Regular Articles

Discrete cosine transform–based local adaptive filtering of images corrupted by nonstationary noise

[+] Author Affiliations
Vladimir V. Lukin

National Aerospace University, 61070, Kharkov, Ukraine

Dmitriy V. Fevralev

National Aerospace University, 61070, Kharkov, Ukraine

Nikolay N. Ponomarenko

National Aerospace University, 61070, Kharkov, Ukraine

Sergey K. Abramov

National Aerospace University, 61070, Kharkov, Ukraine

Oleksiy Pogrebnyak

Instituto Politecnico Nacional, Centro de Investigacion en Computacion, Mexico, DF 07738

Karen O. Egiazarian

Tampere University of Technology, FIN 33101, Tampere, Finland

Jaakko T. Astola

Tampere University of Technology, FIN 33101, Tampere, Finland

J. Electron. Imaging. 19(2), 023007 (May 03, 2010). doi:10.1117/1.3421973
History: Received July 11, 2009; Revised March 19, 2010; Accepted March 19, 2010; Published May 03, 2010; Online May 03, 2010
Text Size: A A A

Open Access Open Access

In many image-processing applications, observed images are contaminated by a nonstationary noise and no a priori information on noise dependence on local mean or about local properties of noise statistics is available. In order to remove such a noise, a locally adaptive filter has to be applied. We study a locally adaptive filter based on evaluation of image local activity in a “blind” manner and on discrete cosine transform computed in overlapping blocks. Two mechanisms of local adaptation are proposed and applied. The first mechanism takes into account local estimates of noise standard deviation while the second one exploits discrimination of homogeneous and heterogeneous image regions by adaptive threshold setting. The designed filter performance is tested for simulated data as well as for real-life remote-sensing and maritime radar images. Recommendations concerning filter parameter setting are provided. An area of applicability of the proposed filter is defined.

Figures in this Article

Nowadays, a great number of image filters intended for removal of noise of a given type exists (e.g., for additive,1 Poisson,2 speckle,3 and impulse4 noise). Other filters are capable to remove different types of mixed noise, such as additive and impulsive,56 multiplicative and impulsive,7 multiplicative and additive,8 etc., under assumption that a mixed noise type and its basic parameters are known in advance or preestimated. In order to perform filtering in an efficient way, one has to have a preliminary knowledge on noise type and statistics.5,910 On the other hand, such a priori information may not always be available in many typical practical situations that often arise in remote sensing, hyperspectral imaging, ultrasound medical diagnostics, cDNA imaging, etc.1115

For many imaging systems, noise is not a stationary process and its statistics and properties could be quite dissimilar in different fragments of a particular image. In practice, dissimilarities in noise statistics can be due to several basic reasons. One reason is the change of imaging conditions or considerably different distances to particular parts of an observed scene as in side-look radar imaging.16 Another reason could be uncontrollable influence of several different sources of noise. For instance, in maritime radars, interference arises due to sea clutter that depends on wind speed, direction of sea waves, incidence angle,16 etc. A third reason could be nonlinear amplification regulations in input circuits before image digitalization as in ultrasound medical devices. Finally, for some imaging systems, adequate models of noise present in the acquired or transferred images have not been established or commonly accepted yet.1720

Therefore, there are typical and atypical situations concerning properties of noise present in images. Typical situations considered thoroughly in image-filtering literature are characterized by the following: noise type is known or its characteristics are known a priori or can be estimated for a given type of noise. On the contrary, less-studied situations addressed in this paper are the following:

  1. Noise is nonstationary in the sense that its basic characteristics [variance, probability density function in some limits] vary from one local fragment of an image to another fragment.
  2. These variations do not have some predictable dependence on local mean as this takes place for (e.g., film-grain or Poisson noise; a simple example of such unpredictable variation could be linear dependence of noise variance on pixel coordinates for one or both axes of an image).
Thus, we deal with a problem of image blind denoising.21 There are several ways to cope with nonstationary noise. A simple way to follow could be applying nonlinear nonadaptive filters that are robust in wide sense9,22 (i.e., they do not need and do not take into account any information on the noise statistics). Examples of such filters are standard median, α-trimmed mean, central weighted median, Wilcoxon filters.9,23 Drawbacks of these filters are well known. Along with noise reduction, these filters can severely distort useful information contained in images.

  1. Noise is nonstationary in the sense that its basic characteristics [variance, probability density function in some limits] vary from one local fragment of an image to another fragment.
  2. These variations do not have some predictable dependence on local mean as this takes place for (e.g., film-grain or Poisson noise; a simple example of such unpredictable variation could be linear dependence of noise variance on pixel coordinates for one or both axes of an image).

These shortcomings led to design locally adaptive filters (see Refs. 7,18,2426 and references therein), including several approaches that employ partial differential equations, anisotropic diffusion, and total variation minimization.2729 They have demonstrated considerable improvement of performance in comparison to nonadaptive nonlinear filters due to exploiting different mechanisms of adaptation to image local content. However, some of them (e.g., Refs. 18,2425), require a priori information on noise type and statistics, other ones are rather complicated.

There are also many techniques that perform adaptive filtering in the domain of an orthogonal transform. To name a few, let us mention discrete wavelet transform and discrete cosine transform (DCT)–based filters for the additive noise case,1,3032 different filters designed for speckle removal,3,3334 and other types of noise.2,3537 However, for most of these filters, it is supposed that noise type is known and its parameters are either known in advance or preestimated with appropriate accuracy by preliminary “global” analysis of an image at hand.36,38a,38b

If noise statistics is unknown and noise is nonstationary, the following set of questions arises:

  1. What can serve as a basis for transform-based locally adaptive filtering of images corrupted by nonstationary noise?
  2. What can be mechanisms of local adaptation?
  3. How accurate can be local estimates of noise statistics and how this accuracy can influence a filter performance?
  4. What methods of local estimation of noise statistics are worth applying?
In our opinion, one answer to the first question could be DCT-based filtering. There are several reasons for this. First, DCT-based filtering is usually applied locally, in blocks of fixed24,31,33,37 or adaptively selected shape and size.36 Second, DCT-based filtering can be easily adapted to a priori known or preestimated noise statistics. DCT-based filters have been earlier proposed and successfully tested for removing multiplicative, Poisson, film-grain, and, in general, signal-dependent noise.2,24,31,33,37 These filters use an estimate of a local mean in an image block for determining noise local variance and setting a local threshold for hard or soft thresholding of DCT coefficients. Third, DCT-based filtering provides excellent noise suppression in homogeneous image regions,31 appropriate texture preservation,24 and good preservation of edges and fine details. Thus, the goal of this paper is to design and test DCT filter modifications applicable to processing images corrupted by nonstationary noise.

  1. What can serve as a basis for transform-based locally adaptive filtering of images corrupted by nonstationary noise?
  2. What can be mechanisms of local adaptation?
  3. How accurate can be local estimates of noise statistics and how this accuracy can influence a filter performance?
  4. What methods of local estimation of noise statistics are worth applying?

The paper is organized as follows. In Section 2, basic principles of DCT-based filtering are described. Main advantages of this approach are mentioned and demonstrated. The ways how to adapt this technique to signal-dependent noise are shown. Section 3 deals with description of two proposed modifications of locally adaptive DCT-based filter. Preliminary testing for simulated images and noise is carried out in Section 4, which also contains data for comparison of the designed methods performance to one of the state-of-the-art methods. Limitations of the method for spatially correlated noise are demonstrated. Real-life image-processing examples are given in Section 5. Finally, we draw conclusions and present possible directions of future work.

All methods of transform-based signal-image-filtering (denoising) rely on the same basic assumption that a signal in transform domain has more sparse representation than noise.30,33,3537,38a,38b,3940 Then, if a component in orthogonal transform domain has a rather large absolute value, it most probably corresponds to information content and should be preserved. If a component has amplitude close to zero (less than a threshold), then it most probably relates to noise and has to be removed (suppressed). Then, filtering can be performed using different orthogonal transforms (wavelets, discrete cosine, Haar, etc.). A choice of basis functions as well as filtering efficiency depend on how high are energy compaction properties of a transform for a given signal (image).

