You are on page 1of 12

Improved minimal inter-quantile distance method for blind estimation of noise variance in images

Vladimir V. Lukina, Sergey K. Abramova, Alexander A. Zelenskya, Jaakko T. Astolab, Benoit Vozelc, Kacem Chehdic a National Aerospace University, 61070, Kharkov, Ukraine; b Tampere University of Technology, Institute of Signal Processing, P.O. Box-553, FIN-33101, Tampere, Finland c University of Rennes 1 - TSI2M, 22 305 Lannion Cedex, BP 80518, France
ABSTRACT
Multichannel (multi and hyperspectral, dual and multipolarization, multitemporal) remote sensing (RS) is widely used in different applications. Noise is one of the basic factors that deteriorates RS data quality and prevents retrieval of useful information. Because of this, image pre-filtering is a typical stage of multichannel RS data pre-processing. Most efficient modern filters and other image processing techniques employ a priori information on noise type and its statistical characteristics like variance. Thus, there is an obvious need in automatic (blind) techniques for determination of noise type and its characteristics. Although several such techniques have been already developed, not all of them are able to perform appropriately in cases when considered images contain a large percentage of texture regions and other locally active areas. Recently we have designed a method of blind determination of noise variance based on minimal inter-quantile distance. However, it occurred that its accuracy could be further improved. In this paper we describe and analyze several ways to do this. One opportunity deals with better approximation of inter-quantile distance curve. Another opportunity concerns the use of image pre-segmentation before forming an initial set of local estimates of noise variance. Both ways are studied for model data and test images. Numerical simulation results confirm improvement of estimate accuracy for the proposed approach. Keywords: inter-quantile distance, noise variance evaluation.

1. INTRODUCTION
An advantage of modern RS systems is that they are able to provide potential users by information valuable for such important applications as meteorology, environment monitoring, pollution detection, agriculture, etc.1,2,3 To ensure wider capabilities of remote sensing, modern RS systems are commonly provided by multichannel (multispectral, hyperspectral) operation facilities, and a general tendency is to increase the channel number. The examples of such systems are AVIRIS4, CHRIS-Proba5, etc. Meanwhile, increasing the number of channels (sub-bands) results in considerably increased complexity of different stages and procedures of RS data processing: filtering, edge and object detection, and compression6-9. Simultaneously with demand to process data as quickly as possible, this explains a need in design of blind methods for different stages of multichannel RS data processing. In this paper, we consider a particular task of blind noise variance evaluation in images. Note that for multichannel RS data this operation is to be carried out component-wise since statistical characteristics of noise in different component (subband) images can vary a lot10,11. Besides, it is worth noting that in many practical situations noise is a dominating
a c

Correspondence to Lukin V.V.: e-mail lukin@xai.kharkov.ua tel./fax +38 0573 151186 Work supported by the European Union. Co-financed by the ERDF and the Regional Council of Brittany, through the European Interreg3b PIMHAI project.
Image and Signal Processing for Remote Sensing XIII, edited by Lorenzo Bruzzone, Proc. of SPIE Vol. 6748, 67481I, (2007) 0277-786X/07/$18 doi: 10.1117/12.738006

Proc. of SPIE Vol. 6748 67481I-1 Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2013 Terms of Use: http://spiedl.org/terms

factor degrading image quality. Finally, one should keep in mind that a priori knowledge of noise type and statistical characteristics is required for the most effective algorithms of image denoising12-14, edge detection15,16, compression of noisy images, etc.11,13,17 For all these applications it is desirable to provide high enough accuracy of noise variance evaluation18, otherwise the performance of image processing might inappropriately worsen. Assume that noise type has been already pre-determined. This can be done, e.g., by means of the already designed methods6,19-21. Then, the basic task is to evaluate the main parameters of the determined dominating factor, desirably in automatic mode. Quite many blind techniques for evaluation of additive or multiplicative noise variance have been already proposed22-32. A practical choice depends upon several factors. The most important factor is a priority of requirements to the technique. So, let us remind the basic requirements. First, a technique should be applicable to images with different structure, i.e., irrespectively, what is a percentage of pixels belonging to image homogeneous regions (IHR). Many existing techniques perform well enough if an image is not too textural. But they fail if the percentage of pixels that belong to IHR is less than 3040% 22,24,26,28. Because of this, considerable efforts have been spent by us for designing the blind methods able to operate well for highly textural images25,27,29,30. Second, a blind technique has to be able to perform properly (with providing appropriate accuracy) in cases of both spatially uncorrelated and correlated noise, desirably even if spatial correlation properties of noise are a priori unknown. While the methods27,29 are more resilient to absence of information on noise spatial correlation properties, the method 25 produces biased estimates of noise variance in case of spatially correlated noise. The third requirement is to provide high accuracy of estimation for a wide range of possible values of noise variance. As a marginal case, a method should produce near-zero estimates for noise-free images. In this sense, the methods27,29,31,32 are characterized by the best accuracy. Moreover, the methods31,32 are able to evaluate fluctuative (additive or multiplicative) noise variance in cases of simultaneous presence of impulse noise. Finally, the fourth requirement is to perform quickly enough. In this sense, the methods31,32 and the corresponding algorithms are slow and require intensive computations. The method27 is faster but the technique29 is one of the most efficient. However, the performance of the latter technique can be further improved. This paper deals with considering modifications that can be done for the technique29 in order to improve its accuracy. The paper is organized as follows. Some common stages of blind evaluation of noise variance are described in Section 2. A model for local variance estimate distribution is also given. Section 3 describes the proposed approximation of inter-quantile curve and its performance is studied for model data. Then, in Section 4, we test this modification for a set of test images artificially corrupted by noise. Performance comparison for several methods is provided. Section 5 contains description of another modification that deals with exploiting segmentation maps. Then, the conclusions follow.

