Fourier Transform. Additive noise. Fourier Tansform. I = S + N. Noise doesn t depend on signal. We ll consider:

Similar documents
Announcements. Filtering. Image Filtering. Linear Filters. Example: Smoothing by Averaging. Homework 2 is due Apr 26, 11:59 PM Reading:

Tutorial 2. COMP4134 Biometrics Authentication. February 9, Jun Xu, Teaching Asistant

arxiv:cs.cv/ Jun 2000

The Fourier Transform

NON-LINEAR CONVOLUTION: A NEW APPROACH FOR THE AURALIZATION OF DISTORTING SYSTEMS

ECE 472/572 - Digital Image Processing. Roadmap. Questions. Lecture 6 Geometric and Radiometric Transformation 09/27/11

ADAPTIVE IMAGE FILTERING

DECOUPLING THEORY HW2

Video Data Analysis. Video Data Analysis, B-IT. Lecture plan:

Lecture Notes on Linear Regression

Quantifying Uncertainty

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix

Error Bars in both X and Y

Impulse Noise Removal Technique Based on Fuzzy Logic

IRO0140 Advanced space time-frequency signal processing

1 Derivation of Rate Equations from Single-Cell Conductance (Hodgkin-Huxley-like) Equations

Digital Signal Processing

Feb 14: Spatial analysis of data fields

IMAGE DENOISING USING NEW ADAPTIVE BASED MEDIAN FILTER

COMPARING NOISE REMOVAL IN THE WAVELET AND FOURIER DOMAINS

NUMERICAL DIFFERENTIATION

Appendix B: Resampling Algorithms

SIMG Solution Set #5

Stanford University CS359G: Graph Partitioning and Expanders Handout 4 Luca Trevisan January 13, 2011

MATH 567: Mathematical Techniques in Data Science Lab 8

Section 8.3 Polar Form of Complex Numbers

The Concept of Beamforming

Transform Coding. Transform Coding Principle

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur

Fingerprint Enhancement Based on Discrete Cosine Transform

1 Matrix representations of canonical matrices

Guided Image Filtering

Multi-dimensional Central Limit Theorem

GEMINI GEneric Multimedia INdexIng

Grover s Algorithm + Quantum Zeno Effect + Vaidman

Line Drawing and Clipping Week 1, Lecture 2

Hidden Markov Models & The Multivariate Gaussian (10/26/04)

Introduction to Antennas & Arrays

IMGS-261 Solutions to Homework #9

Tracking with Kalman Filter

Lecture 3: Shannon s Theorem

COMPUTATIONALLY EFFICIENT WAVELET AFFINE INVARIANT FUNCTIONS FOR SHAPE RECOGNITION. Erdem Bala, Dept. of Electrical and Computer Engineering,

Week 11: Chapter 11. The Vector Product. The Vector Product Defined. The Vector Product and Torque. More About the Vector Product

Bezier curves. Michael S. Floater. August 25, These notes provide an introduction to Bezier curves. i=0

Notes on Frequency Estimation in Data Streams

Multi-dimensional Central Limit Argument

Appendix B. The Finite Difference Scheme

Department of Electrical & Electronic Engineeing Imperial College London. E4.20 Digital IC Design. Median Filter Project Specification

Lecture 4: Universal Hash Functions/Streaming Cont d

Georgia Tech PHYS 6124 Mathematical Methods of Physics I

A Novel Fuzzy logic Based Impulse Noise Filtering Technique

A PAPER CLOCK MODEL FOR THE CESIUM CLOCK ENSEMBLE OF TL

CHAPTER-5 INFORMATION MEASURE OF FUZZY MATRIX AND FUZZY BINARY RELATION

Quantum Mechanics I - Session 4

9 Derivation of Rate Equations from Single-Cell Conductance (Hodgkin-Huxley-like) Equations

CS 468 Lecture 16: Isometry Invariance and Spectral Techniques

Image Denoising by Adaptive Kernel Regression

Lecture 12: Discrete Laplacian

Modelli Clamfim Equazione del Calore Lezione ottobre 2014

VQ widely used in coding speech, image, and video

8.4 COMPLEX VECTOR SPACES AND INNER PRODUCTS

Thermodynamics and statistical mechanics in materials modelling II

Linear Feature Engineering 11

From Biot-Savart Law to Divergence of B (1)