The Karhunen–Loeve transform possesses the best signal decorrelation property in least-squares terms.4142 Unfortunately, its coefficients are data dependent and, thus, no fast algorithm of the Karhunen–Loeve transform exists. DCT is almost as efficient as the Karhunen-Loeve transform in terms of signal decorrelation and energy compaction, especially in the case of signals modeled by Markov processes.4142 There are also fast algorithms for calculating the DCT.4243 Because of these properties, DCT is widely used in image and video compression.4445

Let us recall an operation principle of DCT-based denoising.31,33,37,42 In general form, the two-dimensional DCT coefficients of an N1×N2 matrix A may be defined asDisplay Formula

1D(u,v)=c(u)c(v)n1=0N11n2=0N21A(n1,n2)cos[(2n1+1)uπ2N1]cos[(2n2+1)vπ2N2],
whereDisplay Formula
μ=0,1,,N11,v=0,1,,N21,
Display Formula
c(u)={1N1,u=02N11uN11},
Display Formula
c(v)={1N2,v=02N21vN21}.
The standard scalar (2-D) DCT-based denoising operates in square-shaped blocks of a fixed size N×N and comprises the following steps.31,37 At the first step, the DCT is performed for each image block with values {Iqs: q=n,n+1,,n+N1, s=m,m+1,,m+N1}, to obtain the local spectrum {D(k,l,n,m): k=1,2,,N, l=1,2,,N}. The left upper corner of the image block is located at (n,m) and the indices k and l relate to the DCT (spectral) coefficients. At the second step, a thresholding of the coefficients {D(k,l,n,m)} is carried out for k=1,2,,N, l=1,2,,N. It is possible to apply either soft or hard thresholding.37 In this paper, we have applied the hard thresholding of DCT coefficients for image filtering. According to that, the spectral coefficient D(k,l,n,m) remains unchanged if its absolute value is larger than a predefined threshold T(k,l,n,m); otherwise, it is set to zero. The coefficient D(1,1,n,m), which corresponds to the block mean, is not subjected to the thresholding.

After the thresholding, the set of changed coefficients {DT(k,l,n,m): k=1,2,,N, l=1,2,,N} is obtained for each image block defined by the indices n and m. Next, the inverse DCT is applied to each block of thresholded coefficients and the preliminary filtered image values Ipf(q,s,n,m) are obtained for q=n,n+1,,n+N1, s=m,m+1,,m+N1. Because now almost all image pixels belong to N×N different image blocks, we have for each image pixel (qo,so) usually N×N preliminary filtered pixel values Ipf(qo,so,n,m), where n=qo,qo1,,qoN+1 and m=so,so1,,soN+1. Only the pixels near the image borders are exceptions because, for them, the number of such output estimates is smaller (minimally, one for all four image corners).

Then, at the last step, all output estimates for each pixel must be combined in order to obtain the final filtered image If. The simplest way to do this is to average these estimates. The resulting value for the pixel (qo,so) is thenDisplay Formula

2If(qo,so)=1N×Nn=qoN+1qom=soN+1soIpf(qo,so,n,m).
Partial overlapping of blocks can be used to accelerate processing, but it results in less-efficient filtering.31 For this reason, we exploit fully overlapping blocks.

A threshold T(k,l,n,m) may vary, depending on spatial coordinates defined by indices n and m, or may be frequency dependent46 in the case when noise is spatially correlated with known or preestimated spatial correlation characteristics. We assume that there is no a priori information about noise spatial correlation. Thus, we consider only threshold dependence on n and m.

Concerning the potential of DCT-based filtering with fully overlapping blocks, we would like to recall that it provides performance (in terms of output MSE or PSNR) better or comparable to the best wavelet-based denoising techniques for pure additive and pure multiplicative noise cases.31,33 Furthermore, the main advantage of DCT-based filtering for the considered case of nonstationary noise is that, being carried out blockwise, it can be easily adapted to local statistical properties of noise.33,47 To show how this can be done, let us consider the following two image models. The first one is a the case of film grain noise for which a general observation model isDisplay Formula

3Iqs=Iqstr+(Iqstr)γnqs,
where Iqs, Iqstr, and nqs denote the noisy image sample (pixel) value, true image value, and signal independent noise component, which is characterized by the variance σn2, respectively, for the qs’th sample, γ is a parameter of film grain noise. Then33 for an nm’th block, the threshold is set asDisplay Formula
4T(n,m)=β[Î(n,m)],
where Î(n,m) is the estimate of the local mean for the block and the coefficient β controls the threshold value. Usually, β is set approximately equal to 2.6;31,33 although for better preservation of texture, it is expedient to apply β<2.6.24

Let us confirm this fact with two examples. Figure 1 presents the plots of output MSE on β for the DCT filter described above (8×8pixel blocks with full overlapping). These plots have been obtained for two standard test images (Lenna and Baboon) corrupted by additive white Gaussian noise [γ=0 in Eq. 3] with variances σn2=50 and σn2=100, and by Poisson noise. In the former case, T(n,m) is set as βσn. In the latter case, T(n,m)=β[Î(n,m)]12.

As it is seen, all dependencies have minimums that are observed for β of 2.6 for the image Lenna and 2.3 for the image Baboon. For other standard test images (Goldhill, Peppers, Barbara) minimums are observed for β approximately equal to 2.6. Similar to other filtering algorithms, noise suppression is less efficient if a processed image has a more complex structure and/or the noise variance is smaller. Note that for additive noise with σn2=100, the values of output peak signal-to-noise ratio (PSNR) provided by the DCT filter with β=2.6 are equal to 35.40, 34.42, and 34.61dB for the images Lenna (512×512), Barbara (512×512), and Peppers (256×256), respectively. This is better than for one of the best state-of-the-art Kervrann’s filter (see Ref. 26, Table IV, PSNR=35.18, 33.79, and 34.07dB, respectively).

Thus, DCT-based filtering has a high potential. However, such performance of the DCT filter is observed under condition of known noise type and statistics. And we are interested in the case of non-stationary noise with unknown characteristics.

As is seen from Eq. 4, the product σn[Î(n,m)] is, in fact, an estimate of noise standard deviation (SD) σ̂(n,m) in a given block. For γ=0 (i.e., for pure additive noise), one gets exactly σn. The threshold is fixed and equal to βσn. Similarly, for pure multiplicative case σ̂(n,m)=σμ(n,m)I¯̂(n,m) where σμ is the multiplicative noise normalized SD. As an estimate of local mean I¯̂(n,m), one can use the DCT spectral coefficient D(1,1,n,m).

Thus, generalizing the presented results, the conclusion is as follows. To set a local threshold, one needs a local estimate of noise standard deviation σ̂(n,m). Consider the following image additive observation modelDisplay Formula

5Iqs=Iqstr+nqsnst,
where Iqstr is a true image value and nqsnst denotes nonstationary noise in a qs’th sample assumed to be zero mean. We suppose that its SD σqs is a function of pixel coordinates defined by pixel indices qs. It is also assumed that spatial variations of σqs are not fast, and for a given block, it is possible to assume that nonstationary noise SD is almost constant for all image pixels that belong to a given block. Then, for each block, one needs to have an estimate σ̂(n,m). This estimate can be used for setting a local threshold proportionally to σ̂(n,m).39,48

DCT Filter Adaptive to Image Local Statistics

Here, we discuss the fourth question given in Section 1 (i.e., what methods of local estimation of noise statistics can be used). Some initial imagination concerning properties of local estimates of noise variance in blocks or scanning windows with fixed size can be got from earlier studies.4952 Generally speaking, noise local SD can be estimated in the spatial or spectral domain.4953 These local estimates are then aggregated (processed) in a robust manner to provide blind estimation of noise characteristics based on a model of noise assumed known a priori. Pure additive or multiplicative noise variances can be blindly estimated.4952 More sophisticated methods allow estimating dependence of local variance on local mean for signal-dependent noise.36,53

The following are useful conclusions drawn from several papers:36,4953

  1. Accuracy of local estimates of noise variance depends on several basic factors, namely, block size, image content, true values of noise variance, spatial correlation of noise, a used method for obtaining local estimates, and its parameters.
  2. Better accuracy (in terms of smaller relative errors) is provided for simpler content (less textural) images if noise variance is rather large and noise is spatially uncorrelated.
  3. The appropriate block (scanning window) size is from 7×7to9×9pixels5 (i.e., 8×8pixel blocks used in DCT-based filtering conform well with this recommendation).
  4. In any case, there are normal local estimates of noise variance obtained in homogeneous image blocks that are close to true values of noise variance and abnormal estimates obtained in heterogeneous image blocks that are considerably larger than the true values of noise variance.
These conclusions indicate that it is hard to expect good accuracy of noise variance evaluation for blocks of a limited size, especially if image content in a given block is quite complex and noise is not intensive. To prove this, consider first local estimates of noise variance in spatial domainDisplay Formula
6σ̂c2(n,m)=q=nn+7s=mn+7(IqsI¯̂nm)263,
where Iqs is qs’th pixel value within nm’th image block and I¯̂nm denotes the nm’th block mean. It is easy to show thatDisplay Formula
7σ̂c2(n,m)Var{Itr(n,m)}+σn2(n,m),
where Var{Itr(n,m)}=q=nn+7s=mn+7[(IqstrI¯nmtr)263], I¯nmtr=q=nn+7s=mn+7Iqstr64, σn2(n,m) denotes the noise variance in the nm’th block. Thus, conventional estimates of local variance [Eq. 6] are sensitive to image content in a block. It is an inherent property of such estimates, and the question is how to exploit this sensitivity in a reasonable manner. It can be a drawback or a positive feature for the considered situation of unknown noise variance. The reasons will be considered next.

  1. Accuracy of local estimates of noise variance depends on several basic factors, namely, block size, image content, true values of noise variance, spatial correlation of noise, a used method for obtaining local estimates, and its parameters.
  2. Better accuracy (in terms of smaller relative errors) is provided for simpler content (less textural) images if noise variance is rather large and noise is spatially uncorrelated.
  3. The appropriate block (scanning window) size is from 7×7to9×9pixels5 (i.e., 8×8pixel blocks used in DCT-based filtering conform well with this recommendation).
  4. In any case, there are normal local estimates of noise variance obtained in homogeneous image blocks that are close to true values of noise variance and abnormal estimates obtained in heterogeneous image blocks that are considerably larger than the true values of noise variance.

