Fast DCT-based image convolution algorithms and application to image resampling and hologram reconstruction

Similar documents
Chapter 3 Optical Systems with Annular Pupils

Planar Nearfield Acoustical Holography in High-Speed, Subsonic Flow

15 Solving the Laplace equation by Fourier method

DOING PHYSICS WITH MATLAB COMPUTATIONAL OPTICS

An Exact Solution of Navier Stokes Equation

School of Electrical and Computer Engineering, Cornell University. ECE 303: Electromagnetic Fields and Waves. Fall 2007

Absorption Rate into a Small Sphere for a Diffusing Particle Confined in a Large Sphere

Tutorial on Strehl ratio, wavefront power series expansion, Zernike polynomials expansion in small aberrated optical systems By Sheng Yuan

Analytical Solutions for Confined Aquifers with non constant Pumping using Computer Algebra

3.1 Random variables

Rigid Body Dynamics 2. CSE169: Computer Animation Instructor: Steve Rotenberg UCSD, Winter 2018

Pearson s Chi-Square Test Modifications for Comparison of Unweighted and Weighted Histograms and Two Weighted Histograms

{ } ( ) ( ) ( 1 ( ) π 2. Trapezoidal rule: a ( ) ( ) Simpson rule: 3/8 Simpson rule: Boole's rule a ( ) ( ) ( ) ( + ) 45 Weddle's Rule Hardy's rule:

Numerical Inversion of the Abel Integral Equation using Homotopy Perturbation Method

Lecture 8 - Gauss s Law

Absolute Specifications: A typical absolute specification of a lowpass filter is shown in figure 1 where:

Course Outline. ECE 178: Image Processing REVIEW. Relationship between pixels. Connected components. Distance Measures. Linear systems-review

Lecture 04: HFK Propagation Physical Optics II (Optical Sciences 330) (Updated: Friday, April 29, 2005, 8:05 PM) W.J. Dallas

HOW TO TEACH THE FUNDAMENTALS OF INFORMATION SCIENCE, CODING, DECODING AND NUMBER SYSTEMS?

Hydroelastic Analysis of a 1900 TEU Container Ship Using Finite Element and Boundary Element Methods

Aalborg Universitet. Load Estimation from Natural input Modal Analysis Aenlle, Manuel López; Brincker, Rune; Canteli, Alfonso Fernández

The Substring Search Problem

An Application of Fuzzy Linear System of Equations in Economic Sciences

Stanford University CS259Q: Quantum Computing Handout 8 Luca Trevisan October 18, 2012

ASTR415: Problem Set #6

Pulse Neutron Neutron (PNN) tool logging for porosity Some theoretical aspects

APPLICATION OF MAC IN THE FREQUENCY DOMAIN

Conservative Averaging Method and its Application for One Heat Conduction Problem

DonnishJournals

COMPUTATIONS OF ELECTROMAGNETIC FIELDS RADIATED FROM COMPLEX LIGHTNING CHANNELS

Revision of Lecture Eight

Fresnel Diffraction. monchromatic light source

Quantum Fourier Transform

Scattering in Three Dimensions

Duality between Statical and Kinematical Engineering Systems

Numerical Integration

Light Time Delay and Apparent Position

ME 3600 Control Systems Frequency Domain Analysis

I. CONSTRUCTION OF THE GREEN S FUNCTION

COUPLED MODELS OF ROLLING, SLIDING AND WHIRLING FRICTION

C/CS/Phys C191 Shor s order (period) finding algorithm and factoring 11/12/14 Fall 2014 Lecture 22

A NEW VARIABLE STIFFNESS SPRING USING A PRESTRESSED MECHANISM

A Backward Identification Problem for an Axis-Symmetric Fractional Diffusion Equation

Mathematical Model of Magnetometric Resistivity. Sounding for a Conductive Host. with a Bulge Overburden

Recent Advances in Chemical Engineering, Biochemistry and Computational Chemistry

New problems in universal algebraic geometry illustrated by boolean equations

Method for Approximating Irrational Numbers

Magnetometer Calibration Algorithm Based on Analytic Geometry Transform Yongjian Yang, Xiaolong Xiao1,Wu Liao