2. COMMON STEPS IN BLIND NOISE VARIANCE EVALUATION AND LOCAL ESTIMATES DISTRIBUTION
In general, one can mention two main approaches to blind evaluation of noise variance. As said earlier, estimation in spectral domain 12,25 suffers from problems that arise in case of spatially correlated noise. In turn, methods operating in spatial domain better cope with this frequently met phenomenon. Thus, let us concentrate on the latter approach. According to it, an image under interest at the first step is divided into N b overlapping or non-overlapping blocks of a rather small size that tessellate this image. Typically a block size is 5x5 or 7x7 pixels, the latter is preferable for spatially correlated noise. At the second step, the local mean I l and the local variance estimates l2 , l = 1,..., N b are calculated for each l-th block. If the considered dominant noise is multiplicative, the only difference is that one has to calculate 2 local relative variance estimate as l = l2 / I l2 . Without loosing generality, let us further suppose that noise is pure

additive and we deal with a set of estimates l2 , l = 1,..., N b . Then, at the third stage, the obtained set of such estimates is to be processed in some way. At this stage there is a variety of possible variants. However, almost all of the designed methods are based on the following fact. The local estimates calculated for blocks that belong to IHRs (normal estimates) do not differ too much from the true value of noise variance 2 tr and they form a mode of local estimates distribution. Other local estimates obtained in image heterogeneous (locally active) regions like edge and detail neighborhoods, textural regions, etc, are usually considerably larger than the true value and such estimates produce a heavy one-sided tail of local estimates distribution. The shape and parameters

Proc. of SPIE Vol. 6748 67481I-2 Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2013 Terms of Use: http://spiedl.org/terms

of such tail depend upon image properties, noise variance, block size, etc. Examples of histograms of such distributions have been provided in our previous papers7,23,29. Moreover, it has been shown that normal estimates have practically Gaussian distribution if block size is large enough (e.g., 7x7)23,30. Thus, as in our earlier papers, let us use the following model for probability density function (PDF) of local estimates

( x ) = p Gaussian ( m, 2 ) + (1 p ) Uniform ( 0, M ) ,

(1)

2 2 where Gaussian ( m, 2 ) denotes the Gaussian PDF with the mean m ( m = tr or m = for additive and multiplicative 2 2 noise, respectively) and variance 2 ; do not mix 2 with tr or ), Uniform ( 0, M ) is the uniform PDF within the

limits from 0 to M, p is the parameter characterizing the percentage of normal local estimates. The variance 2 in (1) is determined by the used block size, noise PDF and variance, and spatial correlation properties23,30. Other parameters of the model (1) are the following: M describes the range of possible variation of local estimates obtained in image heterogeneous regions (obviously, a local estimate is non-negative). For modeling different practical situations, it is possible to vary the ratio M / m. For given block size and noise PDF, 2 is approximately proportional to m 2 . The PDF Uniform ( 0, M ) relates to variance estimates obtained in heterogeneous blocks. We used uniform PDF although other appropriate model PDFs can be used without considerable influence on accuracy of noise variance estimates. The aforementioned peculiarities of local estimates distribution have been put into basis of an idea that our task is to get some accurate and robust estimate of this distribution mode and accept it as a final estimate of noise variance. The recent reseach has been concentrated on analysis and design of such estimation techniques including the use of the sample myriad23, bootstrap based approach27 and inter-quantile distance minimization29. This last modification has resulted in obtaining quite accurate estimates even if the percentage p of local estimates calculated in homogeneous image blocks is small, up to p=0.1.

3. THE DESIGNED MODIFIED INTER-QUANTILE METHOD


Let us first briefly consider the approaches27,29 based on quantile and inter-quantile estimates. Assume that as an estimate of noise variance we can use some quantile of the set of estimates l2 , l = 1,..., N b . The best (optimal) quantile index nopt is such for which its bias and variance are minimal. However, it is a priori unknown and, in fact, it depends upon N b and p 27,29. To prove the existence of such quantile, consider several sets of parameters (Cases) of the model (1) presented in Table 1. These sets mimic different possible relationships between the parameters of the model (1). In particular, the Cases 5 and 6 are the most unfavourable for majority of techniques of blind noise variance estimation since for them p is rather small (this corresponds to highly textured images). Table 1. The considered sets of parameters of the model (1) Model p m 2 Parameters M /m Case 1 0.9 100 800 100 Case 2 0.8 300 7200 100 Case 3 0.65 100 800 100 Case 4 0.5 200 3200 100 Case 5 0.15 300 7200 100 Case 6 0.1 400 12800 100 Case 7 0.65 400 12800 10
2000 1800 1600 1400 1200 1000 800 600 400 200 0 0

Case 1 Case 2 Case 3 Case 4 Case 5 Case 6 Case 7

10

20

30

40

50

n%

60

Fig. 1. The aggregate errors for the quantile estimate as the functions of quantile index n% for different Cases

Proc. of SPIE Vol. 6748 67481I-3 Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2013 Terms of Use: http://spiedl.org/terms

Let us characterize the accuracy of estimation techniques by the following three quantitative parameters: the bias = m m , the variance 2 =