Let us now concentrate on estimation of noise local SD in the spectral (DCT) domain. For a given block, it is possible to estimate σ̂(n,m) in a spectral domain as39,48Display Formula

8σ̂(n,m)=1.483med{|D(k,l,m,n)|:k=1,,8;l=1,,8;exceptk=l=1},
where med{ } denotes the median of a data sample. This estimate is similar to the median of absolute deviations used as a robust data scale estimator.9 Here [in Eq. 8], the noise SD (scale) estimate σ̂(n,m) exploits properties of transform coefficients when mixed Gaussian–Laplacian or Gaussian scale mixture models for their description are used.54 Two properties of orthogonal transforms and robust estimation are exploited in Eq. 8. First, noise after an orthogonal transform spreads between all coefficients, and, if noise is independent identically distributed (i.i.d), its power spreads uniformly where spectral coefficients occur to be Gaussian random variables. Then, the SD of spectral coefficients is proportional to the noise SD. Second, the use of the robust estimate (median) provides less sensitivity to outliers where the outliers in the considered case are induced by image content.

However, although the used estimator (sample median) is robust, the image local content (details, texture, edges) present in a given block leads to the positive bias of the estimate σ̂(n,m).52 Note that if noise is spatially correlated, considerable bias can be observed for estimates obtained according to Eq. 651 (these estimates in homogeneous regions are, on average, smaller than the true value of noise SD).

Let us analyze how great the influence of image local content and other factors could be on accuracy of local estimates. Consider a test image [e.g., the standard test image Barbara, (Fig. 2)] corrupted by white additive noise with a zero mean and constant variance σn2=100. The values of σ̂(n,m) from Eq. 6 are visualized in Fig. 2; larger values are lighter (magnification by 7 is used for better visualization). It is seen that the values σ̂(n,m) for blocks that correspond to homogeneous image regions vary a little. They are mostly rather small and approximately equal to σn of additive noise. The values σ̂(n,m) for the blocks that correspond to image heterogeneities such as edges, textures, and details, are commonly slightly larger due to influence of image content.

Behavior of σ̂c(n,m) is similar [see Fig. 2 magnification is by 3.5], but its sensitivity to image heterogeneities in blocks is considerably larger. This can be also seen from the comparison of histograms of σ̂(n,m) and σ̂c(n,m) presented in Figs. 3 respectively. In both cases, normal local estimates (which are obtained in homogeneous image blocks) are grouped near the true value equal to 10. But for the histogram in Fig. 3 there are more abnormal local estimates and they are, on average, larger.

This analysis allows proposing a locally adaptive DCT-based filter that will be further referred as LA DCT-1. This filter algorithm is as follows:

  1. For each given block, estimate σ̂(n,m) according to Eq. 8 and set a local threshold as T(n,m)=βσ̂(n,m).
  2. All other operations (DCT in blocks, thresholding, inverse DCT, and averaging for overlapping blocks) are performed as described in Section 2.
We prefer to use σ̂(n,m), but not σ̂c(n,m), because σ̂(n,m) is less sensitive to heterogeneities. Note that the use of a larger threshold in DCT-based filtering leads to oversmoothing. As follows from the algorithm description, the proposed filter LA DCT-1 adapts to noise local characteristics. This is adaptation mechanism 1.

  1. For each given block, estimate σ̂(n,m) according to Eq. 8 and set a local threshold as T(n,m)=βσ̂(n,m).
  2. All other operations (DCT in blocks, thresholding, inverse DCT, and averaging for overlapping blocks) are performed as described in Section 2.

Improved Locally Adaptive DCT Filter

In this section, we consider one more way (mechanism 2) to further improve performance of the locally adaptive DCT-based filter. Our approach is based on the following general idea and assumptions. Suppose that we have succeeded in discriminating homogeneous and heterogeneous regions of an image. Because in homogeneous regions, local estimates σ̂(n,m) are quite close to true values of local SD of noise, then it is a correct decision to set the local threshold equal to 2.6σ̂(n,m). On the contrary, if a given block corresponds to an image heterogeneous region, then, most probably, σ̂(n,m) is larger than the local SD of noise. Then, it is reasonable to set the local threshold less than 2.6σ̂(n,m) as β(n,m)σ̂(n,m), where β(n,m)<2.6. In general, there are many possible ways to set β(n,m). It can be fixed and equal to some βhet2.6 in a simplest case or it can be determined in a more complicated manner.

Thus, a primordial task is to design some discriminator for homogeneous and heterogeneous blocks. It is easy to resolve this task if one knows the local SD of noise in advance.7 However, it becomes more complicated if the local SD of nonstationary noise is unknown. To get around this shortcoming, let us exploit the properties of local estimates σ̂(n,m) (in spectral domain) and σ̂c(n,m) (in spatial domain) established in Section 3a.

Note that usually7σ̂c(n,m) is defined in square-shaped scanning windows, where window side size is odd (e.g., 5 or 7).7,23 Because here we deal with DCT-based filtering in 8×8 blocks, it is possible to calculate σ̂c(n,m) either according to Eq. 6 or as follows:Display Formula

9σ̂c2(n,m)=q=n+1n+Qs=m+1n+S(IqsI¯̂qs)2QS1,I¯̂qs=q=n+1n+Qs=m+1n+SIqsQS,Q=S=6.
The map of the estimates σ̂c(n,m) for Q=S=6 is visualized in Fig. 2 for the test image Barbara corrupted by zero mean additive noise with the constant variance σ2=100. Because the estimate σ̂c(n,m) is much more sensitive to image heterogeneities in a block than the estimate σ̂(n,m), let us exploit this difference for discriminating homogeneous and heterogeneous image blocks. For this purpose, consider a ratio R(n,m)=[σ̂c(n,m)σ(n,m)]. Its example for the test image Barbara and additive Gaussian noise with σn2=100 is visualized in Fig. 4. We represent the ratio map as an image Rvis(n,m)=[40R(n,m)], n=1,,Nim7, m=1,,Mim7, where [•] means rounding to the nearest positive integer not larger than 255, Nim×Mim is the image size, 40 is a magnification coefficient used to visualize better the maps R(n,m).

Joint visual analysis of the test image Barbara [Fig. 2] and the ratio map [Fig. 4] shows that quite large ratios (essentially larger than unity, indicated by brighter color pixels) are observed for blocks located in the heterogeneous image regions. Figure 4 presents the ratio map for the same test image but corrupted by Poisson noise. The obtained maps are rather similar. This indicates that the method of analyzing local activity (heterogeneity) based on R(n,m) is applicable to different types of noise. This observation allows expectation that the proposed principle will also work well enough for any nonstationary noise.

There are several possible ways to exploit the aforementioned property. The simplest one is to adaptively set the local thresholdDisplay Formula

10T(n,m)={βσ̂(n,m),ifR(n,m)<TRβhetσ̂(n,m),ifR(n,m)TR},
where β=2.6 as in LA DCT-1, TR is a preset threshold, βhet<2.6 is a factor that determines the hard threshold for DCT coefficients in edge/detail neighborhoods and in textural regions. The DCT-based filter described by Eqs. 5,910 (further referred as to LA DCT-2) belongs to the class of locally adaptive hard-switching filters,7 where hard switching relates to the parameter β. This filter implies both adaptions to noise characteristics (mechanism 1) and to image content (mechanism 2).

LA DCT-2 uses two new parameters, TR and βhet. Consider distributions of the ratio R(n,m), which is a random variable. Suppose that the estimates σ̂c(n,m) and σ̂(n,m) are obtained in a homogeneous image region. Then, their means under condition of i.i.d. Gaussian noise are the same. Thus, it is possible to expect that the distribution of the ratio R(n,m) should have a mode in the neighborhood of unity. As follows from the ratio maps in Figs. 4, σ̂c(n,m) is usually larger than σ̂(n,m) in image heterogeneous regions. Therefore, the distribution of the ratio R(n,m) may have a heavy “right-hand” tail. The histogram of the obtained R(n,m) values for the test image Barbara (pure additive Gaussian noise, σn2=100) is presented in Fig. 5. The histogram for the image Lenna corrupted by Poisson noise is shown in Fig. 5. Both distributions, as expected, have modes in the neighborhoods of unity, and they possess a heavy right-hand tail. Similar shapes of R(n,m) value distributions have been observed for other conventional test images as Peppers, Goldhill, etc. The detailed analysis of values of these histograms shows that the distribution mode, which corresponds to homogeneous image regions, is about unity. Being random, the values R(n,m) form the quasi-Gaussian part of this distribution. As in any discrimination or detection task, a larger preset discrimination threshold provides better correct detection of homogeneous regions but larger probability of recognizing heterogeneous region as homogeneous. Thus, threshold setting is a compromise. Analysis of histogram data shows that a trade-off in discriminating blocks that correspond to homogeneous and heterogeneous regions can be provided by setting TR1.35.

