Johann Radon Institute for Computational and Applied Mathematics: www.ricam.oeaw.ac.at 1/39 The Gaussian Kernel, Regularization.

Slides:



Advertisements
Similar presentations
Ter Haar Romeny, MICCAI 2008 Regularization and scale-space.
Advertisements

Ter Haar Romeny, FEV Geometry-driven diffusion: nonlinear scale-space – adaptive scale-space.
Ter Haar Romeny, ICPR 2010 Introduction to Scale-Space and Deep Structure.
Johann Radon Institute for Computational and Applied Mathematics: 1/38 Mathematische Grundlagen in Vision & Grafik ( ) more.
Johann Radon Institute for Computational and Applied Mathematics: 1/35 Signal- und Bildverarbeitung, Image Analysis and Processing.
November 12, 2013Computer Vision Lecture 12: Texture 1Signature Another popular method of representing shape is called the signature. In order to compute.
Chapter 3 Image Enhancement in the Spatial Domain.
Ter Haar Romeny, Computer Vision 2014 Geometry-driven diffusion: nonlinear scale-space – adaptive scale-space.
Johann Radon Institute for Computational and Applied Mathematics: 1/33 Signal- und Bildverarbeitung, Image Analysis and Processing.
Computer vision: models, learning and inference Chapter 13 Image preprocessing and feature extraction.
Signal- und Bildverarbeitung, 323
Data mining and statistical learning - lecture 6
Environmental Data Analysis with MatLab Lecture 11: Lessons Learned from the Fourier Transform.
ECE 472/572 - Digital Image Processing Lecture 8 - Image Restoration – Linear, Position-Invariant Degradations 10/10/11.
Properties of continuous Fourier Transforms
Image Filtering CS485/685 Computer Vision Prof. George Bebis.
1 Image filtering Hybrid Images, Oliva et al.,
Edge detection. Edge Detection in Images Finding the contour of objects in a scene.
Lecture Notes for CMPUT 466/551 Nilanjan Ray
Image processing. Image operations Operations on an image –Linear filtering –Non-linear filtering –Transformations –Noise removal –Segmentation.
Lecture 3 Review of Linear Algebra Simple least-squares.
Quality Control Procedures put into place to monitor the performance of a laboratory test with regard to accuracy and precision.
Texture Reading: Chapter 9 (skip 9.4) Key issue: How do we represent texture? Topics: –Texture segmentation –Texture-based matching –Texture synthesis.
Basic Image Processing January 26, 30 and February 1.
Digital Image Processing, 2nd ed. © 2002 R. C. Gonzalez & R. E. Woods Chapter 4 Image Enhancement in the Frequency Domain Chapter.
Johann Radon Institute for Computational and Applied Mathematics: Summary and outlook.
Johann Radon Institute for Computational and Applied Mathematics: 1/49 Signal- und Bildverarbeitung, silently converted to:
Elec471 Embedded Computer Systems Chapter 4, Probability and Statistics By Prof. Tim Johnson, PE Wentworth Institute of Technology Boston, MA Theory and.
Today – part 2 The differential structure of images
Image Representation Gaussian pyramids Laplacian Pyramids
Johann Radon Institute for Computational and Applied Mathematics: 1/48 Mathematische Grundlagen in Vision & Grafik ( ) more.
© Copyright McGraw-Hill CHAPTER 6 The Normal Distribution.