(m

and the aggregate error = 2 + 2 . Here m denotes the used estimate of

noise variance; is the expectation for an ensemble of realizations. We have also calculated the parameter rel as

rel = ( / m) 100% . Simulations have been carried out for 10000 realizations.
Detailed analysis of behavior of the bias, variance and aggregate error on n% = 100n / N b (n denotes quantile index) is

presented in the paper27. Therefore, here we represent only the dependences of the aggregate error on n% (Fig. 1). They all have global minima for nopt % /100 p / 2 , and the estimates of noise variance obtained for the corresponding quantile nopt = N b nopt % /100 are practically unbiased. The values nopt % or, respectively, nopt are to be estimated. An efficient way to do this is to apply minimal inter-quantile distance (range) IQR approach29. Assume that by sorting an original set of local variance estimates l2 , l = 1,..., N b in ascending order we have obtained a sample X (t ) , t = 1,..., N b where X (t ) denotes the t-th order statistic. Then it can be expected that for PDF (1) the difference X (t + s ) X (t ) for fixed s (where s is an even integer) is the smallest in the neighbourhood of distribution mode. This assumption occurred to be true 29, and Fig. 2,a illustrates this for two different values of s: s / N b = 0.1; s / N b = 0.2 for Case 2. As seen, for both values of s / N b the plots have global minima for 100(t + s / 2) / N b 40 , i.e. for (t + s / 2) / N b 0.4 p / 2 . And according to the plot of in Fig. 2 the quantile that has the smallest for Case 2 is just n% = 40.

Then, by finding such test ( t = 1,..., N b s ) for which X (t + s ) X (t ) is the smallest one can get an estimate of distribution mode and, respectively, noise variance. However, two questions arise: 1) how to produce a distribution mode estimate X (i.e., m in the model (1)) knowing the interval X (test ) ; X (test + s ) , and 2) what is the optimal or, at least, reasonable choice of s? In the paper29, we have considered the simplest way to obtain X : X = X ( test + s / 2) where s is even.
100

X
80 60 40 20 0 0 10

(t+s)

-X

(t)

s/Nb=0.1 s/Nb=0.2

100(t+s/2)/N b
20 30 40 50 60 70 80 90 100

Fig. 2. The plots of (t + s / 2) = X (t + s ) X (t ) vs 100(t + s / 2) / N b (Case 2) The way to select s was two-stage. More in detail, the estimation procedure29 (algorithm AIQR) was the following: init 1) calculate X (t + s ) X (t ) ( t = 1,..., N b s ) for s = [0.1N b ] 2 and find such test for which X (t + s ) X (t ) is the smallest;
init 2) calculate s qopt = 1.444t est + 0.067 N b

, if s qopt > N b 1, s qopt = [N b 1]2 where []2 denotes rounding off to a nearest

even integer;

Proc. of SPIE Vol. 6748 67481I-4 Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2013 Terms of Use: http://spiedl.org/terms

3) calculate X (t + s
fin

qopt

fin X (t ) ( t = 1,..., N b s qopt ), find test for which X (t + s

qopt

X (t ) is the smallest and obtain the final

estimate as X (test + s

qopt

/ 2)

This algorithm has exploited the fact that optimal s depends upon parameters of the model (1). This is clearly seen from analysis of plots presented in Fig. 3. As seen, optimal ratios for which the aggregate errors are the smallest take place for s opt / N b p .
300 250 200 150 100 50 0 0 10 20 30 40 50 60 70 80 90 s % 100

IQR

Case 1 Case 2 Case 3 Case 4 Case 5 Case 6 Case 7

Fig. 3. The plots of IQR as the functions of 100 s / N b for the considered Cases. One of the main disadvantage of AIQR estimator is a rather large estimation variance 29. Our research has shown that the init reason of such behavior of the AIQR is rather large variance of the estimate test obtained for s = [0.1N b ] 2 at the first
init step of the algorithm. Note that test is directly involved in calculation of s qopt at the second step of the algorithm and

the calculated s qopt is used at the third step. In other words, errors at the first step propagate to the next ones. Then, init one can expect that if the estimates test will be more accurate, the final estimates of noise variance will be more accurate as well. One observation that follows from examples in Fig. 2 is that the curves (t + s / 2) are noisy. To suppress this noise (to reduce the curve fluctuations and, hence, to decrease the variances of initial and the final estimates) one should use some preprocessing procedure. We propose to use a square (second order polynomial) LMS regression in the neighborinit hood of the curve minimum (rough quantile number estimation test + s / 2 ). Then the coordinate of approximation curve
flt minimum test is used to obtain corrected estimate X (test ) .
flt

100 80 60 40 20 0 0 10

X (t+s) -X (t)

s/Nb=0.1 s/Nb=0.2

100(t+s/2)/N b
20 30 40 50 60 70 80 90 100

Fig. 4. The plots of (t + s / 2) = X (t + s ) X (t ) vs 100(t + s / 2) / N b and their approximations (Case 2) The examples of regression curves for s / N b = 0.1 and s / N b = 0.2 are shown in Fig.4 (thick solid lines). One particular question is how to determine a range of the values t in which approximation is to be done? Our investigations have

Proc. of SPIE Vol. 6748 67481I-5 Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2013 Terms of Use: http://spiedl.org/terms

init init shown that optimal range for LMS approximation can be determined as 0.1(test + s / 2);1.9(test + s / 2) . Such choice for majority of Cases provides the best accuracy of resulting estimates of inter-quantile curve global minimum. The method based on such approach of inter-quantile range filtering is further referred as IQRF.