Performance Analysis for Test Images in Case of Spatially Uncorrelated Noise

Let us carry experiments with different types of noise and test images. Recall that if noise type and characteristics are a priori known, it is possible to carry out filtering in a priori adjusted or “ideal” manner. For example, if one knows that noise is additive with fixed σn2, then it is possible to apply the standard DCT-based filter31 with the fixed T(n,m)=2.6σn. Similarly, if, e.g., noise is Poissonian, filtering can be done with T(n,m)=βI¯̂(n,m) (see Section 2). Quantitative performance characteristics for such filters that presume availability of full a priori information and its use, IdDCT, can serve as benchmarks of what can be reached if a priori information is available and is exploited in the DCT-based filtering.

Let us analyze the filter performance in conventional terms of output MSEout=(1NimMim)n=1Nimm=1Mim(InmfInmtr)2 where MSE is mean square error, Nim×Mim is the image size, Inmf is the nm’th pixel value for the processed (filtered) image. The simulation results have been obtained for the traditional set of test gray-scale images (512×512pixels), for Gaussian additive (σn2=100) and Poisson noises. These simulation results are collected in Table 1. Two values of β have been used, namely, β=2.6 and 2.3 as motivated by dependences in Fig. 1 and analysis performed in Sections 23.

Table 1 also contains the values of MSEout for four nonadaptive nonlinear filters: the standard median with the scanning window sizes 5×5 and 3×3pixels, the 3×3α-trimmed mean (ATM) filter with trimmed two largest and two smallest values in the scanning window, and the 3×3 center-weighted median (CWM) filter with the central pixel weight equal to 3.23

Analysis of data in Table 1 shows the following:

  1. The IdDCT-based filters always perform better (produces smaller MSEout) than the proposed LA DCT-1; worse performance of LA DCT-1 is explained by inaccuracies of the used local estimates of noise SD.
  2. The IdDCT-based filter with β=2.6 performs better than for β=2.3 for four considered test images; the only exception is the test image Baboon which is the most textural.
  3. The difference of MSEout for the IdDCT-based filter and LA DCT-1 for the same β is not large; the only exception is again the image Baboon. The reason is that for this image the estimates σ̂(n,m) are mostly considerably larger than the true values of local SD of noise, and this leads to oversmoothing.
  4. The results for Gaussian and Poisson noise are in good agreement with each other, which is evidence that the proposed LA DCT-1 is able to carry out denoising well enough, except in the case of highly textural images.
  5. The fact that for the test image Baboon both IdDCT-based filter and LA DCT-1 produce smaller MSEout for smaller β shows that one way to improve filtering performance is to detect textural regions in images and to set smaller β for them. This idea has been put behind the method LA DCT-2 (see Section 3b).
  6. The values MSEout for all considered nonadaptive nonlinear filters are considerably larger than for LA DCT-1 in almost all cases; this additional time demonstrates drawbacks of these nonadaptive filters.
Concerning LA DCT-2, we have carried out numerical simulations for a set of values TR (close to 1.35) and a set of βhet varied in rather wide limits. The simulation data are presented in Table 2 for the five test images corrupted by additive Gaussian noise with variance equal to 100 and Poisson noise.

  1. The IdDCT-based filters always perform better (produces smaller MSEout) than the proposed LA DCT-1; worse performance of LA DCT-1 is explained by inaccuracies of the used local estimates of noise SD.
  2. The IdDCT-based filter with β=2.6 performs better than for β=2.3 for four considered test images; the only exception is the test image Baboon which is the most textural.
  3. The difference of MSEout for the IdDCT-based filter and LA DCT-1 for the same β is not large; the only exception is again the image Baboon. The reason is that for this image the estimates σ̂(n,m) are mostly considerably larger than the true values of local SD of noise, and this leads to oversmoothing.
  4. The results for Gaussian and Poisson noise are in good agreement with each other, which is evidence that the proposed LA DCT-1 is able to carry out denoising well enough, except in the case of highly textural images.
  5. The fact that for the test image Baboon both IdDCT-based filter and LA DCT-1 produce smaller MSEout for smaller β shows that one way to improve filtering performance is to detect textural regions in images and to set smaller β for them. This idea has been put behind the method LA DCT-2 (see Section 3b).
  6. The values MSEout for all considered nonadaptive nonlinear filters are considerably larger than for LA DCT-1 in almost all cases; this additional time demonstrates drawbacks of these nonadaptive filters.

As follows from analysis of the obtained results, for most images it is reasonable to set TR1.3, βhet1.5. This choice produces falling into the neighborhood of minimal MSEout for both noise models and for the four considered test images. The only exception is the test image Baboon, for which it is reasonable to use smaller βhet.

Reduction of MSEout by approximately 10–30% is observed for LA DCT-2 in comparison to LA DCT-1 with fixed β (compare the data in Tables 12). Improvement in comparison to LA DCT-1 with fixed β is larger for more complex images. The values of MSEout are almost the same as were observed for the IdDCT filter. Only for test image Baboon are the results not good enough.

We have examined the a reason for this. It occurred that the values σ̂(n,m) in image texture regions are about four times larger than the true values of local SD of noise σ̂(n,m). Thus, even if βhet is set to 0.8, the local threshold is larger than 3σ̂(n,m). Because of this, some oversmoothing takes place even for LA DCT-2.

In addition to hard switching of β according to Eq. 10, we have also analyzed another (soft) algorithm of threshold adaptation as follows:Display Formula

11T(n,m)=2.6σ̂(n,m)R(n,m)=2.6σ̂(n,m)σ̂c(n,m)σ̂(n,m)=2.6σ̂2(n,m)σ̂c(n,m).
We have tested this method for the same test images and noise models as presented in Table 2. Very similar results as for the LA DCT-2 with optimally set TR1.3, βhet1.5 have been obtained. One reason why it is difficult to improve performance of the locally adaptive DCT-based filters is that there are many blocks for which the ratio R(n,m)=[σ̂c(n,m)σ(n,m)] is close to unity although these blocks are, in fact, heterogeneous. This means that, in the future, it is worth studying other local parameters than R(n,m) that should be able to better discriminate homogeneous and heterogeneous image blocks. In particular, gaussianity tests in spatial or spectral domain are worth studying.55

Let us compare the data obtained for LA DCT-2 to other locally adaptive filters. The best known results of removing noise with a priori unknown characteristics are presented in the paper of Kervrann and Boulanger,26 who perform a thorough comparison of their method to other filtering methods and demonstrate superior performance of their approach. Thus, we present our results obtained for the recommended TR1.3, βhet1.5 and Kervrann and Boulanger’s data for the same test images and noise variances (Table 3). Noise is additive white Gaussian. Filter performance is characterized by output PSNR.

As is seen, the method26 produces slightly better results for larger noise variance and simpler test images. In turn, our method LA DCT-2 provides larger PSNR for smaller SD of noise and more textural images. The advantage of the proposed method LA DCT-2 is that image processing is quite fast. This is due to the fact that two basic operations used in the proposed techniques are DCT and data sorting where both can be realized using fast algorithms.42

Filter Performance in Spatially Correlated Noise Environment

The case of spatially uncorrelated (i.i.d.) noise has been studied. However, in practice, it is possible that noise can be spatially correlated. Furthermore, 2-D autocorrelation function of spatially correlated noise can be unknown in advance.7,51

We have analyzed how the proposed LA DCT-2 filter performs in the situation of spatially correlated noise. For this purpose, we have simulated additive Gaussian zero mean spatially correlated noise with variance 100. Spatially correlated noise has been modeled by applying the 3×3 window mean filter to originally i.i.d. zero mean Gaussian 2-D noise and adjusting a desired variance. The results are presented in Table 4 for three values of TR and several values of βhet. The test images are the same as earlier.

Analysis of the data presented in Table 4 shows the following:

  1. For rather simple images (Lenna, Peppers, Barbara, Goldhill), there is some reduction of noise, but it is considerably smaller than in the earlier considered case of i.i.d. noise. For the image Baboon, it is still difficult to provide appropriate performance of the filter.
  2. For the case of spatially correlated noise, the recommendation to set TR1.3, βhet1.5 is incorrect. The results obtained for TR=1.37 and βhet=2.6 occurred to be better than for TR1.3, βhet1.5 for four test images.
Thus, additional analysis of why it happens is needed. Some reasons are clear; they come from studies done recently.44,56 In particular, it has been demonstrated that frequency-dependent thresholds should be used in DCT-based filter if noise is spatially correlated. Denoising with frequency-dependent thresholding can be carried out if either spatial correlation properties of noise are known in advance or preestimated. Thus, it becomes necessary to adapt not only to local statistics of noise and image local content, but also to spatial correlation characteristics of noise.

  1. For rather simple images (Lenna, Peppers, Barbara, Goldhill), there is some reduction of noise, but it is considerably smaller than in the earlier considered case of i.i.d. noise. For the image Baboon, it is still difficult to provide appropriate performance of the filter.
  2. For the case of spatially correlated noise, the recommendation to set TR1.3, βhet1.5 is incorrect. The results obtained for TR=1.37 and βhet=2.6 occurred to be better than for TR1.3, βhet1.5 for four test images.