Lecture 3: Probability Distributions

Lecture 4: September 12

Feature Selection & Dynamic Tracking F&P Textbook New: Ch 11, Old: Ch 17 Guido Gerig CS 6320, Spring 2013

Composite Hypotheses testing

A linear imaging system with white additive Gaussian noise on the observed data is modeled as follows:

Lecture 10 Support Vector Machines. Oct

Digital Modems. Lecture 2

PHYS 450 Spring semester Lecture 02: Dealing with Experimental Uncertainties. Ron Reifenberger Birck Nanotechnology Center Purdue University

Statistical analysis using matlab. HY 439 Presented by: George Fortetsanakis

princeton univ. F 13 cos 521: Advanced Algorithm Design Lecture 3: Large deviations bounds and applications Lecturer: Sanjeev Arora

Lecture Topics VMSC Prof. Dr.-Ing. habil. Hermann Lödding Prof. Dr.-Ing. Wolfgang Hintze. PD Dr.-Ing. habil.

12 MATH 101A: ALGEBRA I, PART C: MULTILINEAR ALGEBRA. 4. Tensor product

Inner Product. Euclidean Space. Orthonormal Basis. Orthogonal

Change Detection: Current State of the Art and Future Directions

A NEW DISCRETE WAVELET TRANSFORM

White Noise Reduction of Audio Signal using Wavelets Transform with Modified Universal Threshold

Linear Classification, SVMs and Nearest Neighbors

Week 2. This week, we covered operations on sets and cardinality.

What would be a reasonable choice of the quantization step Δ?

The Feynman path integral

Lecture 10 Support Vector Machines II

6 Supplementary Materials

Stochastic Analysis of Image Acquisition, Interpolation and Scale-space Smoothing

SIO 224. m(r) =(ρ(r),k s (r),µ(r))

Convergence of random processes

A Quantum Gauss-Bonnet Theorem

Turbulence classification of load data by the frequency and severity of wind gusts. Oscar Moñux, DEWI GmbH Kevin Bleibler, DEWI GmbH

princeton univ. F 17 cos 521: Advanced Algorithm Design Lecture 7: LP Duality Lecturer: Matt Weinberg

STATS 306B: Unsupervised Learning Spring Lecture 10 April 30

A random variable is a function which associates a real number to each element of the sample space

COS 511: Theoretical Machine Learning. Lecturer: Rob Schapire Lecture #16 Scribe: Yannan Wang April 3, 2014

Lecture 3: Dual problems and Kernels

8 Derivation of Network Rate Equations from Single- Cell Conductance Equations

Pop-Click Noise Detection Using Inter-Frame Correlation for Improved Portable Auditory Sensing

Maximal Margin Classifier

Support Vector Machines. Vibhav Gogate The University of Texas at dallas

Transcription:

Flterng Announcements HW2 wll be posted later today Constructng a mosac by warpng mages. CSE252A Lecture 10a Flterng Exampel: Smoothng by Averagng Kernel: (From Bll Freeman) m=2 I Kernel sze s m+1 by m+1 Convoluton: R= K*I R(, j) = m/2 m/2 h= m/2 k= m/2 R K( h, k) I( h, j k) Propertes of convoluton Let f,g,h be mages and * denote convoluton f * g( x, y) = f ( x u, y v) g( u, v) dudv Commutatve: f*g=g*f Assocatve: f*(g*h)=(f*g)*h Lnear: for scalars a & b and mages f,g,h (af+bg)*h=a(f*h)+b(g*h) Dfferentaton rule f g ( f * g) = * g = f * x x x

