Вы находитесь на странице: 1из 6

Normalized convolution for Image Interpolation

Pi19404
February 4, 2014

Contents

Contents
Normalized convolution for Image Interpolation
0.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3
3 6

2 | 6

Normalized convolution for Image Interpolation

Normalized convolution for Image Interpolation


0.1 Introduction
In this article we will look at the concept for normalized convolution for image interpolation and how a sparsely and ununiformaly sampled grid can be used for image reconstruction.

 

Convolution is a neighborhood operation. For 1D case,the convolution of a signal f (t) with a filter h(t) can be expressed as

y (n) =

Xh k f n k
( ) (
k

(1)

      

Thus result of convolution at point n is weighted sum of samples in the neighborhood of sample point n. Interpolation is process of estimation of value of signal at unknown point based on set of known points. Many times we require analysis of irregularily sampled data, which is more compilated than regularily sampled data.It is often required to reconstruct the irregularily sampled signal or resample it onto a regular grid. One method to do this is to use interpolation techniques obtain a regularily sampled signal.The missing values in the regularily sampled grid are computed using intepolation which is implemented using convolution operations. Let us consider a signal 1D signal x(t) = [x1 ; 0; 0; x4 ; x5 ; 0; 0] and h(t) = [1=3; 1=3; 1=3]. The result of convolution is given by y (t)

x5 =3; x4 + x5 =3; x5 =3; x1 =3]

= [x1 =3; x1 =3; x4 =3; x4 +

convolution can be expressed as weighted average about a local neighborhood.

3 | 6

Normalized convolution for Image Interpolation

 

In general Let x1 ; : : : ; xn be set of values and w1 ; : : : ; wn be the associated weights The convolution is given by

xavg

P w i f xi Pwi
i

( ) (
i

( )

(2)

The convolution can be made more effective by a normalized operation which takes into account the missing samples.One of applications of normalized convolution which is a method of interpolating irregularily sampled data points.The conceptual basis for the method is the signal/certainty philosophy separating the values of a signal from the certainty of the measurements. Normalized Convolution can, for each neighborhood of the signal, geometrically be interpreted as a projection into a subspace which is spanned by some analysis functions The idea of normalized convolution is to associate each signal with a certainty component which expresses the level of confidence in the reliability of each measure of the signal. Certainty associated with missing samples is 0 ,while that of known samples is 1. Thus we can express a map c(t) associated with signal f (t) which has the same dimensions as the signal f (t).The certanity map associated with a signal is simple the locations at which samples are to be found. In case of the above examples the certainty map is given by

        

c(t) = [1; 0; 0; 1; 1; 0; 0]

Having both signal and associated certainty map leads to main concept behind normalized convolution

yc (t) = [1=3; 1=3; 1=3; 2=3; 2=3; 1=3; 1=3]

Let us consider the convolution of certainty map by filter h(t)

In normalized convolution we have two distinct weights ,a certainty c(t) which is associate with the signal and applicability a(t) which is associatedd with neighborhood. the applicability function determines neighborhood of convolution as well as weights associated with neighborhood pixel.

4 | 6

Normalized convolution for Image Interpolation

For 1D signals it can be expressed as

xavg

sumk f (x k ) c(x k ) a(k ) k c(x k )a(k ) (fc) x xavg = ca

(3) (4)

   

in case of above examples we get

y (t) = [x1 ; x1 ; x4 ; (x4 + x5 )=2; (x4 + x5 )=2; x5 ; x1 ]


This is an approximation of original signal,where components x1 ; x5 have been retained while the other components are interpolated values in the neighborhood. The division of certainty matrix ensures than the signal values remain within a valid range and is primary reason for better performance. when signal value increases due to applicability function ,the certainty values also increases by the same ammount and this will compensate for the increase in the signal values and provide a normalized measure.This behavior is also observed when the signal value is reduced. Let us consider a gaussian applicability function.Only parameter that neeeds to be controlled is the standard deviation  of the gaussian function and the aperture/neighborhood size of the gaussian. Due to normalization property we get a much better result than a standard convolution would yeild The matlab code for the 2D normalized convolution is given below.

  
1 2 3 4 5 6 7 8 9 10

im=double(imread('lena.png')); figure(1);colormap(gray);imagesc(im); cert = double(rand(size(im)) > 0.8); imcert = im.*cert; figure(2);colormap(gray);imagesc(imcert); %applicability function 3x3 gaussian with variance of 1 x = ones(7,1)*(-3:3) y = x'; a = exp(-(x.^2+y.^2)/4); figure(3);mesh(a);

5 | 6

Normalized convolution for Image Interpolation

(a) orignal

(b) sampled

(c) filter

(d) gaussian interpola- (e) normalized convotion lution


11 12 13 14 15 16 17 18 19 20

imlp = conv2(imcert, a, 'same'); figure(4);colormap(gray);imagesc(imlp); %convolution with certainty map G=conv2(cert,a,'same'); %normalized convolution c = imlp./G; figure(5);colormap(gray);imagesc(c);

6 | 6

Вам также может понравиться