Another reason is that statistical characteristics of σ̂c(n,m) and σ̂(n,m) as well as their ratio R(n,m) change if noise is spatially correlated with respect to the i.i.d. noise case. Let us analyze histograms of R(n,m) for spatially correlated noise. Two typical histograms are presented in Fig. 6. Their shapes are similar to the shapes of the histograms represented in Fig. 5. However, there is an obvious difference. The modes of the histograms in Fig. 6 occur not in the neighborhood of unity but for larger values.

This phenomenon has its explanation. For spatially correlated noise, both estimates σ̂c(n,m) and σ̂(n,m) in homogeneous blocks (that form histogram maximum) become biased, smaller than the true value of noise SD.51 But the bias Δc of the estimates σ̂c(n,m) can be about (0.10.2)σn2, whereas for the estimates σ̂(n,m) the bias Δ can reach (0.70.8)σn2.51 Thus, the estimates σ̂c(n,m) in image homogeneous blocks are, on average, larger than the estimates σ̂(n,m). This leads to mode shifting to values larger than unity.

Let us analyze the mode estimates r̂ for distributions of R(n,m). The obtained data are presented in Table 5. The method49 has been applied for mode estimation.

As is seen, for spatially uncorrelated noise the estimates r̂ are quite close to unity for all test images for both additive i.i.d. Gaussian and Poissonian noise. One interesting observation is that, for images with more complicated structure (e.g., Baboon), the estimates r̂ are larger. Similarly, if noise is spatially correlated, the estimates r̂ are considerably larger than 1.0 for all five considered test images.

Two conclusions follow from this analysis:

  1. Practical cases when an image is either corrupted by spatially correlated noise or is textural can be discriminated from cases when an image is rather simple and corrupted by i.i.d. noise. This can be done by comparing r̂ to some threshold Tspc (e.g., approximately equal to 1.15: if r̂<Tspc), image filtering by LA DCT-2 can be efficient (see data in Section 4a), whereas in the opposite case, a more complicated filtering should be applied.
  2. If noise is spatially correlated, it seems reasonable to use TR>1.3; adaptation to noise spatial correlation is needed.
Thus, we have determined practical situations for which the designed filter LA DCT-2 is worth applying. A decision of whether or not filtering has to be applied can be undertaken automatically by analyzing r̂ for an image at hand. Design of adaptive DCT-based filters for complex-structure images corrupted by spatially correlated noise is a topic of future research.

  1. Practical cases when an image is either corrupted by spatially correlated noise or is textural can be discriminated from cases when an image is rather simple and corrupted by i.i.d. noise. This can be done by comparing r̂ to some threshold Tspc (e.g., approximately equal to 1.15: if r̂<Tspc), image filtering by LA DCT-2 can be efficient (see data in Section 4a), whereas in the opposite case, a more complicated filtering should be applied.
  2. If noise is spatially correlated, it seems reasonable to use TR>1.3; adaptation to noise spatial correlation is needed.

Let us present two examples of applying the proposed filters to real-life data. One example is a set of polarimetric radar data, presented as real valued (floating point) data arrays. These data have been obtained by a maritime coastal radar. The images are presented in Fig. 7. HH means that a horizontally polarized signal is transmitted and then received; VH relates to a vertically polarized emitted and horizontally polarized received signal.

A sensed area mainly corresponds to sea surface (that occupies basic part of images and form background), one large rock (small rocky island placed in the image central part, very bright pixels), several small ones (left lower corner, groups of bright pixels), and a shadowed zone behind the large rock. The horizontal axis of images corresponds to range and vertical axis relates to azimuth of the radar. During data acquisition, internal gain control was used to partly compensate dependence of backscattered signal mean on distance.

Visual analysis of these images allows concluding that background intensity varies depending on range. More detailed analysis has been done to confirm this conclusion. First, histograms (Fig. 8) have been obtained for manually selected six “homogeneous” regions (marked by frames in images in Fig. 7) for two different sectors and three different mean distances. Analysis of sample histograms in Fig. 8 shows that the histogram shape changes with range. The distributions are asymmetric with respect to their means. There is a heavy tail to the right side from the distribution mode. Distribution modes are also different for different ranges. Outliers are seldom (occur with quite small probabilities), their values differ from the mode values by approximately ten times.

To carry out more thorough analysis, we have determined the following parameters for each image homogeneous region:

  1. Minimal, maximal, and mean mi values for each i’th fragment, i=1,,6
  2. Variance σi2 and relative variance σreli2 calculated as variance divided by squared mean
  3. Interquantile estimate of relative variance σIQi249 (this estimate is practically insensitive to the presence of outliers, trend, and heterogeneities in data)

  1. Minimal, maximal, and mean mi values for each i’th fragment, i=1,,6
  2. Variance σi2 and relative variance σreli2 calculated as variance divided by squared mean
  3. Interquantile estimate of relative variance σIQi249 (this estimate is practically insensitive to the presence of outliers, trend, and heterogeneities in data)

Table Grahic Jump Location
Output MSE for “ideal” DCT-based (IdDCT) filter, LA DCT-1, and nonlinear nonadaptive filters.
Table Grahic Jump Location
Values of MSEout for several βhet and TR for the considered test images.
Table Grahic Jump Location
PSNR for the compared LA-DCT-2 and Kervrann’s filters.
Table Grahic Jump Location
Values of MSEout for several βhet and TR for the considered test images corrupted by spatially correlated noise.
Table Grahic Jump Location
Estimates r̂ for the test images corrupted by different types of noise.
Graphic Jump LocationF1 :

Output MSE versus β for two standard test images: (a) Lenna and (b) Baboon, corrupted by additive white Gaussian noise with variances 50 (dotted line) and 100 (dashed line) and by Poisson noise (solid line).

Graphic Jump LocationF2 :

(a) Noise-free test image Barbara, (b) visualized σ̂(n,m) of test image Barbara (values have been multiplied by 7) and (c) σ̂c(n,m) (values have been multiplied by 3.5).

Graphic Jump LocationF3 :

Histograms of (a) σ̂(n,m) and (b) σ̂c(n,m) for the image Barbara, σn2=100.

Graphic Jump LocationF4 :

Magnified ratio image Rvis(n,m)=[40R(n,m)] for the test image Barbara (a) corrupted by additive noise and (b) corrupted by Poisson noise.

Graphic Jump LocationF5 :

Histograms of the ratio R(n,m) for the test images: (a) Barbara corrupted by zero mean pure additive Gaussian noise with σn2=100 and (b) Lenna corrupted by Poisson noise.

Graphic Jump LocationF6 :

Histograms of the ratio R(n,m) for the test images: (a) Lenna and (b) Barbara corrupted by zero mean spatially correlated additive Gaussian noise with σn2=100.

Graphic Jump LocationF7 :

Data registered for (a) HH and (b) VH polarization of radar signal.

Graphic Jump LocationF8 :

Sample histograms for the image in Fig. 7 (a) for small distance fragment 1 and (b) for middle distance fragment 3.

The results obtained for the image in Fig. 7 are given in Table 6. For more distant fragments (5 and 6), the values mi and σi2 are larger than for less distant fragments, especially for the fragments 1 and 2. In turn, the standard and robust (interquantile) estimates of relative variance for more distant fragments are smaller. This shows that noise is neither pure additive nor pure multiplicative. A specific “distance-dependent” noise is observed. This conclusion has been confirmed by results of analysis carried out for the image in Fig. 7. The observed nonstationarity results from joint influence of several factors, namely, specific features of the receiver amplifier gain control used, varying incidence (grazing) angle of backscattering from wavy sea surface, different mutual geometry of sea wave direction and radar azimuth, etc. This is the case when it is difficult to separate the influence of these factors. We assume that nonstationary noise local statistics does not change abruptly [i.e., it is possible to consider it practically constant for fragments (blocks) of relatively small size, let us say 8×8 blocks commonly used in DCT-based filtering].

Table Grahic Jump Location
Statistical characteristics of the selected “homogeneous” regions for the image in Fig. 7.

Figure 9 presents the processed images for the maritime radar data carlier given in Fig. 8. Note that in this case, a two-stage procedure has been applied. At the first stage, the CW Median filter23 with the 3×3 scanning window and the center element weight Wc=5 has been used to remove outliers. Then, at the second stage, the LA DCT-1 has been used. Comparing images in Figs. 7,9, one can conclude that noise is well suppressed and useful information in images is preserved well enough, although some oversmoothing of fine details is observed.

Graphic Jump LocationF9 :

Output images for (a) HH and (b) VH polarization of radar signal.

Let us give another example of a real-life image processing. The 224th subband image of the Lunar Lake AVIRIS5759 image is presented in Fig. 10. Noise present in this image is visually seen. The output image for LA DCT-1 is shown in Fig. 10. Noise is well suppressed, although sharp edges are slightly smeared. The ratio image Rvis(n,m)=[40R(n,m)] is visualized in Fig. 10. The most sharp edges and details are marked by brighter color pixels in this map. This allows better preservation of sharp edges and details in the output image by the designed LA DCT-2 [see Fig. 10TR=1.35, βhet=1.7, r̂=1.03].