Addtve nose I = S + N. Nose doesn t depend on sgnal. We ll consder: I = s + n wth E( n ) = 0 s determnstc. n, n n, n j j ndependent for n dentcally dstrbuted n j Fourer Transform 1-D (sgnal processng) 2-D (mage processng) Consder 1-D Tme doman Frequency Doman Real Complex Consder tme doman sgnal to be expressed as weghted sum of snusod. A snusod cos(ut+φ) s characterzed by ts phase φ and ts frequency u The Fourer sgnal s a functon gvng the weghts (and phase) as a functon of frequency u. Fourer Tansform Dscrete Fourer Transform (DFT) of I[x,y] Fourer bass element 2π e ( ux+vy ) Transform s sum of orthogonal bass functons Inverse DFT Vector (u,v) Magntude gves frequency Drecton gves orentaton. x,y: spatal doman u,v: frequence doman Implemented va the Fast Fourer Transform algorthm (FFT) Here u and v are larger than n the prevous slde. And larger stll...

Usng Fourer Representatons Domnant Orentaton Phase and Magntude e θ = cosθ + sn θ Lmtatons: not useful for local segmentaton Fourer of a real functon s complex dffcult to plot, vsualze nstead, we can thnk phase and magntude Phase s the phase complex Magntude s the magntude complex Curous fact all natural mages have about the same magntude hence, phase seems to matter, but magntude largely doesn t Demonstraton Take two pctures, swap the phase s, compute the nverse - what does the result look lke? Ths s the magntude cheetah pc Ths s the phase cheetah pc

Ths s the magntude zebra pc Ths s the phase zebra pc Reconstructon wth zebra phase, cheetah magntude Reconstructon wth cheetah phase, zebra magntude The Fourer Transform and Convoluton If H and G are mages, and F(.) represents Fourer, then F(H*G) = F(H)F(G) Thus, one way of thnkng about the propertes of a convoluton s by thnkng of how t modfes the frequences mage to whch t s appled. In partcular, f we look at the power spectrum, then we see that convolvng mage H by G attenuates frequences where G has low power, and amplfes those whch have hgh power. Varous Fourer Transform Pars Important facts scale functon down scale up.e. hgh frequency = small detals The FT of a Gaussan s a Gaussan. compare to box functon Ths s referred to as the Convoluton Theorem

Smoothng by Averagng Kernel: An Isotropc Gaussan The pcture shows a smoothng kernel proportonal to exp x2 + y 2 2σ 2 (whch s a reasonable model of a crcularly symmetrc fuzzy blob) Smoothng wth a Gaussan Kernel: Effcent Implementaton Both, the BOX flter and the Gaussan flter are separable: Frst convolve each row wth a 1D flter Then convolve each column wth a 1D flter. Other Types of Nose Impulsve nose randomly pck a pxel and randomly set ot a value saturated verson s called salt and pepper nose Some other useful flterng technques Medan flter Ansotropc dffuson Quantzaton effects Often called nose although t s not statstcal Unantcpated mage structures Also often called nose although t s a real repeatable sgnal.

Medan flters : prncple Medan flters: Example for wndow sze of 3 Method : 1. rank-order neghborhood ntenstes 2. take mddle value Input Sgnal Medan Fltered sgnal 1,1,1,7,1,1,1,1?,1,1,1.1,1,1,? non-lnear flter no new grey levels emerge... Advantage of ths type of flter s that t Elmnates spkes (salt & peper nose). Medan flters : example flters have wdth 5 : Medan flters : analyss medan completely dscards the spke, lnear flter always responds to all aspects medan flter preserves dscontnutes, lnear flter produces roundng-off effects DON T become all too optmstc Medan flter : mages 3 x 3 medan flter : Medan flters : Gauss revsted Comparson wth Gaussan : sharpens edges, destroys edge cusps and protrusons e.g. upper lp smoother, eye better preserved

Example of medan 10 tmes 3 X 3 medan Flters are templates Applyng a flter at some pont can be seen as takng a dotproduct between the mage and some vector Flterng the mage s a set of dot products Insght flters look lke the effects they are ntended to fnd flters fnd effects they look lke patchy effect mportant detals lost (e.g. ear-rng)