Using Laplace Transform to Evaluate Improper Integrals Chii-Huei Yu

Research Article On Alzer and Qiu s Conjecture for Complete Elliptic Integral and Inverse Hyperbolic Tangent Function

ON INDEPENDENT SETS IN PURELY ATOMIC PROBABILITY SPACES WITH GEOMETRIC DISTRIBUTION. 1. Introduction. 1 r r. r k for every set E A, E \ {0},

No. 6 Cell image ecognition with adial hamonic Fouie moments 611 ecognition expeiment is pefomed. Results show that as good image featues, RHFMs ae su

FREE TRANSVERSE VIBRATIONS OF NON-UNIFORM BEAMS

Identification of the degradation of railway ballast under a concrete sleeper

Math 124B February 02, 2012

Application of homotopy perturbation method to the Navier-Stokes equations in cylindrical coordinates

$ i. !((( dv vol. Physics 8.02 Quiz One Equations Fall q 1 q 2 r 2 C = 2 C! V 2 = Q 2 2C F = 4!" or. r ˆ = points from source q to observer

Quantum Mechanics II

EQUATIONS OF MOTION LUCA GUIDO MOLINARI

IN SITU SOUND ABSORPTION COEFFICIENT MEASUREMENT OF VARIOUS SURFACES

Graphs of Sine and Cosine Functions

MAGNETIC FIELD AROUND TWO SEPARATED MAGNETIZING COILS

Analysis and Optimization of a Special Type of Dielectric Loaded Resonant Cavity for Mobile Communication Filters

( ) [ ] [ ] [ ] δf φ = F φ+δφ F. xdx.

2 x 8 2 x 2 SKILLS Determine whether the given value is a solution of the. equation. (a) x 2 (b) x 4. (a) x 2 (b) x 4 (a) x 4 (b) x 8

Chapter 5 Linear Equations: Basic Theory and Practice

DIMENSIONALITY LOSS IN MIMO COMMUNICATION SYSTEMS

NOTE. Some New Bounds for Cover-Free Families

11) A thin, uniform rod of mass M is supported by two vertical strings, as shown below.

Modeling Fermi Level Effects in Atomistic Simulations

On the integration of the equations of hydrodynamics

University Physics (PHY 2326)

Transverse Wakefield in a Dielectric Tube with Frequency Dependent Dielectric Constant

LC transfer of energy between the driving source and the circuit will be a maximum.

2 Governing Equations

Localization of Eigenvalues in Small Specified Regions of Complex Plane by State Feedback Matrix

Swissmetro: design methods for ironless linear transformer

Analytical evaluation of 3D BEM integral representations using complex analysis

Right-handed screw dislocation in an isotropic solid

ANALYSIS OF PRESSURE VARIATION OF FLUID IN AN INFINITE ACTING RESERVOIR

Chem 453/544 Fall /08/03. Exam #1 Solutions

Chebyshev Polynomial Based Numerical Inverse Laplace Transform Solutions of Linear Volterra Integral and Integro-Differential Equations

ELASTIC ANALYSIS OF CIRCULAR SANDWICH PLATES WITH FGM FACE-SHEETS

transformation Earth V-curve (meridian) λ Conical projection. u,v curves on the datum surface projected as U,V curves on the projection surface

A scaling-up methodology for co-rotating twin-screw extruders

Application of Parseval s Theorem on Evaluating Some Definite Integrals

QUALITATIVE AND QUANTITATIVE ANALYSIS OF MUSCLE POWER

The second law of thermodynamics - II.

A DETAILED STUDY OF THE HIGH ORDER SERIAL RESONANT INVERTER FOR INDUCTION HEATING

Multiple Criteria Secretary Problem: A New Approach

Lecture 10. Vertical coordinates General vertical coordinate

A Relativistic Electron in a Coulomb Potential

Fractional Tikhonov regularization for linear discrete ill-posed problems

QUANTUM ALGORITHMS IN ALGEBRAIC NUMBER THEORY

Power efficiency and optimum load formulas on RF rectifiers featuring flow-angle equations

radians). Figure 2.1 Figure 2.2 (a) quadrant I angle (b) quadrant II angle is in standard position Terminal side Terminal side Terminal side