Graphic Jump LocationF10 :

(a) Original one-channel image of AVIRIS hyperspectral data, (b) the output of LA DCT-1 filter, (c) the ratio Rvis(n,m) image, and (d) the output of the hard switching LA DCT-2 filter.

It is shown that there are practical situations where noise is nonstationary and limited a priori information on its statistics is available. To filter images effectively under these conditions, two DCT-based filtering techniques to suppress nonstationary spatially uncorrelated noise have been proposed and studied.

The first technique is a locally adaptive filter based on local estimation of noise variance in blocks and setting the corresponding threshold proportionally to the obtained estimate of noise SD. It performs well enough for rather simple images. The second locally adaptive filter employs, in addition, the analysis of the ratio R(n,m) and adapts to image content in a block. This leads to decreasing MSEout and improving edge/detail/texture preservation in processed images. As a result, the performance of this filter is comparable to performance of the best state-of-the-art methods.

The recommendations concerning proper selection of the filter adaptation parameters are given. The designed filters have been applied to real-life images and have demonstrated excellent results. The spatially correlated noise case has been studied as well. The proposed methods do not perform well enough for this case and should be further modified. A way to discriminate the cases of spatially uncorrelated and correlated noise has been proposed. The design of locally adaptive filters for spatially correlated noise case may be a subject of our future work.

We are thankful to anonymous reviewers for their valuable comments and propositions. The work of O. Pogrebnyak was partially supported by Instituto Politecnico Nacional as a part of the research project SIP No. 20101090. Images have been kindly provided by Dr. A.V. Popov, of National Aerospace University, Kharkov, Ukraine.

Shui  P.-L., “ Image denoising algorithm via doubly local wiener filtering with directional windows in wavelet domain. ,” IEEE Signal Process. Lett..  1070-9908 12, (10 ), 681–684  (Oct. (2005)).
Foi  A., , Katkovnik  V., , and Egiazarian  K., “ Signal-dependent noise removal in point-wise shape-adaptive DCT domain with locally adaptive variance. ,” in  Proc. of European Signal Processing Conf. (EUSIPCO 2007). , CD-ROM,  EURASIP ,  Poznan, Poland  ((2007)).
Solbo  S., and Eltoft  T., “ Homomorphic wavelet-based statistical despeckling of SAR images. ,” IEEE Trans. Geosci. Remote Sens..  0196-2892 42, (4 ), 711–721  ((2004)).
Eng  H.-L., and Ma  K.-K., “ Noise adaptive soft switching median filter. ,” IEEE Trans. Image Process..  1057-7149 10, (2 ), 242–251  ((2001)).
Barner  K. and Arce  G., Eds.,  Nonlinear Signal and Image Processing: Theory, Methods, and Applications. ,  CRC Press ,  Boca Raton  ((2003)).
Plataniotis  K. N., and Venetsanopoulos  A. N.,  Color Image Processing and Applications. ,  Springer-Verlag ,  New York  ((2000)).
Melnik  V., “ Nonlinear locally adaptive techniques for image filtering and restoration in mixed noise environments. ,” PhD Thesis, Tampere University of Technology, Tampere, Finland ((2000)), ⟨http://www.atilim.edu.tr/~roktem/Research/interests.htm⟩.
Lukin  V. V., , Ponomarenko  N. N., , Abramov  S. K., , Vozel  B., , Chehdi  K., , and Astola  J., “ Filtering of radar images based on blind evaluation of noise characteristics. ,” Proc. SPIE.  0277-786X 7109, , 71090R  (Sept. (2008)).
Astola  J., and Kuosmanen  P.,  Fundamentals of Nonlinear Digital Filtering. ,  CRC Press ,  Boca Raton, FL  ((1997)).
Touzi  R., “ A review of speckle filtering in the context of estimation theory. ,” IEEE Trans. Geosci. Remote Sens..  0196-2892 40, (11 ), 2392–2404  ((2002)).
Barton  D. K.,  Radar System Analysis and Modeling. ,  Artech House ,  Boston  ((2005)).
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., , and Tarnavsky  Y. V.,  Methods and Algorithms for Pre-processing and Classification of Multichannel Radar Remote Sensing Images. ,  TTY Monistamo ,  Tampere, Finland  (Dec. (2004)).
Lopez-Martinez  C., and Pottier  E., “ On the extension of multidimensional speckle noise model from single-look to multilook SAR imagery. ,” IEEE Trans. Geosci. Remote Sens..  0196-2892 45, (2 ), 305–320  ((2007)).
Chang  C.-I., Ed.,  Hyperspectral Data Exploitation: Theory and Applications. ,  Wiley ,  Hoboken, NJ  ((2007)).
Lukac  R., , Plataniotis  K. N., , Smolka  B., , and Venetsanopoulos  A. N., “ cDNA microarray image processing using fuzzy vector filtering framework. ,” J. Fuzzy Sets Syst.. 152, (1 ), 17–35  (May (2005)).
Kulemin  G. P.,  Millimeter-Wave Radar Targets and Clutter. ,  Artech House ,  Boston  ((2003)).
Oliver  C., and Quegan  S.,  Understanding Synthetic Aperture Radar Images. ,  SciTech Publishing ,  Raleigh, NC  ((2004)).
Lee  J. S., , Wen  J. H., , Ainsworth  T. I., , Chen  K. S., , and Chen  A. J., “ Improved sigma filter for speckle filtering of SAR imagery. ,” IEEE Trans. Geosci. Remote Sens..  0196-2892 47, (1 ), 202–213  (Jan. (2009)).
Barducci  A., , Guzzi  D., , Marcoionni  P., , and Pippi  I., “ CHRIS-Proba performance evaluation: signal-to-noise ratio, instrument efficiency and data quality from acquisitions over San Rossore (Italy) test site. ,” in  Proc. 3rd ESA CHRIS/Proba Workshop. , CD-ROM,  ESA-ESRIN ,  Frascati, Italy  ((2005)).
Koivisto  P., , Astola  J., , Lukin  V., , Melnik  V., , and Tsymbal  O., “ Removing impulse bursts from images by training based filtering. ,” EURASIP J. Appl. Signal Process..  1110-8657 2003, (3 ), 223–237  ((2003)).
Rabic  T., “ Robust estimation approach to blind denoising. ,” IEEE Trans. Image Process..  1057-7149 14, (11 ), 1755–1766  (Nov. (2005)).
Huber  P.,  Robust Statistics. ,  Wiley ,  Hoboken, NJ  ((1981)).
Sun  T., , Gabbouj  M., , and Neuvo  Y., “ Center weighted median filters: some properties and their applications in image processing. ,” Signal Process..  0165-1684 35, (3 ), 213–229  ((1994)).
Tsymbal  O. V., , Lukin  V. V., , Ponomarenko  N. N., , Zelensky  A. A., , Egiazarian  K. O., , and Astola  J. T., “ Three-state locally adaptive texture preserving filter for radar and optical image processing. ,” EURASIP J. Appl. Signal Process..  1110-8657 8, , 1185–1204  (May (2005)).
Yaroslavsky  L. P., “ Local criteria and local adaptive filtering in image processing: a retrospective view. ,” in  Proc. Int. Workshop on Local and Non-Local Approximation in Image Processing (LNLA 2008). , CD-ROM,  EURASIP ,  Lausanne, Switzerland  ((2008)).
Kervrann  C., and Boulanger  J., “ Local adaptivity to variable smoothness for exemplar-based image regularization and representation. ,” Int. J. Comput. Vis..  0920-5691 79, (1 ), 45–69  ((2008)).
Perona  P., and Malik  J., “ Scale space and edge detection using anisotropic diffusion. ,” IEEE Trans. Pattern Anal. Mach. Intell..  0162-8828 12, (7 ), 629–639  ((1990)).
Rudin  L., , Osher  S., , and Fatemi  E., “ Nonlinear total variation based noise removal algorithms. ,” Physica D.  0167-2789 60, , 259–268  ((1992)).
Spokoiny  V. G., “ Estimation of a function with discontinuities via local polynomial fit with an adaptive window choice. ,” Ann. Stat..  0090-5364 26, (4 ), 141–170  ((1998)).
Mallat  S.,  A Wavelet Tour of Signal Processing. ,  Academic Press ,  San Diego  ((1998)).
Lukin  V. V., , Oktem  R., , Ponomarenko  N., , and Egiazarian  K., “ Image filtering based on discrete cosine transform. ,” Telecommun. Radio Eng.. 66, (18 ), 1685–1701  ((2007)).
Motwani  M. C., , Gadiya  M. C., , Motwani  R. C., , and Harris  F. C., “ Survey of image denoising techniques. ,” in  Proc. Global Signal Processing Expo and Conf. (GSPx 2004). , CR-ROM,  Global Technology Conferences, Inc. ,  Santa Clara, CA  ((2004)).
Oktem  R., , Egiazarian  K., , Lukin  V., , Ponomarenko  N., , and Tsymbal  O., “ Locally adaptive DCT filtering for signal-dependent noise removal. ,” EURASIP J. Adv. Signal Process.. 2007, , 42472  ((2007)).
Klaine  L., , Vozel  B., , and Chehdi  K., “ An integro differential method for adaptive filtering of additive or multiplicative noise. ,” in  Proc. Int. Conf. on Acoustics, Speech, and Signal Processing (ICASSP 2005). , pp. 1001–1004 ,  IEEE ,  Piscataway, NJ  ((2005)).
Argenti  F., , Torricelli  G., , and Alparone  L., “ Signal dependent noise removal in the undecimated wavelet domain. ,” in  Proc. Int. Conf. on Acoustics, Speech, and Signal Processing (ICASSP 2002). , pp. 3293–3296 ,  IEEE ,  Piscataway, NJ  ((2002)).
Foi  A., “ Pointwise shape-adaptive DCT image filtering and signal-dependent noise estimation. ,” PhD Thesis, Tampere University of Technology, Tampere, Finland (Dec. (2007)).
Oktem  R., “ Transform domain algorithms for image compression and denoising. ,” PhD Thesis, Tampere University of Technology, Tampere, Finland ((2000)).
Sendur  L., and Selesnick  I. W., “ Bivariate shrinkage with local variance estimation. ,” IEEE Signal Process. Lett..  1070-9908 9, (12 ), 438–441  ((2002));
Fevralev  D. V., , Lukin  V. V., , Totsky  A. V., , Egiazarian  K., , and Astola  J., “ Combined bispectrum filtering technique for signal shape estimation with DCT based adaptive filter. ,” in  Proc. of Int. Workshop on Spectral Methods and Multirate Signal Processing (SMMSM 2006). , CD-ROM, pp. 133–140 , TICSP, Florence, Italy ((2006)).
Coifman  R., and Donoho  D. L., “ Translation invariant denoising. ,” in  Wavelets and Statistics. , Antoniadis  A., Ed.,  Springer-Verlag ,  Berlin  ((1995)).
Salomon  D.,  Data Compression: The Complete Reference. ,  Springer ,  New York  ((2007)).
Yaroslavsky  L. P.,  Digital Holography and Digital Image Processing: Principles, Methods, Algorithms. ,  Kluwer ,  Dordrecht  ((2004)).
Huang  H., , Lin  X., , Rahardja  S., , and Yu  R., “ A method for realizing reversible type-IV discrete cosine transform (IntDCT-IV). ,” in  Proc. of 7th Int. Conf. on Signal Processing (ICSP'04). , Vol. 1, , pp. 101–104 ,  IEEE ,  Piscataway, NJ  ((2004)).
Wallace  G., “ JPEG still image compression standard. ,” Commun. ACM.  0001-0782 34, (4 ), 30–44  ((1991)).
Bovik  A.,  Handbook on Image and Video Processing. ,  Academic Press ,  New York  ((2000)).
Ponomarenko  N., , Lukin  V., , Zelensky  A. A., , Astola  J., , and Egiazarian  K., “ Adaptive DCT-based filtering of images corrupted by spatially correlated noise. ,” Proc. SPIE.  0277-786X 6812, , 68120W  ((2008)).
Lukin  V. V., , Fevralev  D. V., , Abramov  S. K., , Peltonen  S., , and Astola  J., “ Adaptive DCT-based 1-D filtering of Poisson and mixed Poisson and impulsive noise. ,” in  Proc. Int. Workshop on Local and Non-Local Approximation in Image Processing (LNLA 2008). , CD-ROM,  EURASIP ,  Lausanne, Switzerland  ((2008)).
Lukin  V. V., , Fevralev  D. V., , Ponomarenko  N. N., , Pogrebnyak  O. B., , Egiazarian  K. O., , and Astola  J. T., “ Local adaptive filtering of images corrupted by nonstationary noise. ,” Proc. SPIE.  0277-786X 7245, , 724506  ((2009)).
Lukin  V. V., , Abramov  S. K., , Zelensky  A. A., , Astola  J., , Vozel  B., , and Chehdi  B., “ Improved minimal interquantile distance method for blind estimation of noise variance in images. ,” Proc. SPIE.  0277-786X 6748, , 67481I  ((2007)).
Lukin  V. V., , Koivisto  P. T., , Ponomarenko  N. N., , Abramov  S. K., , and Astola  J. T., “ Two-stage methods for mixed noise removal. ,” in  Proc. EURASIP Workshop on Nonlinear Signal and Image Processing (NSIP). , CD-ROM,  EURASIP ,  Sapporo, Japan  ((2005)).
Abramov  S., , Lukin  V., , Vozel  B., , Chehdi  K., , and Astola  J., “ Segmentation-based method for blind evaluation of noise variance in images. ,” J. Appl. Remote Sens.. 2, , 023533  (Aug. (2008)).
Ponomarenko  N. N., , Lukin  V. V., , Abramov  S. K., , Egiazarian  K. O., , and Astola  J. T., “ Blind evaluation of additive noise variance in textured images by nonlinear processing of block DCT coefficients. ,” Proc. SPIE.  0277-786X 5014, , 178–189  ((2003)).
Liu  C., , Szeliski  R., , Kang  S. B., , Zitnick  C. L., , and Freeman  W. T., “ Automatic estimation and removal of noise from a single image. ,” IEEE Trans. Pattern Anal. Mach. Intell..  0162-8828 30, (2 ), 299–314  ((2008)).
Portilla  J., , Strela  V., , Wainwright  M., , and Simoncelli  E. P., “ Image denoising using Gaussian scale mixtures in the wavelet domain. ,” IEEE Trans. Image Process..  1057-7149 12, (11 ), 1338–1351  ((2003)).
Ponomarenko  N., , Fevralev  D., , Roenko  A., , Krivenko  S., , Lukin  V., , and Djurovic  I., “ Edge detection and filtering of images corrupted by nonstationary noise using robust statistics. ,”  Proc. 10th Int. Conf. on the Experience of Designing and Application of CAD Systems in Microelectronics (CADSM 2009). ,  Polyana-Svalyava, Ukraine , pp. 129–136 , IEEE ((2009)).
Ponomarenko  N., , Lukin  V., , Djurovic  I., , and Simeunovic  M., “ Pre-filtering of multichannel remote sensing data for agricultural bare soil field parameter estimation. ,”  Proc. Int. Conf. BioSense 2009. , University of Novi Sad,  Novi Sad, Serbia  (Oct. (2009)).
AVIRIS Home page, ⟨http://aviris.jpl.nasa.gov/⟩ (accessed Jan 6, (2007)).
Ponomarenko  N., , Lukin  V., , Zriakhov  M., , and Kaarna  A., “ Preliminary automatic analysis of characteristics of hyperspectral AVIRIS images. ,”  Proc. 11th Int. Conf. on Mathematical Methods in Electromagnetic Theory (MMET 2006). ,  Kharkiv, Ukraine , pp. 158–160  ((2006)).
Christophe  E., , Leger  D., , and Mailhes  C., “ Quality criteria benchmark for hyperspectral imagery. ,” IEEE Trans. Geosci. Remote Sens..  0196-2892 43, (9 ), 2103–2114  (Sept. (2005)).