Digital Image Processing Chapter # 4 Image Enhancement in Frequency Domain Digital Image Processing Chapter # 4 Image Enhancement in Frequency Domain.
Transforms. 5*sin (2  4t) Amplitude = 5 Frequency = 4 Hz seconds A sine wave.
Image Processing © 2002 R. C. Gonzalez & R. E. Woods Lecture 4 Image Enhancement in the Frequency Domain Lecture 4 Image Enhancement.
ECE 8443 – Pattern Recognition ECE 8423 – Adaptive Signal Processing Objectives: Deterministic vs. Random Maximum A Posteriori Maximum Likelihood Minimum.
1 PATTERN COMPARISON TECHNIQUES Test Pattern:Reference Pattern:
Curve-Fitting Regression
ECE 8443 – Pattern Recognition LECTURE 07: MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION Objectives: Class-Conditional Density The Multivariate Case General.
SUPA Advanced Data Analysis Course, Jan 6th – 7th 2009 Advanced Data Analysis for the Physical Sciences Dr Martin Hendry Dept of Physics and Astronomy.
Course 2 Image Filtering. Image filtering is often required prior any other vision processes to remove image noise, overcome image corruption and change.
Axioms Arjan Kuijper Axioms; PhD course on Scale Space, Cph 1-5 Dec Axioms of measurements and scale I1I2I3O KYBL1F1APN L2F2 Convolution.
CSC508 Convolution Operators. CSC508 Convolution Arguably the most fundamental operation of computer vision It’s a neighborhood operator –Similar to the.
1 Methods in Image Analysis – Lecture 3 Fourier CMU Robotics Institute U. Pitt Bioengineering 2630 Spring Term, 2004 George Stetten, M.D., Ph.D.
Fourier Transform.
Chapter 20 Statistical Considerations Lecture Slides The McGraw-Hill Companies © 2012.
CHAPTER 2.3 PROBABILITY DISTRIBUTIONS. 2.3 GAUSSIAN OR NORMAL ERROR DISTRIBUTION  The Gaussian distribution is an approximation to the binomial distribution.
Basic Theory (for curve 01). 1.1 Points and Vectors  Real life methods for constructing curves and surfaces often start with points and vectors, which.
Computer Graphics & Image Processing Chapter # 4 Image Enhancement in Frequency Domain 2/26/20161.
Machine Vision Edge Detection Techniques ENT 273 Lecture 6 Hema C.R.
Computer vision. Applications and Algorithms in CV Tutorial 3: Multi scale signal representation Pyramids DFT - Discrete Fourier transform.
Instructor: Mircea Nicolescu Lecture 7
Filters– Chapter 6. Filter Difference between a Filter and a Point Operation is that a Filter utilizes a neighborhood of pixels from the input image to.
Miguel Tavares Coimbra
PDE Methods for Image Restoration
Technique for Heat Flux Determination
- photometric aspects of image formation gray level images
LECTURE 09: BAYESIAN ESTIMATION (Cont.)
SIGNALS PROCESSING AND ANALYSIS
Linear Filters and Edges Chapters 7 and 8
Image Enhancement in the
Computational Data Analysis
Digital Image Processing
Outline Linear Shift-invariant system Linear filters
CSCE 643 Computer Vision: Image Sampling and Filtering
Introduction: A review on static electric and magnetic fields
Basic Image Processing
Digital Image Processing Week IV
Intensity Transformation
16. Mean Square Estimation
Presentation transcript:

Johann Radon Institute for Computational and Applied Mathematics: 1/39 The Gaussian Kernel, Regularization

Johann Radon Institute for Computational and Applied Mathematics: 2/39 The Gaussian Kernel The Gaussian kernel Normalization Cascade property, selfsimilarity The scale parameter Relation to generalized functions Separability Relation to binomial coefficients The Fourier transform of the Gaussian kernel Central limit theorem Anisotropy The diffusion equation Taken from B. M. ter Haar Romeny, Front-End Vision and Multi-scale Image Analysis, Dordrecht, Kluwer Academic Publishers, Chapter 3

Johann Radon Institute for Computational and Applied Mathematics: 3/39 The Gaussian Kernel The  determines the width of the Gaussian kernel. In statistics, when we consider the Gaussian probability density function it is called the standard deviation, and the square of it,  2, the variance. The scale can only take positive values,  >0. The scale-dimension is not just another spatial dimension.

Johann Radon Institute for Computational and Applied Mathematics: 4/39 Normalization The term in front of the Gaussian kernel is the normalization constant. With the normalization constant this Gaussian kernel is a normalized kernel, i.e. its integral over its full domain is unity for every .

Johann Radon Institute for Computational and Applied Mathematics: 5/39 Normalization This means that increasing the  of the kernel reduces the amplitude substantially. The normalization ensures that the average greylevel of the image remains the same when we blur the image with this kernel. This is known as average grey level invariance.

Johann Radon Institute for Computational and Applied Mathematics: 6/39 Cascade property, self-similarity The shape of the kernel remains the same, irrespective of the . When we convolve two Gaussian kernels we get a new wider Gaussian with a variance  2 which is the sum of the variances of the constituting Gaussians: The Gaussian is a self-similar function. Convolution with a Gaussian is a linear operation, so a convolution with a Gaussian kernel followed by a convolution with again a Gaussian kernel is equivalent to convolution with the broader kernel.

Johann Radon Institute for Computational and Applied Mathematics: 7/39 The scale parameter In order to avoid the summing of squares, one often uses the following parameterization: 2  2 = t To make the self-similarity of the Gaussian kernel explicit, we can introduce a new dimensionless (natural) spatial parameter: … and obtain the natural Gaussian kernel

Johann Radon Institute for Computational and Applied Mathematics: 8/39 Relation to generalized functions The Gaussian kernel is the physical equivalent of the mathematical point. It is not strictly local, like the mathematical point, but semi-local. It has a Gaussian weighted extent, indicated by its inner scale . Focus on some mathematical notions that are directly related to the sampling of values from functions and their derivatives at selected points. These mathematical functions are the generalized functions, i.e. the Dirac Delta-function, the Heaviside function and the error function.

Johann Radon Institute for Computational and Applied Mathematics: 9/39 Dirac delta function  (x) is everywhere zero except in x = 0, where it has infinite amplitude and zero width; its area is unity.

Johann Radon Institute for Computational and Applied Mathematics: 10/39 The error function The integral of the Gaussian kernel from -  to x is the error function, or cumulative Gaussian function: The result is so re-parameterzing is needed: -> natural coordinates!

Johann Radon Institute for Computational and Applied Mathematics: 11/39 The Heavyside function When the inner scale  of the error function goes to zero, we get the Heavyside function or unitstep function. The derivative of the Heavyside function is the Delta-Dirac function. The derivative of the error function is the Gaussian kernel.

Johann Radon Institute for Computational and Applied Mathematics: 12/39 Separability The Gaussian kernel for dimensions higher than one, say N, can be described as a regular product of N one- dimensional kernels. g 2 D ( x ; y;¾ 2 ) = g 1 D ( x;¾ 2 ) + g 1 D ( y;¾ 2 )

Johann Radon Institute for Computational and Applied Mathematics: 13/39 Relation to binomial coefficients The coefficients of this expansion are the binomial coefficients ('n over m')

Johann Radon Institute for Computational and Applied Mathematics: 14/39 The Fourier transform the Fourier transform: the inverse Fourier transform: The Fourier transform of the Gaussian function is again a Gaussian function, but now of the frequency .

Johann Radon Institute for Computational and Applied Mathematics: 15/39 Central limit theorem The central limit theorem: any repetitive operator goes in the limit to a Gaussian function. Example: a repeated convolution of two blockfunctions with each other.

Johann Radon Institute for Computational and Applied Mathematics: 16/39 Anisotropy The Gaussian kernel as specified above is isotropic, which means that the behavior of the function is in any direction the same. When the standard deviations in the different dimensions are not equal, we call the Gaussian function anisotropic.

