|
||||||
Inst. för Matematik | KTH | | ||||||
|
||||||
This is an early meating, in fact one day before the official semester start. It is not a problem for new student to start at a later occation. We need to get some idea how many will attend the course, so we get an appropriate size lecture room. Preliminary we plan lectures on Tuesdays 1-3 In this first couse meeting the were given some introductatary information about the administration of the course. Also ther were given some very general and very short description of wavelets as wave packages and some application of it to imageprocessing were demonstrated. In this wavelet course we will assume that the students have some knowledge in two basic areas in mathematics: 1. Linera algebra, with basis systems , specially orthormlal basis. 2. Fourier series and/or Fourier transforms. The student are assumed to have accress to Matlab during the course In this course we will jump back and forth between four scenarious (or types) of functions: (2 x 2 = 4 kombinations) Countinuous / Discrete Non-period / Periodic functions and their "fourier" transforms. (Some of which topics most student have already seen and other topics may be new for most students) Also, if I get enough time we plan to describe the system of Haar functions. Lecture 1 (September 4) A short repition from first
introductary meeting about the four senarious of functions , the Homework assignmet 1 was handed
out, it is also available here. Lecture 2 (September 11) We saw that the expansion of function with the Haar filter corresponds to filter operations on the sequence space l^2 with two filter the Lowpass filter h = (1,1)/sqrt(2) and the Highpass filter g=(-1,1)/sqrt(2) , arranging the iterated filter operation in the Wavelet filter tree . The filter h and g generatate the translation invariant ON-sets {T^2k g}_k resp. {T^2k g}_k . Those two ON-sets are mutually orthogonal and makes together an ON- basis for l^2. The Low- and High- pass filter operation can be seen as changing coordinates system to this new ON-basis. This change of coordinates can be seen as locally doing an 45 degree in many copies of R². By repeated local rotations (with specially selected angles) one will construct longer filters h and g with will generate similar translation invariant ON-sets and ON- basis. Usually one selects the rotation angles so that the constructed filter fullfill some moment conditions. Lecture 3 (September 18) Given a filter h of finite length such that {T^2k h}_k is a ON-set we saw how to get the complimentary filter g such that {T^2k h}_k union {T^2k g}_k will be a ON-basis for l^2. We saw also how to obtain such filters h of arbitary length 2m by m local rotations A local rotations is defined on pairvise coodinates (a_2k, a_2k+1) by the matrix cos A sin A R_A= ( ) each local rotation is interwined by a translation T: -sin A cos A Thus h = R_A1 T RA_2......TR_Am Dirac_0 and g = R_A1 T RA_2......TR_Am Dirac_1. gives ON-filter of length 2m. The rotation angels A1,...,Am are chose usually so that g satisfies some moment conditions. Homwork number 2 is available here. Inlevering utsatt till 9 oktober pga sen WEB-publisering. Lecture 4 (September 25) First some hints about doing the filtration in matlab. Matlab start indexing arrays with index one. Often Low- and highpass filters involves filters where some non-zero elements have negative index. These implies some shifting or other tricks is necessary when using Matlab. If one consider a given finites set of data surrounded by zero elements, the Low- and Highpass filtering with result in a totally larger set of coefficients then the data points. To get around this one often consider the data as a set of periodic data. This means that the convolution in Matlab for Low and Highpass filter must be followed by a periodization. We saw how the ON-conditions of the filters h and g can be written in term of their Fourier transform och Z-transforms. Multi-scale analysis. We stated sixth conditions for multi-scale analysis, for a family of subspace V_j of L^2 (R) and saw how this leads to a a filter h such that {T^2k h}_k is a ON-set. We also saw how to use the complimentaty fileter g to construct the corresponding ON- wavelet basis. Lecture 5 (October 2) Here is a list of topics from lecture 5. Homework number 3 was handed out. Lecture 6 (October 9) Using the scale equation we can construct the scaling function by iteration . This is the cascade algorithm for filters h of length 4 can be found at http://plan9.bell-labs.com/who/wim/cascade/ In this lecture we also saw how an image which is represented of a matrix with grey-scale pixelvalues may be filtered using lowpass and highpass filter in the x- and the y directions. We saw that using this procedure in several levels we with will get the wavelet coefficients of this matrix and that most of the coefficient will be very small. By setting the small coefficient to zero, we can represent the image by a small part of'the coefficients. This leads to compression algorithms. Lecture 7 (October 16) Mathlab hints: use the built in Matlab functions sum and reshape to perform periodizations without loops. We look at bi-orthonormal basis, and biorthonormal wavelet filters. bi-orthonormal filter can be spit into a product of filters of length 3-1. a so called lifting. Using this liftings gives a very fast implementation of the bi-orthonormal filtering procedure. Convolution of filter related to convolution of scalingfunctions by the cascade algorithm. Th spline filter b_n=(1,1) * ....* (1,1) i.e iterated convolution with n (1,1) filters. Let b_2n ,be the bi-orthogonal wavelet scaling filter (length = 2n+1), then its dual filter b^*_2n of length 2n-1, can be found by solving a few linear equations. The filter I_n = b_2n * (b^_2n)^ will be a so called interpolet filter of length 4n-1 (whose dual is the Dirac_0-filter.) With some numerical efforts one may split I_n as I_n=h_n * (h_n)^~. In this way we have constructed an orthonormal wavelet filter h_n of length 2n. Homework number 4 was handed out. Lecture 8 Estimate of size of wavelet coefficients. Moment conditions on wavelet functions and scaling functions. Lowpass and higpass filter of polynomial sequences. The moments of a wavelet with m times continuous derivatives. Tresholding of wavelet coefficientents Lecture 9 The time-frequency plane, the uncertainty principle. The continuous wavelet transform. Homework number 5 was handed out. No lectures on November 6 and November 13 Here is the last homework ( with a correction of index in a summation formula) Homework number 6 Kom ihåg att fylla i kursutvärderingen. En länk till den finns här
|
||||||