扫码添加小助手 扫码关注B站 获取更多内容 了解更多课程
COMP 9517 2021 T1
Lecture Schedule
• Introduction & Image Formation • Image Processing
• Feature Representation
• Image Segmentation
• Pattern Recognition
• Motion and Tracking
• Deep Learning
• Recognition Applications
R/B/T/E/D/U
Student Assessment (2020 T3)
Assignments Release: Week 1 Due: Week 3 Marks: 10%
Group Project Release: Week 5 Due: Week 10 Marks: 40%
Lab Work Release: Week 2,3,5,7,8
9517 Due: Week 3,4,6,8,9 Marks: 10%
Exam Marks: 40%
Student Assessment (2021 T1)
Assignments Release: Week 2 Due: Week 4 Marks: 10%
Group Project Release: Week 5 Due: Week 10 Marks: 40% (15% – Personal; 25% – Group)
Lab Work Release: Week 2, 3, 4, 5
9517 Due: Week 3, 4, 5, 6 Marks: 10%
Exam Marks: 40%
What is Computer Vision
Computer vision is the interdisciplinary scientific field that develops theories and methods allowing computers to extract high-level information from digital images or videos. Physics-based mathematical and statistical models as well as machine-learning methods are used to assist in the task.
This course provides an introduction to fundamental concepts and an opportunity to develop a real-world application of computer vision.
Introduction
Current state on computer vision:
• Image classification
• Detection (e.g., face detection, object detection…) • Motion and tracking
• Smart cars
• Video surveillance
Critical Issues
• Sensing: how do sensors obtain images of the world?
• Encoded Information: how do images yield information of the scene,
such as colour, texture shape, motion, etc.?
• Representation: what representations are appropriate to describe objects?
• Algorithms: what algorithms process image information and construct scene descriptions?
High & Low Level Vision
• LowLevelVision
• Sensing:imagecaptureanddigitization
• Pre-processing:improveimagequality:suppressnoise,enhanceobject
features, edge extraction
• Imagesegmentation:separateobjectsfrombackground,partitionimageinto
objects of interest
• Description:computefeatureswhichdifferentiateobjectsalsocalledfeature
extraction
• Classification:assignlabelstoimagesegments(regions)
• HighLevelVision
• Recognition:identificationofobjects
• Interpretation:assignmeaningtogroupsofrecognizedobjects • Sceneanalysis
Image Formation
Mapping between image and world coordinates • Pinholecameramodel
• Projectivegeometry
• Vanishingpointsandlines • Projectionmatrix
Projective Geometry
What is lost? • Length
• Angles
What is preserved?
• Straightlinesarestillstraight
Pinhole Camera
Projection
Radial Distribution
Image Formation
Colour Image
Image in Computer
Colour spaces: RGB
Colour spaces: HSV
Colour spaces:YCbCr
Digitization and Sampling
• Digitization:convertsandanalogimagetodigitalimage • Samplingdigitizesthecoordinatesxandy:
• spatialdiscretizationofpicturepicturefunctionF(x,y)
• useagridofsamplingpoints,normallyrectangular:imagesampledatpoints.
Spatial Resolution
• Spatial Resolution: number of pixels per unit of length
• Resolution decreases by one half-see right
• Human faces can be recognized at 64 x 64 pixels
per face
• Appropriate resolution is essential:
• Too little resolution, poor recognition
• Too much resolution, slow and wastes
memory
Quantisation
• Quantisation digitizes the intensity or amplitude values, i.e., F(x, y)
• Called intensity or gray level quantisation
• Gray-level resolution:
• Usually has 16, 32, 64, …, 128, 256 levels
• Numbers of levels should be high enough for human perception of shading
details – human visual system require about 100 levels for a realistic image.