Consider now the performance of the designed algorithm. First, let us compare the accuracy of initial estimates for IQR29 and IQRF. The data for different values of p are given in Table 2. We have controlled such additional statistical parameters: the mean n % of n% = 100(test + s / 2) / N b and its variance n2 . Besides, we present the results for optimal
%

values of parameter s = 100s / N b where s is the optimal value of s that produces the smallest aggregate error. Below for all considered statistical parameters we use subscripts to denote their relation to the corresponding method.
opt % opt opt

As seen, the obtained values of the optimal quantile

nIQR % are characterized by better accuracy smaller variances

n2

IQR %

and

2 IQR

rel . Bias values for both estimators are close to zero. As the result, the values IQR for the new estimator

IQRF are smaller than for the IQR estimator. Table 2. Accuracy of the IQR and novel IQRF estimators for the different p ( m =10; 2 =8; M m =100)

p
0.90 0.75 0.65 0.50 0.25 0.15 0.10

IQR estimator

IQRF estimator
2 IQR

opt %

nIQR %
45.02 37.73 32.80 25.44 13.23 8.32 5.87

2 n IQR %

IQR
0.013 0.019 0.022 0.023 0.022 0.023 0.023

IQR
0.015 0.019 0.023 0.031 0.063 0.115 0.186

rel IQR

opt %

nIQRF %
45.00 37.82 32.91 25.58 13.24 8.29 5.82

2 n IQRF %

IQRF
-0.0069 -0.0069 -0.0068 -0.0010 -0.0089 -0.0085 -0.0204

2 IQRF

rel IQRF IQRF

88.0 73.4 63.7 48.2 23.7 13.7 9.1

0.098 0.253 0.338 0.420 0.436 0.373 0.326

0.015 0.019 0.022 0.030 0.063 0.114 0.185

1.23 1.38 1.52 1.76 2.51 3.39 4.31

89.5 74.5 64.5 49.5 24.6 14.6 9.2

0.178 0.206 0.287 0.333 0.354 0.302 0.244

0.014 0.018 0.021 0.027 0.058 0.098 0.157

0.014 0.018 0.021 0.027 0.056 0.098 0.158

1.18 1.34 1.45 1.64 2.37 3.13 3.97

Based on IQRF estimation, an automatic inter-quantile range filtered (AIQRF) estimator exploiting the same idea as AIQR estimator has been proposed29. As well as the AIQR estimator, the AIQRF consists of two stages obtaining the initial estimate and the final one. More in details, it is the following: 1) apply to the sample at hand the IQRF estimator with parameter s = [0.1N b ] 2 and obtain the initial estimation of
I 2) calculate quasi-optimal inter-quntile range s qopt = 1.835test + 0.079 N b , if s qopt > N b 1, s qopt = [ N b 1]2 ; 2 flt 3) apply to the sample the IQRF estimator with s = s qopt , find test II and obtain the final estimate as X
flt ( test II )

I flt minimal inter-quantile range left boundary test = test I s 2 ;

Table 3. Accuracy of the AIQR, IQRF and AIQRF estimators for the different Cases
Case
1 2 3 4 5 6 7

AIQR estimator

IQRF estimator (with optimal s)


rel AIQR

AIQRF estimator

AIQR
-0.09 -0.39 -0.17 -0.43 -3.59 0.50 -0.56

2 AIQR

AIQR
6.28 56.85 7.56 35.95 198.82 294.90 105.36

IQRF
0.04 -0.40 -0.17 -0.30 -0.13 -0.85 -0.31

2 IQRF

IQRF
1.44 14.95 2.17 11.48 90.56 259.86 32.98

rel IQRF

AIQRF
-0.13 -0.50 -0.18 -0.39 -1.60 -0.97 -0.48

2 AIQRF

AIQRF
1.61 16.80 2.28 11.95 97.75 254.36 34.59

rel AIQRF

6.27 56.70 7.53 35.76 185.97 294.65 105.05

2.51 2.51 2.75 3.00 4.70 4.29 2.57

1.44 14.78 2.14 11.39 90.55 259.13 32.88

1.20 1.29 1.47 1.69 3.17 4.03 1.44

1.59 16.55 2.24 11.81 95.18 253.41 34.36

1.27 1.37 1.51 1.73 3.30 3.99 1.47

Simulation results for the old estimator AIQR29 and the designed estimators for different Cases are presented in Table 3.

Proc. of SPIE Vol. 6748 67481I-6 Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2013 Terms of Use: http://spiedl.org/terms

As seen, the proposed modified estimator AIQRF outperforms previously designed algorithm AIQR and provides comrel parable results with IQRF in case of optimal selection of s. The relative aggregate error BM for AIQRF is considerably (by approximately two times) reduced in comparison to AIQR.

