> ", A " "<n. :-t. 'O "" (l, -...I. c-j "';> \ J, \l\ , iy-j: "a-j C 'C) \.,S \\ \ /tv '-> .,,J c,. <tt .

Size: px

Start display at page:

Download "> ", A " "<n. :-t. 'O "" (l, -...I. c-j "';> \ J, \l\ , iy-j: "a-j C 'C) \.,S \\ \ /tv '-> .,,J c,. <tt ."

Abel Wells
6 years ago
Views:

1 ~

3 l' '" > A " '.",... "<n J. 'O "" (l,, :t I "...I cj '. "';> \ J, \l\, iyj: c It "aj C 'C) \.,S ( /tv... hj N \\ \ + <tt I '>.,,J c,..$ < 4S'

4 .,. r 1 """.., "" "",.. Q?., 0... < Jt... 2 :3 (c 1\4 1,. ", J ' 3 A 3 3,.. r \7 I : I + 1\ ", <<+ L:J \\. ' 4! t c."t + 3" \..., Q!1 ri " % j 1\ 3 ;: q\ if 1 ::> " ')) 11 3.,J\.,.g 0. '+ Q. (':!:, '"...J '',&.,..

5 Objectives of Image Coding. Representation of an image with acceptable quality, using as small a number of bits as possible Applications:. Reduction of channel bandwidth for image transmission. Reduction of required storage

6 Issues in Image Coding 1. What tocode? a. Image density b. Image transformcoefficients c. Image model parameters 2. How to assign reconstruction levels a. Uniform spacing between reconstruction levels b. Nonuniform spacing between reconstruction levels 3. Bit assi :nment "'" a. Equallength bit assignment to each reconstruction level b. Unequallength bit assignment to each reconstructionlevel 111'II9I IDUI'CI In. of bi1i /:;" I

7 1\1ethods of Reconstruction Level Assignments Assumptions:. Image intensity is to be coded. Equallength bit assignment Scalar Case 1. Equal spacing of reconstruction levels (Unifonn Quantization) (Ex): Image intensity f: Number of reconstruction levels: 4 (2 bits for equal bit assignment) 1; 1j 1j II. d, d& dj J,.,. I I I I I,.1, I I I I,,, f

8 f tunifonft qu"'ti_ ~; ; I1pn 10.3 Eumple or unirorm quantizer. The Dumber or reconstructioa levels is 4./ is assumed to be betweed0 and 1. and I is the result or quantizinl/. The reconstruction levels and decisiod, bounlhries are denoted by r. and d.. respectively.

9 Scalar Case (cont.) 2. Spacing based on some error criterion ri : ~construction levels (32, 96,.160, 224) di : decision boundaries (0, 64, 128, 192, 256) Optimally choose ri and di. To do this, assume f MJll < f < f max r" \'". J ~ the number of reconstruction levels p (j): probability density function for f Minimize fmax ri ' di : E = E [(f 1)2] = J (f 1)2 P (f) df f =f min ==> 0 => r. 2d. rj J J 'l = 0 => LloydMaxQuantizer These are not simple linear equations. J dj+1 d. f P (f) df rj = ~ J j+l p(f) df d. J

10 Scalar Case (cont.) Solution to the Optimization Problem TABLE10.1 PLACEMENTOF RECONSTRUCTIONANDDECISIONLEVELS FORLLOYDMAXaUANTlZER. FORUNIFORMPDF,p,(fo)ISASSUMED UNIFORM BETWEEN 1 AND 1. THE GAUSSIAN PDFIS ASSUMEDTO HAVEMEANOF0 ANDVARIANCEOF1. FORTHELAPLACIANPDF,...'21'01 p,( fo) = 2; e; with C1 a 1. Uniform Gaussian Laplacian Bits r; d; rj d, rj dj ??oo :I: :I: ??oo ??oo :I: :r: ??oo QO ??oo ??oo ??oo ??oo QO :I: ??oo :I: :I: ??oo ??oo :I:) :I:) ??oo ??oo ??oo ??oo ??oo 00 00

11 ; Q.4528 l I , f1curc 10.5 EJample of a Uoyd.Mu quantizer. The Dumber of ru:oastrvc:tioo levels is 4. and the probability dciisityfunc:tioofor f is GalllliaD with mcaa of 0 and variance of 1. o 10 Q 2 J J, UoydMIx 30 I quii'i\iz8cion L flprc 10.6 eomparisod of IYCfIIC distonion D E1.<I.m IS a tune. tion of L, the awnbcrof~ levels, for a Wliform quidwr (doctcd tide) ajid the Uoyd.Mu quidliur (loiid line). The wrtica111is ill0... D. The probability dcmity (uaetioiiif 0( 1. to be Oa18iaD widl ftriadci

12 Scalar Case (cont.) For some densities, the optimal solution can be viewed as follows: f g Uniform,.. Nonlinearity. r Nonline.rity' quantizer Figure 10.7 Nonuniformquantization by companding,,..,.. f p (g) is flat density equal1y likely between gmax' gmin o.clliry OcnIDry Fo...ard TransCormauon lnyei'1c Transformallo" p(fj. C:ft'1)8',? &Xi'{} '.ierlb{;j '.';:6 crill:;) R.ayleilh '(fj. {:S},.iup{f} /..j':4'ln[l/ci ijd"1 II,(n.2{1!1),. Htap {.I}] /ao I I InC' I 2" CJ II 120 J:.. ",.,(Iup 101}] /<0 /.!. II In 11 2; ] ;<0

13 Veetor Case fl' f 2: two image intensities One approach: Separate level assignment for fl' f 2. previous discussions apply Another approach: Joint level assignment for (f l' f 2). typically more efficient than separate level assignment ~ rs, Minimize reconstruction levels/decisionboundaries: = f f «(f 1 j 1)2 + (f 2 j 2)2).P (f l' f 2) 'df 1. df 2 11/2. efficiency depends on the amount of conelationbetween f } and f 2. finding joint density is difficult (Ex): Extreme casevector level assignment for 256x 256 Joint density P(x}. x2'.... x256)is difficult to get Law of diminishing returns comes into play

14 Codeword Design: Bit Allocation. After quantization, need to assign binray codword to each quantization symbol.. Options: uniform: equal number of bits to each symbol + inefficient ~ nonuniform: short codewords to more probable symbols, and longer codewords for less probable ones.. For nonuniform, code has to be uniquely decodable: Example: L. ==.4, rl + 0, r2 + 1, rg + 10, r suppose we receive 100. Can decode it two ways: either rgrl or r2rlrl. Not uniquely decodable. One codeword is a prefix of another one.. Use Prefix codes instead to achieve unique decodability. l 1

15 Overview (. Goal: Design variable length codewords lo that the average bit rate is minimized.. Define entropy to be:. L H = E Pilog2Pi i=1 Pi is the probabiliyt that the ith symbol is ai. Since 2:f IPi == 1, we have 0 < H < log2l. Entropy: average amount of information a message contains. Example: L == 2, PI == 1, P2 == 0 + H == 0 A symbol contains NO information. PI == P2 == 1/2 + H == 1 Maximum possible value. ~ () p 2

16 Overview. Information theory: H is the theoretically minimum possible average bite rate.. In practice, hard. to achieve Example: L symbols, Pi == I + H == log2~ uniform length coding can achieve this.. Huffman coding tells you how to do nonuniform bit allocation to different codewords so that you get unique decodability and get pretty close to entropy. 3

Objectives of Image Coding

Objectives of Image Coding Representation of an image with acceptable quality, using as small a number of bits as possible Applications: Reduction of channel bandwidth for image transmission Reduction