Uniform Circular Motion

Simulation and Discretization of Fractional Order Systems

Vector d is a linear vector function of vector d when the following relationships hold:

Safety variations in steel designed using Eurocode 3

Transcription:

Fast DCT-based image convolution algoithms and application to image esampling and hologam econstuction Leonid Bilevich* a and Leonid Yaoslavsy** a a Depatment of Physical Electonics, Faculty of Engineeing, Tel Aviv Univesity, 69978, Tel Aviv, sael ABSTRACT Convolution and coelation ae vey basic image pocessing opeations with numeous applications anging fom image estoation to taget detection to image esampling and geometical tansfomation. n eal time applications, the cucial issue is the pocessing speed, which implies mandatoy use of algoithms with the lowest possible computational complexity. Fast image convolution and coelation with lage convolution enels ae taditionally caied out in the domain of Discete Fouie Tansfom computed using Fast Fouie Tansfom algoithms. Howeve standad DFT based convolution implements cyclic convolution athe than linea one and, because of this, suffes fom heavy bounday effects. We intoduce a fast DCT based convolution algoithm, which is vitually fee of bounday effects of the cyclic convolution. We show that this algoithm have the same o even lowe computational complexity as DFT-based algoithm and demonstate its advantages in application examples of image abitay tanslation and scaling with pefect discete sinc-intepolation and fo image scaled econstuction fom hologams digitally ecoded in nea and fa diffaction zones. n geometical esampling the scaling by abitay facto is implemented using the DFT domain scaling algoithm and DCT-based convolution. n scaled hologam econstuction in fa diffaction zones the Fouie econstuction method with simultaneous scaling is implemented using DCT-based convolution. n scaled hologam econstuction in nea diffaction zones the convolutional econstuction algoithm is implemented by the DCT-based convolution. A linea (apeiodic) convolution. TRODUCTO ~ a a h hna n (, K, ) n is a fundamental basic opeation in digital image pocessing. t is implemented by fast algoithms in tansfom domain, usually in DFT domain though FFT algoithms. Howeve, these algoithms compute a cyclic (cicula) convolution: athe than a linea one. a h n hna( n) mod (, K, Because in digital pocessing signals ae always given by a finite numbe of thei samples, any digital convolution algoithm equies one o anothe method fo detemination of signals and convolution enels outside thei boundaies. n ode to cope with this poblem, seveal DFT domain digital convolution methods wee suggested. n paticula, in Ref. [] zeo padding of both the signal and the enel is ecommended and in Ref. [] zeo padding of the signal and mio eflection of the enel is suggested. Howeve these methods, widely used in digital signal pocessing, suffe fom heavy bounday effects caused by discontinuities in signals due to zeo padding. A method of digital convolution that is vitually fee fom such bounday effects was suggested in [3-5]. Accoding to this method, signals ae extended to *bilevich@eng.tau.ac.il; http://www.eng.tau.ac.il/~bilevich/ **yao@eng.tau.ac.il; http://www.eng.tau.ac.il/~yao/ ) () ()

double length by thei mio eflection and the enel is zeo-padded to double length. This esults in the following convolution algoithm: [ DCT( DCT[ a ] R{ DFT[ SH( ZP ( h ))]}) + DcST( DCT[ a ] { DFT[ SH( ( h ))]} )] ~ a ZP whee DFT, DCT, DCT and DcST ae the Discete Fouie Tansfom, Discete Cosine Tansfom, nvese Discete Cosine Tansfom and nvese Discete Cosine-Sine Tansfom defined by: h a a, (3) DFT α DFT a a, (4) ( ) DCT + α DCT a a cos π, (5) ( DCT ) ( DCT ) ( DCT ) ( + ) α α + α cos π, (6) DCT ( DST ) ( DST ) ( DST ) ( + ) α α + α sin π ; (7) DcST ZP denotes a zeo-padding of { h to double length and SH denotes a shift by intege pat of the agument: R denotes a eal pat and factional tanslation algoithm [6]. SH } [ ZP ( h )] [ ZP ( h )] ( ) mod, whee denotes, (8) denotes an imaginay pat. The algoithm given by Eq. (3) was used to implement a p -. BOUDARY EFFECT SAFE DGTAL COVOLUTO DCT DOMA Algoithm given by Eq. (3) equies computing the nvese Discete Cosine-Sine Tansfom DcST which is not commonly available. Howeve DcST can be easily computed via DCT [7], and the Eq. (3) can be conveted into the following fom: + DCT( DCT[ a ] R{ DFT[ SH( ZP ( h ))]} ) ( ) { DcST( [ DCT[ a ] { DFT[ SH( ZP ( h ))]}] )}. This equation assumes computing DFT of the zeo-padded convolution enel of double length. Futhemoe, this equation can be conveted into the following fom that equies computing ove oiginal samples: DCT { [ ( ˆ DCT DCT a DCT )] DcST[ DCT DcST ( ˆ h + a h )]} [ [ ( ˆ DCT DCT a DCT )] { DCT[ ( DCT DcST ( ˆ h + a h ) ]} ], DcST whee and ae the type- Discete Cosine Tansfom and type- Discete Cosine-Sine Tansfom defined by: (9) ()

