.graftek.com