4. ANALYSIS OF THE PROPOSED METHOD PERFORMANCE FOR TEST IMAGES


Consider now the accuracy of the proposed method for a set of conventional test images (Peppers, Barbara, Baboon, 2 Goldhill). Additive noise (Gaussian, i.i.d.) with zero mean and different variances tr (50, 100, 400) has been added to these images. Estimates of noise variance have been obtained for a large number of noise realizations and then statistical characteristics of estimates have been calculated. Also, we have processed noise-free images for which an estimate is characterized by the only parameter its bias . The obtained results are given in Table 4. Table 4. Accuracy of the AIQR and proposed AIQRF techniques for different test images and variances of additive noise AIQR estimator AIQRF estimator
2 tr
2 nAIQR % n AIQR % AIQR

Image

2 AIQR

AIQR
165.69 175.14 74.83 49.81 63.74 211.31 1337 1888 5845 452.79 679.43 1272

rel AIQR

2 nAIQRF % n AIQRF % AIQRF

2 AIQRF

rel AIQRF AIQRF

0 50 100 400 0 50 100 400 0 50 100 400 0 50 100 400

16.99 28.87 32.19 37.46 7.72 19.13 21.80 28.63 6.87 11.27 13.95 22.31 5.35 21.92 26.87 36.49

2.52 1.79 2.41 0.74 1.23 1.96 0.87 0.79 1.71 2.46 1.96 2.03

9.79 12.80 13.10 6.21 4.24 6.99 7.68 12.42 28.33 36.44 43.25 75.63 4.54 21.15 25.93 35.04

1.98 3.50 36.33 0.94 4.78 57.11 9.90 16.73 125.84 5.43 7.09 44.19

25.74 13.23 2.16 14.12 7.98 3.63 73.13 43.45 19.11 42.56 26.07 8.92

18.90 31.53 34.92 40.30 8.95 20.89 23.55 31.55 8.66 12.78 15.63 25.16 5.42 24.76 29.95 39.65

0.26 0.12 0.13 0.08 0.10 0.21 0.21 0.26 0.40 0.48 0.34 0.20

10.44 14.87 16.49 16.62 4.56 8.83 10.67 27.11 32.36 41.27 49.92 98.55 4.61 25.01 31.47 48.31

0.36 0.61 5.43 0.26 0.96 11.64 3.80 6.70 36.47 1.31 1.97 7.27

221.4 272.4 281.5 78.30 114.8 746.5 1707 2499 9748 626.7 992.1 2341

29.76 16.51 4.20 17.70 10.72 6.83 82.63 49.99 24.68 50.07 31.50 12.10

GoldHill

Baboon

Barbara

Peppers

Let us analyze the obtained data. As seen, for noise free images both considered methods produce noise variance estimates that are not equal to zero (bias values AIQR and AIQRF are positive). There are several reasons behind this fact. First, real life images practically always contain some noise (even those ones commonly used in experiments) and/or they do not have ideally homogeneous regions. Second, local estimate distribution for noise-free images does not obey the model (1) on which we have relied in design of our methods for blind evaluation of noise variance. One interesting observation is that both AIQR and AIQRF are larger for images that contain more textured regions. Since nAIQR % and
nAIQRF % can serve as some estimates of the percentage p of IHRs, it follows that the images Baboon and GoldHill are the

most textural. This is in good agreement with visual inspection of the considered test images. For noisy images we have got, in some sense, quite surprising results. Recall that for simulation data for the model (1) rel BM for AIQRF was considerably smaller than for AIQR (Table 3 in Section 3). However, for the data in Table 4 the situation is the opposite for practically all test images and additive noise variance values. Analysis shows that this is due 2 to the fact that AIQR is smaller than AIQRF . Note that at the same time 2 AIQR is considerably larger than AIQRF . In other words, due to using polynomial approximation we have reduced estimate variance but increased its bias.

Proc. of SPIE Vol. 6748 67481I-7 Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2013 Terms of Use: http://spiedl.org/terms

To understand why this has happened, we have, first, returned to the model (1) again. Its analysis shows that approximation of normal local estimate distribution by Gaussian Gaussian ( m, 2 ) is not absolutely adequate. In fact, this distribution is slightly asymmetric23 and its approximation by the second-order polynomial produces slightly biased estimates of distribution mode. Another reason could be that another term in the model (1), Uniform ( 0, M ) , can not be well enough approximated by uniform distribution for practical distributions of local estimates for real life images. Thus, we decided to somehow additionally reduce this negative influence of abnormal estimates obtained for blocks placed in image heterogeneous regions.

5. METHOD BASED ON USING PRE-SEGMENTATION MAPS


Suppose that we have managed somehow to carry out discrimination between image homogeneous and heterogeneous 2 regions and, after this, have obtained a set of local variance estimates lh , l = 1,..., N bh only for blocks that belong to the determined IHRs. Obviously, in this case a number of such blocks N bh occurs smaller than the maximally possible number of blocks that can tessellate an image. Then, a question is how to perform discrimination between image homogeneous and heterogeneous regions without knowing noise type and variance. There exist many image segmentation methods able to indicate and localize quite large image homogeneous regions characterized by compactness of pixels that are referred to the corresponding areas 33. Many methods of image segmentation require a priori information on noise type and variance34,35 and they can not be directly applied in our case since we do not know noise variance a priori. Fortunately, there are image segmentation techniques that do not require knowing noise type and statistical characteristics a priori36. Among them, we retain a recent unsupervised one, based on a variational classification method of observed pixels, following a preliminary optimized histogram transformation by gravitational clustering. In the first stage, the original histogram is greatly reduced to highlight only pertinent modes of the observed image. This is achieved by progressively decreasing the dispersion of the initial modes with regard to their relative centres of gravity. Then, the best thresholds and the best modes are obtained by alternate optimization of some energy of multi-thresholding. This energy measures the quality of a map of homogeneous regions as a function of the intra-area variance of the detected regions. An area is decided to be uniform (homogeneous) if the dispersion of its grey levels is sufficiently low. In the second stage, a supervised variational classification method to which it is sufficient to give the previously obtained set of representative modes of the classes and which takes into account an a priori homogeneity constraint is applied to obtain the final result. Let us give an example of segmentation method operation. Fig. 5,a presents the original (noise-free) test image Baboon. Its noisy version is given in Fig. 5,b (additive noise variance is equal to 100). Image segmentation result obtained by the method36 is represented in Fig. 5,c. Pixels that belong to the same region (segmentation class) are shown by the same intensity in gray scale representation. For getting discrimination maps, we have applied the following post-processing of segmented image. Within 5x5 scanning window, it has been tested to how many segmentation classes the pixels belong. It has been done using the following rule for all scanning window positions: a) if all scanning window pixels belong to the same segmentation class, then this scanning window position corresponds to homogeneous region (and the corresponding pixel of discrimination map DM ij = 128 ); b) if scanning window pixels belong to two different segmentation classes, then this scanning window position corresponds to edge/detail neighborhood (and the corresponding pixel of discrimination map DM ij = 0 ); c) if scanning window pixels belong to three or more segmentation classes, then it is supposed that this scanning window position corresponds to texture (and the corresponding pixel of discrimination map DM ij = 255 ). The obtained discrimination map is presented in Fig. 5,d. Pixels that belong to detected homogeneous regions are shown by gray color. As seen, the proposed two-stage procedure of image processing (segmentation+discrimination) determines homogeneous and quasi-homogeneous areas quite well. And the percentage of pixels that belong to these areas for the test image Baboon is rather small. Note that image segmentation and, thus, discriminations results depend upon noise level.