Johann Radon Institute for Computational and Applied Mathematics: 17/39 The diffusion equation The Gaussian function is the solution of the linear diffusion equation: The diffusion equation can be derived from physical principles: the luminance can be considered a flow, that is pushed away from a certain location by a force equal to the gradient. The divergence of this gradient gives how much the total entity (luminance in our case) diminishes with time.  L =  (D  L)

Johann Radon Institute for Computational and Applied Mathematics: 18/39 Summary The normalized Gaussian kernel has an area under the curve of unity. Two Gaussian functions can be cascaded, to give a Gaussian convolution result which is equivalent to a kernel with the variance equal to the sum of the variances of the constituting Gaussian kernels. The spatial parameter normalized over scale is called the dimensionless 'natural coordinate'. The Gaussian kernel is the 'blurred version' of the Dirac Delta function. The cumulative Gaussian function is the Error function, which is the 'blurred version' of the Heavyside stepfunction. The central limit theorem states that any finite kernel, when repeatedly convolved with itself, leads to the Gaussian kernel. Anisotropy of a Gaussian kernel means that the scales, or standard deviations, are different for the different dimensions. The Fourier transform of a Gaussian kernel acts as a low-pass filter for frequencies. The Fourier transform has the same Gaussian shape. The Gaussian kernel is the only kernel for which the Fourier transform has the same shape. The diffusion equation describes the expel of the flow of some quantity (intensity, temperature, …) over space under the force of a gradient.

Johann Radon Institute for Computational and Applied Mathematics: 19/39 The Gaussian Kernel, Regularization

Johann Radon Institute for Computational and Applied Mathematics: 20/39 Differentiation and regularization Regularization Regular tempered distributions and testfunctions An example of regularization Relation regularization and Gaussian scale space Taken from B. M. ter Haar Romeny, Front-End Vision and Multi-scale Image Analysis, Dordrecht, Kluwer Academic Publishers, Chapter 8

Johann Radon Institute for Computational and Applied Mathematics: 21/39 Regularization Regularization is the technique to make data behave well when an operator is applied to them. Such data could e.g. be functions, that are impossible or difficult to differentiate, or discrete data where a derivate seems to be not defined at all. From physical principles – images are physical entities - this implies that when we consider a system, a small variation of the input data should lead to small change in the output data. Differentiation is a notorious function with 'bad behavior'.

Johann Radon Institute for Computational and Applied Mathematics: 22/39 Regularization Differentiation is not well-defined.

Johann Radon Institute for Computational and Applied Mathematics: 23/39 Regularization A solution is well-defined in the sense of Hadamard if the solution –Exists –Is uniquely defined –Depends continuously on the initial or boundary data The operation is the problem, not the function. And what about discrete data?

Johann Radon Institute for Computational and Applied Mathematics: 24/39 Regularization How should the derivative of this thing look like? Regularize the data or the operation?

Johann Radon Institute for Computational and Applied Mathematics: 25/39 Laurent Schwartz: use the Schwartz space with smooth test functions –Infinitely differentiable –decrease fast to zero at the boundaries Construct a regular tempered distribution –i.e. the integral of a test function and “something” The regular tempered distribution now has the nice properties of the test function. It can be regarded as a probing of “something” with a mathematically nice filter. Regular tempered distributions and testfunctions

Johann Radon Institute for Computational and Applied Mathematics: 26/39 Regular tempered distributions and testfunctions Smooth test functions –Infinitely differentiable –decrease fast to zero at the boundaries For example a Gaussian. The regular tempered distribution = The filtered image Now everything is well-defined, since integrating is well-defined. Do everything “under the integral” No data smoothing needed.

Johann Radon Institute for Computational and Applied Mathematics: 27/39 An example of regularization

Johann Radon Institute for Computational and Applied Mathematics: 28/39 Regularization and Gaussian scale space When data are regularized by one of the methods above that 'smooth' the data, choices have to be made as how to fill in the 'space' in between the data that are not given by the original data. In particular, one has to make a choice for the order of the spline, the order of fitting polynomial function, the 'stiffness' of the physical model etc. This is in essence the same choice as the scale to apply in scale-space theory. The well known and much applied method of regularization as proposed by Tikhonov and Arsenin (often called 'Tikhonov regularization') is essentially equivalent to convolution with a Gaussian kernel.