α K ( DCT ) DCT ( a ) a + ( ) a + a cos π (, K,,,, ), () α ( DST ) DcST ( a ) a sin π (, K,,, K, ) () and ĥ denotes a nomalized enel: hˆ h h if if if, K,. (3) Since the algoithm Eq. () opeates on signals of the oiginal length (and not double one) and fast FFT-type algoithms ae available fo computation of diffeent types of DCT/DcST tansfoms [8], the DCT convolution Eq. () can be computed in efficient way suited fo eal-time applications. 3. APPLCATOS The convolution algoithm Eq. () can be applied to vaious applications including esampling and hologam econstuction. As epesentative applications, we ll descibe a method of scaling and two methods of scaled hologaphic econstuction. 3. Scaling The input signal a has to be scaled by abitay scaling facto σ. n ode to map the cente of the input image into the cente of the scaled image, we have to intoduce the tanslational centeing facto : ( σ ) ( )σ (4) (the distance between the cente of the input image and the cente of the scaled image is equal to ). The scaled signal is computed in the following way [, 9]: ( σ ) ( )( ) SDFT (, ( ) ) ( a ) σ σ whee SDFT is the Shifted Discete Fouie Tansfom [4] defined by: ( ) σ, ( u)( v) ( u ) u, v, v uv α SDFT a a (5). (6) n the scaling method Eq. (5) the poblem of computation of scaled signal boils down to computation of convolution that is implemented in DCT domain using Eq. (). The peculia choice of phase factos in the Eq. (5) ensues that fo the eal input image the computed scaled image is eal and centeed coectly. Fo the case of zoom-out ( σ < ) the method Eq. (5) does not equie the low-pass pe-filteing of the input signal (because this pe-filteing is built-in in the DCT convolution Eq. (). The intepolation fomula of the scaling algoithm Eq. (5) is given by:

whee sincd σ n is a discete sinc function defined by: an sincd ; ; σ n, (7) Mx sin π sincd M ; ; x. (8) x sin π Examples of images obtained by the scaling method Eq. (5) with scaling facto σ implemented by the DFT convolution [] and by the DCT convolution Eq. () ae shown in Figue. As one can see, bounday effect in fom of heavy oscillations at the image bodes ae pesent in the esult of DFT convolution and ae absent in the esult of DCT convolution. Figue. Compaison of image scaling algoithms. Left: the oiginal image. Middle: the scaled image computed using DFT-domain convolution ( σ ). Right: the scaled image computed using DCT-domain convolution ( σ ). 3. Hologam econstuction with scaling n numeical econstuction of hologams it is fequently equied to econstuct images with diffeent scale facto commensuable with illumination wave length. The two most impotant methods of numeical econstuction of Fesnel hologams ae the Fouie econstuction algoithm (fo fa diffaction zones) and the convolutional econstuction algoithm (fo nea diffaction zones) [5, ]. The suggested above digital convolution algoithm can natually be employed fo solving this tas. Fo the Fouie econstuction algoithm, the econstuction and scaling can be pefomed simultaneously in one step while fo the convolutional econstuction algoithm the econstuction and scaling have to be pefomed in two steps. Fouie econstuction algoithm with scaling Fo Fouie econstuction algoithm, σ -scaled econstucted samples of the object wave font ae computed fom the α by the following fomula [5, ]: input hologam samples { }