Proc. of SPIE Vol. 6748 67481I-8 Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2013 Terms of Use: http://spiedl.org/terms

The obtained discrimination map DM ij can be further used for blind evaluation of noise variance. We propose to 2 calculate local estimates lh , l = 1,..., N bh only for those scanning window (block) positions for which DM ij = 128 . As seen, segmentation method localizes image homogeneous (or, at least, quasi-homogeneous) regions shown by gray color well enough.

-.

fr -

c d Fig. 5. The results of image region discrimination: a) the original (noise-free) image; b) the noisy image; c) the segmentation result; d) the obtained discrimination map Then, the method based on image pre-segmentation and homogeneous region detection that we have called AIQRFHRD contains the following stages:

Proc. of SPIE Vol. 6748 67481I-9 Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2013 Terms of Use: http://spiedl.org/terms

E4: ?T

1) perform image pre-segmentation, find blocks that fully belong to the pre-determined IHRs, and obtain for them 2 a set of local variance estimates lh , l = 1,..., N bh ; 2 2) make up sorting of the estimates lh , l = 1,..., N bh in ascending order with obtaining X ( t ) , t = 1,..., N bh ;
init 3) calculate X (t + s ) X (t ) ( t = 1,..., N bh s ) for s = [0.1N b ] 2 and find such test for which X (t + s ) X (t ) is the

smallest; 4) approximate X (t + s ) X (t ) curve (as the function of index t) in the neighborhood flt init init init 0.1(test + s / 2);1.9(test + s / 2) of test + s 2 and find its minimum test I analytically as described in Section 3;
I 5) calculate quasi-optimal inter-quantile range s qopt = 1.835test + 0.079 N bh , if s qopt > N bh 1, s qopt = [ N bh 1]2 ; 2

I flt use this minimum to calculate the first stage quantile index test = test I s 2 ;

6) repeat one time the steps 3 and 4 for s = s qopt and determine the second stage inter-quantile curve minimum
flt II flt test II and quantile index test = test II s qopt 2 ;

7) obtain the final estimate as X

II ( test + s qopt / 2)

Consider now the performance of this method for the test images. The obtained simulation data are presented in Table 5 and they can be easily compared to results given for other techniques in Table 4. This comparison shows the following. First, for noise-free images, estimates have become more accurate (the values AIQRFH are smaller than AIQR and

AIQRF for the corresponding images). Besides, the values nAIQRFH % are larger than nAIQRF % for all considered images
2 and noise variance values. This indirectly shows that the majority of obtained estimates lh , l = 1,..., N bh can be considered normal (histogram analysis of these estimates is presented in our paper 30 and it shows that this conclusion is true). In case of noisy images, estimates of noise variance have become more accurate than for the technique AIQRF. Bias absolute values for the AIQRF-HRD have considerably decreased in comparison to the method AIQRF. There is no obvious tendency to be observed in comparing variance values for both techniques. However, comparing the relative aggregate errors for the corresponding situations (the same image and the same noise variance), we can state that accuracy of AIQRF-HRD is sufficiently better. Table 5. Accuracy of the proposed AIQRF technique with homogenous region detection (AIQRF-HRD) for different test images and variances of additive noise AIQRF-HRD estimator
Image
2 tr

nAIQRFH %

n2

AIQRFH %

AIQRFH
3.92 3.73 2.08 -5.28 25.24 31.84 34.16 30.21 3.39 14.67 14.22 -6.98

2 AIQRFH

AIQRFH
14.64 6.90 424.43 1016 1172 976.02 225.13 227.92 1038

rel AIQRFH

0 50 100 400 0 50 100 400 0 50 100 400

31.34 45.01 45.92 45.38 24.37 36.58 40.46 45.31 26.97 40.22 43.18 44.05

1.50 1.34 17.41 1.71 1.39 1.50 14.57 11.59 47.63

0.76 2.60 396.54 2.91 5.87 63.22 9.96 25.59 989.99

7.65 2.63 5.15 63.75 34.23 7.81 30.01 15.10 8.05

Concluding our analysis, we would like to mention that the method AIQRF-HRD is rather fast and computationally efficient. Segmentation of images that contain less than 106 pixels can be performed at modern computers faster than in

GoldHill

Baboon

Barbara

Proc. of SPIE Vol. 6748 67481I-10 Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2013 Terms of Use: http://spiedl.org/terms

one second. Other used operations do not require too much efforts as well. It is quite easy to determine blocks that be2 long to pre-determined image homogeneous regions and to calculate for them lh , l = 1,..., N bh . Sorting in modern DSP applications is a widely used operation and fast algorithms for its implementation exist 15. Finding a global minimum of X (t + s ) X (t ) , curve approximations by second-order polynomials and determination of a coordinate of approximation curve minimum are trivial algorithmic operations as well.