Grahic Jump LocationImage not available.

Vladimir V. Lukin graduated from Kharkov Aviation Institute (now National Aerospace University) in 1983, with his Diploma with honor in radio engineering. Since then he has been with the Department of Transmitters, Receivers and Signal Processing at National Aerospace University. He defended the thesis of Candidate of Technical Science in 1988 and Doctor of Technical Science in 2002 in DSP for remote sensing. Since 1995, he has been in cooperation with Tampere University of Technology. Currently, he is department vice-chairman and professor. His research interests include digital signal/image processing, remote sensing data processing, image filtering, and compression.

Grahic Jump LocationImage not available.

Dmitriy V. Fevralev graduated from National Aerospace University in 2002 with a Diploma in radio engineering. Since then he has been with the Department of Transmitters, Receivers and Signal Processing at National Aerospace University. He defended his thesis of Candidate of Technical Science in 2008 in DSP for bispectral analysis in radar systems. Currently, he is research fellow and part-time assistant with the Department of Transmitters, Receivers and Signal Processing. His research interests include digital signal/image processing, bispectral analysis of radar signals, and image filtering.

Grahic Jump LocationImage not available.

Nikolay N. Ponomarenko graduated from National Aerospace University in 1995 and received his Diploma with honor in computer science. Since then he has been with the Department of Transmitters, Receivers and Signal Processing at National Aerospace University. He defended his thesis of Candidate of Technical Science in 2004 in DSP for remote sensing. He also defended his thesis of Doctor of Technology at Tampere University of Technology, Finland, in 2005, on image compression. Currently, he is senior researcher and part-time associate professor with the Department of Transmitters, Receivers and Signal Processing. His research interests include digital signal/image processing, remote sensing data processing, image filtering, and compression.

Grahic Jump LocationImage not available.