[ σ ] σ ( µ σ + w) α ( µ σ ) w µ ( ) σ Theefoe the poblem of computation of scaled hologam econstuction boils down to the poblem of computation of convolution that can be implemented in DCT domain using Eq. (). Examples of Fouie hologam econstuction algoithm without scaling ( σ ) and with scaling ( σ ) ae shown in Figue.. (9) Figue. Fouie econstuction algoithm with scaling. Left: hologam econstuction without scaling ( ). Right: scaled hologam econstuction ( σ ). σ Convolutional econstuction algoithm with scaling Fo the convolutional algoithm, econstucted samples of the object wave ae computed fom the input hologam α by the following fomula [5, ]: samples { } whee fincd is a discete finc function defined by: a α fincd( ; µ ; w), () +

fincd ; q; x q x. () n this case, the poblem of hologam econstuction boils down to the poblem of computation of convolution that can be implemented in DCT domain using Eq. (). The function fincd is computed fom Eq. () using DFT. The scaled ( σ ) hologam econstuction a ~ is computed fom the hologam econstuction using scaling algoithm Eq. (5). Examples of convolutional hologam econstuction algoithm without scaling ( shown in Figue 3. a σ ) and with scaling ( σ ) ae Figue 3. Convolution econstuction algoithm with scaling. Left: hologam econstuction without scaling ( σ ). Right: scaled hologam econstuction ( σ ). 4. COCLUSOS The bounday effect safe DCT-domain convolution algoithm is pesented and applications of this algoithm to image escaling and scaled hologaphic econstuction ae povided. Thans to the availability of fast FFT-type algoithms fo computing tansfoms involved in the algoithm, the suggested DCT-domain convolution epesents a valuable altenative to DFT-domain convolution in eal-time video pocessing applications. REFERECES [] Stocham, T. G., High-speed convolution and coelation with applications to digital filteing, in [Digital Pocessing of Signals], Gold, B. and Rade, C. M., eds., 3 3, McGaw-Hill, nc. (969). [] Rabine, L. R., Schafe, R. W., and Rade, C. M., The chip z-tansfom algoithm and its application, Bell Syst. Tech. J. 48(5), 49 9 (May-June 969).

[3] Yaoslavsy, L., Bounday effect fee and adaptive discete signal sinc-intepolation algoithms fo signal and image esampling, Appl. Opt. 4(), 466 475 (July 3). [4] Yaoslavsy, L., [Digital Hologaphy and Digital mage Pocessing: Pinciples, Methods, Algoithms], Kluwe Academic Publishes (4). [5] Yaoslavsy, L., Discete tansfoms, fast algoithms, and point spead functions of numeical econstuction of digitally ecoded hologams, in [Advances in Signal Tansfoms: Theoy and Applications], Astola, J. and Yaoslavsy, L., eds., EURASP Boo Seies on Signal Pocessing and Communications 7, 93 4, Hindawi (7). [6] Yaoslavsy, L., Fast discete sinc-intepolation: A gold standad fo image esampling, in [Advances in Signal Tansfoms: Theoy and Applications], Astola, J. and Yaoslavsy, L., eds., EURASP Boo Seies on Signal Pocessing and Communications 7, 337 45, Hindawi (7). [7] Wang, Z., A fast algoithm fo the discete sine tansfom implemented by the fast cosine tansfom, EEE Tans. Acoust., Speech, Signal Pocessing 3(5), 84 85 (Oct. 98). [8] Bitana, V., Yip, P., and Rao, K. R., [Discete Cosine and Sine Tansfoms: Geneal Popeties, Fast Algoithms and ntege Appoximations], Elsevie Ltd. (7). [9] Deng, X., Bihai, B., Gan, J., Zhao, F., and Chen, R. T., Fast algoithm fo chip tansfoms with zooming-in ability and its applications, J. Opt. Soc. Am. A 7(4), 76 77 (Ap. ). [] Yaoslavsy, L., ntoduction to digital hologaphy, in [Digital Signal Pocessing in Expeimental Reseach], Yaoslavsy, L. and Astola, J., eds., Bentham E-boo Seies, 87 (9).