6. CONCLUSIONS
In this paper, we first briefly discuss a practical need in design and application of blind methods for noise variance evaluation in multichannel remote sensing. Note that such need can also arise in processing of color photo and medical images. The situations in which the existing methods can fail are considered. The main requirements to blind methods of noise variance evaluation are listed. Then, we briefly describe the model for distribution of local estimates of noise variance for techniques operating in spatial domain. We revisit the approach based on finding minimal inter-quantile distance and discuss the ways how its performance can be improved. The first modification deals with approximation of inter-quantile distance curve by second-order polynomials. It is shown that for model data this modification improves inter-quantile method performance. Problems in processing real life images are considered. Based on this analysis, another modification that assumes image pre-segmentation is introduced. It is demonstrated via simulations that the introduced modifications, in aggregate, produce considerable improvement of the inter-quantile method accuracy. Some aspects of algorithm implementation are discussed as well showing that an algorithm can be rather efficient.

REFERENCES
1. 2. 3. 4. 5. 6. 7. Xiuping Jia, Richards J.A., Gessner W, Ricken D.E. Remote Sensing Digital Image Analysis. An Introduction. 3-rd edition. Berlin: Springer-Verlag, 1999, 400 p. Advances in Environmental Remote Sensing, Edited by F.M. Danson and S.E.Plummer, John Wiley & Son Ltd, 1995. Hoffman R., Markman A. Interpreting Remote Sensing Imagery: Human Factors, CRC Press, LLC, 2001. http://makalu.jpl.nasa.gov/aviris.html. A. Barducci, D. Guzzi, P. Marcoinni, and I. Pippi, CHRIS-Proba performance evaluation: signal-to-noise ratio, instrument efficiency and data quality from acquisitions over San Rossore (Italy) test site, Proceedings of the 3-rd ESA CHRIS/Proba Workshop, Italy, 11 p., 2005. M.-P. Carton-Vandecandelaere, B. Vozel, L. Klaine, K. Chehdi, Application to Multispectral Images of a Blind Identification System for Blur, Additive, Multiplicative and Impulse Noises, Proceedings of EUSIPCO, Toulouse (France), III, pp. 283-286, 2002. Kulemin G.P., Zelensky A.A., Astola J.T., Lukin V.V., Egiazarian K.O., Kurekin A.A., Ponomarenko N.N., Abramov S.K., Tsymbal O.V., Goroshko Y.A., Tarnavsky Y.V., Methods and Algorithms for Pre-processing and Classification of Multichannel Radar Remote Sensing Images, TICSP Series #28, Dec. 2004, ISBN 952-15-1293-8, Tampere, Finland, TTY Monistamo, 116 p. V. Lukin, N. Ponomarenko, A. Kurekin, K. Lever, O. Pogrebnyak, L. Sanchez-Fernandez, Approaches to Classification of Multichannel Images, Proc. of CIAPR, Springer LNCS, 4225, pp. 794-803, 2006. B. Aiazzi, L. Alparone, A. Barducci, S. Baronti, I. Pippi, Information-theoretic assessment of sampled hyperspectral imagers, IEEE Transactions on Geoscience and Remote Sensing, 39, no. 7, pp. 1447-1458, July 2001. P.J. Curran, J.L. Dungan, Estimation of signal-to-noise: a new procedure applied to AVIRIS data, IEEE Trans. on Geoscience and Remote Sensing, GRS-27, no. 7, pp. 620-628, May 1989. N. Ponomarenko, V. Lukin, M. Zriakhov, A. Kaarna, Preliminary automatic analysis of characteristics of hyperspectral AVIRIS images, Proceedings of MMET, Kharkov (Ukraine), pp. 158-160, 2006. S. Mallat, A Wavelet tour of signal processing, Academic Press, San Diego, 1998. Oktem R. Transform Domain Algorithms for Image Compression and Denoising, Thesis for the degree of Doctor of Technology, Tampere (Finland), Tampere University of Technology, 2000, 142 p. R. Oktem, K. Egiazarian, V. Lukin, N. Ponomarenko, O. Tsymbal, Locally Adaptive DCT Filtering for SignalDependent Noise Removal, EURASIP Journal on Advances in Signal Processing, V. 2007, Article ID 42472, 10 p. O. Tsymbal, Multistage Robust Adaptive Filtering of Multichannel Remote Sensing Images: Thesis for the degree of Doctor of Technology, Tampere University of Technology, Tampere, Finland, June 2005.

8. 9. 10. 11. 12. 13. 14. 15.