Johann Radon Institute for Computational and Applied Mathematics: 29/39 Regularization and Gaussian scale space Try to find the minimum of a functional E(g), where g is the regularized version of f, given a set of constraints. This constraint is the following: we also like the first derivative of g to x (g x ) to behave well: we require that when we integrate the square of over its total domain we get a finite result. The method of the Euler-Lagrange equations specifies the construction of an equation for the function to be minimized where the constraints are added with a set of constant factors, one for each constraint, the so- called Lagrange multipliers.

Johann Radon Institute for Computational and Applied Mathematics: 30/39 Regularization and Gaussian scale space The functional becomes The minimum is obtained at Simplify things: go to Fourier space: –1) Parceval theorem: the Fourier transform of the square of a function is equal to the square of the function itself. –2)

Johann Radon Institute for Computational and Applied Mathematics: 31/39 Regularization and Gaussian scale space Therefore: So

Johann Radon Institute for Computational and Applied Mathematics: 32/39 Regularization and Gaussian scale space Back in the spatial domain: This is a first result for the inclusion of the constraint for the first order derivative. However, we like our function to be regularized with all derivatives behaving nicely, i.e. square integrable.

Johann Radon Institute for Computational and Applied Mathematics: 33/39 Continue the procedure:

Johann Radon Institute for Computational and Applied Mathematics: 34/39 Regularization and Gaussian scale space Back in the spatial domain: This is a result for the inclusion of the constraint for the first and second order derivative. However, we like our function to be regularized with all derivatives behaving nicely, i.e. square integrable.

Johann Radon Institute for Computational and Applied Mathematics: 35/39 Regularization and Gaussian scale space General form: How to choose the Lagrange multipliers? –Cascade property / scale invariance for the filters: –Computing per power one obtains

Johann Radon Institute for Computational and Applied Mathematics: 36/39 Regularization and Gaussian scale space This results in (s=½  2,  =1) The Taylor series of the Gaussian in Fourier space:

Johann Radon Institute for Computational and Applied Mathematics: 37/39 Summary Many functions can not be differentiated. –The solution, due to Schwartz, is to regularize the data by convolving them with a smooth test function. –Taking the derivative of this 'observed' function is then equivalent to convolving with the derivative of the test function. A well know variational form of regularization is given by the so-called Tikhonov regularization. –A functional is minimized in sense with the constraint of well behaving derivatives. –Tikhonov regularization with inclusion of the proper behavior of all derivatives is essentially equivalent to Gaussian blurring.

Johann Radon Institute for Computational and Applied Mathematics: 38/39 Next Time Gaussian derivatives –Shape and algebraic structure –Gaussian derivatives in the Fourier domain –Zero crossings of Gaussian derivative functions –The correlation between Gaussian derivatives –Discrete Gaussian kernels –Other families of kernels Natural limits on observations –Limits on differentiation: scale, accuracy and order Deblurring Gaussian blur –Deblurring –Deblurring with a scale-space approach –Less accurate representation, noise and holes Multiscale derivatives: implementations –Implementation in the spatial domain –Separable implementation –Some examples –N-dim Gaussian derivative operator implementation –Implementation in the Fourier domain –Boundaries

Johann Radon Institute for Computational and Applied Mathematics: 39/39 In some weeks The differential structure of images –Differential image structure –Isophotes and flowlines –Coordinate systems and transformations –Directional derivatives –First order gauge coordinates –Gauge coordinate invariants: examples Ridge detection Isophote and flowline curvature in gauge coordinates Affine invariant corner detection –A curvature illusion –Second order structure The Hessian matrix and principal curvatures The shape index Principal directions Gaussian and mean curvature Minimal surfaces, zero Gaussian curvature surfaces –Third order image structure: T-junction detection –Fourth order image structure: junction detection –Scale invariance and natural coordinates –Irreducible invariants Tensor notation