Sergey K. Abramov graduated from National Aerospace University in 2000 with a Diploma with honor in radio engineering. Since then, he has been with the Department of Transmitters, Receivers and Signal Processing at National Aerospace University. He defended his thesis of Candidate of Technical Science in 2003 in DSP for remote sensing. Currently, he is associate professor and part-time senior researcher with the Department of Transmitters, Receivers and Signal Processing. His research interests include digital signal/image processing, blind estimation of noise characteristics, and image filtering.

Grahic Jump LocationImage not available.

Oleksiy Pogrebnyak received his Ph.D. degree from Kharkov Aviation Institute (now National Aerospace University), Ukraine, in 1991. Currently, he is with The Center for Computing Research of National Polytechnic Institute, Mexico. His research interests include digital signal/image filtering and compression and remote sensing.

Grahic Jump LocationImage not available.

Karen O. Egiazarian received her PhD from Moscow M. V. Lomonosov State University, Russia, in 1986, and Doctor of Technology degree from Tampere University of Technology, Finland, in 1994. He is a leading scientist in signal, image, and video processing, with about 300 refereed journal and conference articles, three book chapters, and a book published by Marcel Dekker. His main interests are in the field of multirate signal processing, image and video denoising and compression, and digital logic. He is a member of the DSP Technical Committee of the IEEE Circuits and Systems Society.

Grahic Jump LocationImage not available.

Jaakko Astola received his BS, MS, Licentiate, and PhD in mathematics from Turku University, Finland, in 1972, 1973, 1975, and 1978, respectively. From 1976 to 1977, he was with the Research Institute for Mathematical Sciences of Kyoto University, Kyoto, Japan. Since 1979 to 1987, he was with Lappeenranta University of Tecvhnology (Finland). From 1987, he is with Tampere University of Technology, Tampere, Finland. Currently, he is professor of signal processing and director of Tampere International Center for Signal Processing, academy professor by the Academy of Finland, and IEEE fellow. His research interests include signal/image processing, statistics, and image coding.

© 2010 SPIE and IS&T

Citation

Vladimir V. Lukin ; Dmitriy V. Fevralev ; Nikolay N. Ponomarenko ; Sergey K. Abramov ; Oleksiy Pogrebnyak, et al.
"Discrete cosine transform–based local adaptive filtering of images corrupted by nonstationary noise", J. Electron. Imaging. 19(2), 023007 (May 03, 2010). ; http://dx.doi.org/10.1117/1.3421973


Figures

Graphic Jump LocationF1 :

Output MSE versus β for two standard test images: (a) Lenna and (b) Baboon, corrupted by additive white Gaussian noise with variances 50 (dotted line) and 100 (dashed line) and by Poisson noise (solid line).

Graphic Jump LocationF2 :

(a) Noise-free test image Barbara, (b) visualized σ̂(n,m) of test image Barbara (values have been multiplied by 7) and (c) σ̂c(n,m) (values have been multiplied by 3.5).

Graphic Jump LocationF3 :

Histograms of (a) σ̂(n,m) and (b) σ̂c(n,m) for the image Barbara, σn2=100.

Graphic Jump LocationF4 :

Magnified ratio image Rvis(n,m)=[40R(n,m)] for the test image Barbara (a) corrupted by additive noise and (b) corrupted by Poisson noise.

Graphic Jump LocationF5 :

Histograms of the ratio R(n,m) for the test images: (a) Barbara corrupted by zero mean pure additive Gaussian noise with σn2=100 and (b) Lenna corrupted by Poisson noise.

Graphic Jump LocationF6 :

Histograms of the ratio R(n,m) for the test images: (a) Lenna and (b) Barbara corrupted by zero mean spatially correlated additive Gaussian noise with σn2=100.

Graphic Jump LocationF7 :

Data registered for (a) HH and (b) VH polarization of radar signal.

Graphic Jump LocationF8 :

Sample histograms for the image in Fig. 7 (a) for small distance fragment 1 and (b) for middle distance fragment 3.

Graphic Jump LocationF9 :

Output images for (a) HH and (b) VH polarization of radar signal.

Graphic Jump LocationF10 :

(a) Original one-channel image of AVIRIS hyperspectral data, (b) the output of LA DCT-1 filter, (c) the ratio Rvis(n,m) image, and (d) the output of the hard switching LA DCT-2 filter.

Tables

Table Grahic Jump Location
Statistical characteristics of the selected “homogeneous” regions for the image in Fig. 7.
Table Grahic Jump Location
Estimates r̂ for the test images corrupted by different types of noise.
Table Grahic Jump Location
Values of MSEout for several βhet and TR for the considered test images corrupted by spatially correlated noise.
Table Grahic Jump Location
PSNR for the compared LA-DCT-2 and Kervrann’s filters.
Table Grahic Jump Location
Values of MSEout for several βhet and TR for the considered test images.
Table Grahic Jump Location
Output MSE for “ideal” DCT-based (IdDCT) filter, LA DCT-1, and nonlinear nonadaptive filters.

References

Shui  P.-L., “ Image denoising algorithm via doubly local wiener filtering with directional windows in wavelet domain. ,” IEEE Signal Process. Lett..  1070-9908 12, (10 ), 681–684  (Oct. (2005)).
Foi  A., , Katkovnik  V., , and Egiazarian  K., “ Signal-dependent noise removal in point-wise shape-adaptive DCT domain with locally adaptive variance. ,” in  Proc. of European Signal Processing Conf. (EUSIPCO 2007). , CD-ROM,  EURASIP ,  Poznan, Poland  ((2007)).
Solbo  S., and Eltoft  T., “ Homomorphic wavelet-based statistical despeckling of SAR images. ,” IEEE Trans. Geosci. Remote Sens..  0196-2892 42, (4 ), 711–721  ((2004)).
Eng  H.-L., and Ma  K.-K., “ Noise adaptive soft switching median filter. ,” IEEE Trans. Image Process..  1057-7149 10, (2 ), 242–251  ((2001)).
Barner  K. and Arce  G., Eds.,  Nonlinear Signal and Image Processing: Theory, Methods, and Applications. ,  CRC Press ,  Boca Raton  ((2003)).
Plataniotis  K. N., and Venetsanopoulos  A. N.,  Color Image Processing and Applications. ,  Springer-Verlag ,  New York  ((2000)).
Melnik  V., “ Nonlinear locally adaptive techniques for image filtering and restoration in mixed noise environments. ,” PhD Thesis, Tampere University of Technology, Tampere, Finland ((2000)), ⟨http://www.atilim.edu.tr/~roktem/Research/interests.htm⟩.
Lukin  V. V., , Ponomarenko  N. N., , Abramov  S. K., , Vozel  B., , Chehdi  K., , and Astola  J., “ Filtering of radar images based on blind evaluation of noise characteristics. ,” Proc. SPIE.  0277-786X 7109, , 71090R  (Sept. (2008)).
Astola  J., and Kuosmanen  P.,  Fundamentals of Nonlinear Digital Filtering. ,  CRC Press ,  Boca Raton, FL  ((1997)).
Touzi  R., “ A review of speckle filtering in the context of estimation theory. ,” IEEE Trans. Geosci. Remote Sens..  0196-2892 40, (11 ), 2392–2404  ((2002)).
Barton  D. K.,  Radar System Analysis and Modeling. ,  Artech House ,  Boston  ((2005)).
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., , and Tarnavsky  Y. V.,  Methods and Algorithms for Pre-processing and Classification of Multichannel Radar Remote Sensing Images. ,  TTY Monistamo ,  Tampere, Finland  (Dec. (2004)).
Lopez-Martinez  C., and Pottier  E., “ On the extension of multidimensional speckle noise model from single-look to multilook SAR imagery. ,” IEEE Trans. Geosci. Remote Sens..  0196-2892 45, (2 ), 305–320  ((2007)).
Chang  C.-I., Ed.,  Hyperspectral Data Exploitation: Theory and Applications. ,  Wiley ,  Hoboken, NJ  ((2007)).
Lukac  R., , Plataniotis  K. N., , Smolka  B., , and Venetsanopoulos  A. N., “ cDNA microarray image processing using fuzzy vector filtering framework. ,” J. Fuzzy Sets Syst.. 152, (1 ), 17–35  (May (2005)).
Kulemin  G. P.,  Millimeter-Wave Radar Targets and Clutter. ,  Artech House ,  Boston  ((2003)).
Oliver  C., and Quegan  S.,  Understanding Synthetic Aperture Radar Images. ,  SciTech Publishing ,  Raleigh, NC  ((2004)).
Lee  J. S., , Wen  J. H., , Ainsworth  T. I., , Chen  K. S., , and Chen  A. J., “ Improved sigma filter for speckle filtering of SAR imagery. ,” IEEE Trans. Geosci. Remote Sens..  0196-2892 47, (1 ), 202–213  (Jan. (2009)).
Barducci  A., , Guzzi  D., , Marcoionni  P., , and Pippi  I., “ CHRIS-Proba performance evaluation: signal-to-noise ratio, instrument efficiency and data quality from acquisitions over San Rossore (Italy) test site. ,” in  Proc. 3rd ESA CHRIS/Proba Workshop. , CD-ROM,  ESA-ESRIN ,  Frascati, Italy  ((2005)).