Proc. of SPIE Vol. 6748 67481I-11 Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2013 Terms of Use: http://spiedl.org/terms

16. K.N. Plataniotis, A.N. Venetsanopoulos, Color Image Processing and Applications, Springer-Verlag, NY, 2000. 17. N. Ponomarenko, V. Lukin, M. Zriakhov, K. Egiazarian, J. Astola, Estimation of accesible quality in noisy image compression, Proceedings of EUSIPCO, Florence, Italy, 4 p., 2006. 18. Abramov S.K., Lukin V.V., Ponomarenko N.N., Egiazarian K.O., Pogrebnyak O.B. Influence of multiplicative noise variance evaluation accuracy on MM-band SLAR image filtering efficiency, Proceedings of the Fifth International Kharkov Symposium Physics and Engineering of Millimeter and Sub-Millimeter Waves, Kharkov (Ukraine), 1, pp. 250-252, 2004. 19. Chehdi K., Vozel B., Carton-Vandecandelaere M.P., Kermad C. A blind system to identify and filter degradations, Proc. IEEE-ICSP, Beijing (China), pp. 1987-1993, 2000. 20. Chehdi K., Sabri M. A new approach to identify the nature of noise affecting an image, Proceedings ICAASP-92. 3, pp. 285-288, 1992. 21. B. Vozel, K. Chehdi, L. Klaine, V.V. Lukin, S.K. Abramov, Noise identification and estimation of its statistical parameters by using unsupervized variational classification, Proceedings of ICASSP, Toulouse, France, II, pp 841-844, 2006. 22. Ramponi G., d'Alvise R. Automatic estimation of the noise variance in SAR images for use in speckle filtering, Proceedings of the IEEE-EURASIP Workshop on Nonlinear Signal and Image Processing, Antalya (Turkey), 2, pp. 835838, 1999. 23. Abramov S.K., Lukin V.V., Zelensky A.A., Astola J.T. Blind evaluation of noise variance in images using myriad operation, Proc. of IS&T/SPIE International Conference on Image Processing: Algorithms and Systems, San Jose (USA), SPIE 4667, pp. 192-203, 2002. 24. Rank K., Lendl M., Unbehauen R. Estimation of Image Noise Variance, IEE Proceedings on Vision, Image and Signal Processing, 146(2), pp. 80-84, 1999. 25. Ponomarenko N.N., Lukin V.V., Abramov S.K., Egiazarian K.O., Astola J.T. Blind evaluation of additive noise variance in textured images by nonlinear processing of block DCT coefficients, Proceedings of International Conference "Image Processing: Algorithms and Systems II, Santa Clara (CA, USA), SPIE 5014, pp. 178-189, 2003. 26. A. Zelensky, V. Melnik, V. Lukin et al, Airborne Multichannel Remote Sensing Data Processing Techniques and Software, Proc. of ERIM, San Francisco, USA, Vol. 3, pp. 151-160, 1996. 27. V.V. Lukin, S.K. Abramov, B. Vozel, B. Chehdi, A method for blind automatic evaluation of noise variance in images based on bootstrap and myriad operations, Proceedings of SPIE/EUROPTO Symp. On Satellite Remote Sensing, Bruges, Belgium, SPIE Vol. 5982, pp. 299-310, 2005. 28. V.V. Lukin, S.K. Abramov, A.A. Zelensky, J.T. Astola, Myriad based shift parameter estimation method and its application to image filtering and processing, Proceedings of SPIE Conference Mathematical Methods in Pattern and Image Analysis, San Diego, USA, SPIE Vol. 5916, pp. 1-12, 2005. 29. V.V. Lukin, S.K. Abramov, A.A. Zelensky, J.T. Astola, Use of minimal inter-quantile distance estimation in image processing, Proceedings of SPIE Conference on Mathematics of Data/Image Pattern Recognition, Compression, and Encryption with Applications IX, SPIE Vol. 6315, San Diego, USA, 12 p ., 2006. 30. V.V. Lukin, S.K. Abramov, N.N. Ponomarenko, B. Vozel, K. Chehdi, Methods for blind evaluation of noise variance in multichannel optical and radar images, Telecommunications and Radioengineering, 65 (6), pp. 509-537, 2006. 31. M. Uss, V. Lukin, I. Baryshev, B. Vozel, K. Chehdi, Joint Estimation of Additive and Impulsive Noise Parameters in Remote sensing Images with Fractal Structure, Proceedings of the International Conference Modern Problems of Radioengineering, Telecommunications and Computer Science, Lviv-Slavsko, Ukraine, pp. 232-235, 2006. 32. M. Uss, V. Lukin, S. Abramov, B. Vozel, K. Chehdi, Joint Estimation of Multiplicative and Impulsive Noise Parameters in Remote Sensing Images with Fractal Structure, Proceedings of ICASSP, 1, pp. 581-584, 2007. 33. C. Samson, L. Blanc-Fraud, G. Aubert, and J. Zerubia, A variational model for image classification and restoration, IEEE Transaction on Pattern Analysis and Machine Intelligence, 22(5), pp. 460-472, May 2000. 34. G. Aubert, L. Blanc-Fraud, and R. March. Gamma-convergence of discrete functionals with non-convex perturbation for image classification, Technical Report n4560, Institut National de Recherche en Informatique et en Automatique, September 2002. 35. G. Aubert, P. Kornpobst, Mathematical problems in image processing (Partial differential equations and the calculus of variations), Springer, 2002. 36. L. Klaine, B. Vozel, K. Chehdi, Unsupervised Variational Classification Through Image Multi-Thresholding, Proceedings of 13th EUSIPCO Conference, Antalya, Turkey, 4 p., Sept 2005.

Proc. of SPIE Vol. 6748 67481I-12 Downloaded From: http://proceedings.spiedigitallibrary.org/ on 03/09/2013 Terms of Use: http://spiedl.org/terms

You might also like