Image Analysis
• Manipulation of image data to extract the information necessary for solving an imaging problem
• Consists of preprocessing, data reduction and feature analysis
• Preprocessing removes noise, eliminates irrelevant information
• Data reduction extracts features for the analysis process
• During feature analysis, the extracted features are examined and evaluated for
their use in the application
Image Preprocessing
• Input and output are intensity images
• Aim to improve image, by suppressing distortions and enhancing image features, so
that result is more suitable for a specific application
• Exploit redundancy in image: for example, neighbouring pixels have similar brightness
value
• Two types of image processing/image transformation:
• Spatial domain – refers to the 2D image plane
• Transform domain, mainly frequency domain, work on the Fourier transform of
the image
• Two principal categories in spatial domain processing:
• Intensity transformation (on single pixels)
• Spatial filtering (on pixel and its neighbours)
Spatial Domain Techniques
Spatial Domain Techniques
Image Preprocessing on Spatial Domain
• Some basic gray-level transformation function
• Histogram processing
• Spatial filtering
• Smooth spatial filter
• Sharpen spatial filter
Basic gray-level Transformation
• Image reversal
• Log transformation
• Power transformation
• Contrast Stretching
s=L−1−r
s = clog(1 + r)
s = crγ
Image Reversal
s=L−1−r
∈ [1, L − 1]
sand rrepresentsthepixelvaluesbeforeandafterprocessingrespectively. Produces equivalent of a photo negative
Useful for enhancing white or gray detail in dark regions of image, when black areas are dominany.
Log Transformation
s = clog(1 + r) Where c is constant
• Maps narrow range of low gray-level values into wider range of output values, and opposite for higher gray-level values
• To expand the value of dark pixel or suppress higher gray level value in the picture.
Power Transformation
s = crγ
• Similar to log transformation on input and output
• Family of possible transform by varying
• Useful in displaying an image accurately on a computer screen (for example on web sites!)
by pre-processing images
Contrast Stretching
• One of the simplest piecewise linear transformations
• To increase the dynamic range of gray levels in image
• Used in display devices or recoding mediums to span the full intensity range
(r1, s1) = (r2, s2) r1 = r2
s1 =0,s2 =L−1 (r1, s1) = (rmin, 0)
(r2, s2) = (rmax, L − 1)
Gray-level Slicing
• Highlighting of specific range of gray levels
1. Display high value for all gray levels in range of interest, and low value for all others produces binary image
2. Brighten the desired range of gray levels, while preserving background and other gray-scale tones of image
Bit-plane Slicing
• Highlights contribution made to total image appearance by specific bits
• E.g., for an 8-bit image, there are 8 1-bit planes
• Useful in compression
Histogram Processing
• Histogram Equalization
• To get an image with equally distributed
brightness levels over the whole brightness scale
• Histogram Matching
• To get an image with a specified histogram
(brightness distribution)
Histogram Equalization
Histogram Equalization
Histogram Matching
Example
Example
The difference
• Histogram Equalization is kind of generator of
• In histogram matching, the T (rha)s been given.
T(r)
Image Averaging
Spatial Filtering
Spatial Filtering
Smooth spatial filter
• Neighbourhood Averaging
• Gaussian filter
• Median filter (non-linear filter)
• Max filter (non-linear filter)
• Min filter (non-linear filter)
Neighbourhood Averaging
Gaussian Filter
Gaussian Filter
Order-Statistics Filters
Pooling
• Max/average/median pooling
• Provides translation invariance
• Reduces computations
• Popular in deep convolutional neural networks (deep learning)
Sharpening Spatial Filters-Edge Detection
• Goal is to highlight fine detail, or enhance detail that has been blurred
• Spatial differentiation is the tool- strength of response of derivative operator is
proportional to degree of discontinuity of the image at the point where operator is
applied
• Image differentiation enhances edges, and deemphasizes slowly varying gray-level
values.
• Common sharpening spatial filters-edge detection
• Differential detection
• Gradient Operator
• The Laplacian
• The Sobel
• Non-sharpening mask
Derivative definitions
Summary of Derivative
• First-order derivatives produce thicker edges, have stronger response to gray-level step
• Second-order derivatives produce stronger response to fine detail (thin lines, isolated points), produce double response at step changes in gray level
Gradient Operator
The Sobel
The Laplacian
•
Non-sharpening mask
The procedure:
• Blurring the original image
• Obtaining the mask via minus the original image with the blurred image
• Plus the mask on the original
̃ gmask(x, y) = f(x, y) − f(x, y)
g(x, y) = f(x, y) + k × gmask(x, y)