Feature Extraction for Image Analysis

Feature Extraction for image analysis

Feature Extraction for image analysisDr.S.DomnicAssistant ProfessorDepartment of Computer ApplicationsNational Institute of TechnologyTiruchirappalli.

Overview of presentationBasics of ImageImage FeaturesFeature Extraction TechniquesApplicationsBasics of imageAn analog image is 2D F(x,y) which has infinite precision in spatial parameters x and y, infinite precision in intensity at each spatial point (x,y).

An digital image is 2D array of intensity values with finite precision.Basics of imageIntensity value Color value Color value 0 represents black color 255 represents white colorFor ex: gray scale image

Basics of imageImage Types

RGB Gray Scale Black & White

Range 0 to 224-1 0 to 28-1 0-1

Basics of image

Basics of image

Decomposing an image in to its bitsBasics of imageEffects of differing image resolution

Basics of image

Effects of gray level resolution256 levels16 graylevels2 levels

Image in different walks of life

Remote sensingMedical fieldForensic fieldImage featuresImage analysis is the extraction of meaningful information from digital image by means of digital image processing techniquesfeatureImage OperatorsHistogram Thresholding Template ConvolutionAveraging Statistical operators (median filter, mode filter..)

Image OperatorsHistogram

Image Operators Thresholding

Image OperatorsTemplate Convolution Weighting Coefficients

Image OperatorsAveraging

Types of featuresLow level featureEdge, curvature, texture, orientation, etc..High level featureShape, patch, region etc

EdgeA change in intensity values, it can be revealed by differencing adjacent points.

First Order -Horizontal edge detector vertical edge

First Order -Vertical edge detector horizontal edge

Edge

To detect horizontal and vertical edge together

Template for first order difference

Edge20CurvatureCurvature is the rate of change in edge direction.Points where the edge direction changes rapidly are corners.

ShapeShape is a high level feature.Facial features: eyes, ears, mouth Complex picture can be decomposed into simple shapesInvariance properties: (i) Invariant to illumination (bright or dark) (ii) Invariant to position, location (iii) Invariant to rotation or orientation

Shape extraction by subtraction and thresholding

Shape extraction by Template matching

Shape extraction by Template matching

Shape extraction by Hough transformIt is a technique to locate shape in images.It has been used to extract lines, circles and ellipses.

Object descriptorsObjects are represented as a collection of pixels in an image.Describe the properties of the group of pixels. Object characterization can be done by (i) Region/ and shape descriptorBasic shape measures are : area, perimeter, compactness, moments (Zernike moment)

Object descriptors (description of perimeter or boundary)Based on frequency analysis, Fourier coefficients are used to describe shape. (i) define curve (ii) Four.coef

Object descriptors (Region descriptor)Region or area descriptor area, density of the region

TextureImage Texture gives us information about the spatial arrangement of color or intensities in an image or selected region of an image.SegmentationGabor filterFourier transform

Texture image example

Fourier TransformMathematical transformation employed to transform between spatial domain (time) and frequency domain.The Fourier transform is also a reversible operation.Shape analysis (object description)Texture analysisCompression

Fourier seriesThe idea of a Fourier series is that any (reasonable) function, f(x), that is periodic on the interval (ie: f(x + n) = f(x) for all n) can be decomposed into contributions from sin(nx) and cos(nx).

The Fourier series can be written as:f(x) = a0 /2+ a1 cos (x) + a2 cos (2x) + a3 cos (3x) + . . . + an cos (nx)+ b1 sin (x) + b2 sin (2x) + b3 sin (3x) + . . . + bn sin (nx)

Where:cos (nx) and sin (nx) are periodic on the interval 2*PI for any integer n.The an and bn coeffcients measure the strength of contribution from each

Periodic signal

sin( x+ P )=sin( x )Fourier series approximationFirst four partial sum of the Fourier series for a square wave:

f(x) = a0 /2+ a1 cos (x) + b1 sin (x).

f(x) = a0 /2+ a1 cos (x) + a2 cos (2x) + b1 sin (x) + b2 sin (2x).

f(x) = a0 /2+ a1 cos (x) + a2 cos (2x) + a3 cos (3x) + b1 sin (x) + b2 sin (2x) + b3 sin (3x).

Fourier series approximation (periodic signal for T / 2pi period)

Complex form:

Fourier TransformFor any non-periodic function and assume T-> , rewrite previous general Fourier series equation and get:

When T->

Fourier Transform propertiesShift Invariance:If we shift all the features by a fixed amount, the magnitude of its Fourier transform does not change:

Fourier Transform propertiesRotation:The Fourier Transform of an image rotates when the source image rotates.

Fourier Transform propertiesFrequency Scaling:Scaling the original image, the spectrum will spread from the origin consistent with an increase in spatial frequency

ApplicationsAgricultureMedical Image analysisSurveillanceetc..

Thank you