A Mar III

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Local Adaptive Filtering of Images Corrupted by Nonstationary Noise

Vladimir V. Lukina, Dmitriy V. Fevraleva, Nikolay N. Ponomarenkoa, Oleksiy B. Pogrebnyakb, Karen O. Egiazarianc, Jaakko T. Astolac National Aerospace University, 61070, Kharkov, Ukraine; b Instituto Politecnico Nacional, Mexico DF, Mexico; c Tampere University of Technology, FIN 33101, Tampere, Finland
ABSTRACT
In various practical situations of remote sensing image processing it is assumed that noise is nonstationary and no a priory information on noise dependence on local mean or about local properties of noise statistics is available. It is shown that in such situations it is difficult to find a proper filter for effective image processing, i.e., for noise removal with simultaneous edge/detail preservation. To deal with such images, a local adaptive filter based on discrete cosine transform in overlapping blocks is proposed. A threshold is set locally based on a noise standard deviation estimate obtained for each block. Several other operations to improve performance of the locally adaptive filter are proposed and studied. The designed filter effectiveness is demonstrated for simulated data as well as for real life radar remote sensing and marine polarimetric radar images. Keywords: non-stationary noise, locally adaptive filtering, remote sensing.
a

1. INTRODUCTION
It is a quite typical situation in remote sensing, polarimetric radar systems1-3, hyperspectral imaging4, ultrasound medical diagnostics5, genomic imaging6, etc. that an image is noisy and it has to be pre-filtered before further processing. In order to perform filtering in an efficient way, it is desirable to have a preliminary knowledge on (or a prediction of) noise type and statistics7-9. However, such a priori information is not always available. For some types of imaging systems like synthetic aperture radar the noise type and characteristics like for fully developed speckle can be known in advance10,11 and they can be quite stable for entire image. But for many other imaging systems noise characteristics can vary from one observation (image data registration session) to another or from one sub-band of hyperspectral image to another12. Moreover, noise statistics and properties could be quite dissimilar in different fragments of a particular image. Such dissimilarities in noise statistics can occur due to several reasons. One possibility is changing of imaging conditions, for example, illumination or different distance to particular parts of an observed scene like in side look radar imaging9. The noise dissimilarities could be also attributed to uncontrollable influence of several different sources of noise12. For instance, in marine radars internal noise is mixed with interference due to sea clutter that depends on wind speed, direction of sea waves, incidence angle13, etc. Quite many imaging and signal processing systems employ nonlinear amplification regulations in input circuits before image digitalization. Due to unknown properties of such regulations, noise in output images possesses statistical properties that can be hardly predicted even if input noise statistics is known. A good example of such systems could be ultrasound medical devices. For some imaging systems, adequate models of noise present in the formed images have not been established yet. In such cases of limited information about noise statistics one way to follow could be applying nonlinear non-adaptive filters that are robust in the sense that they do not take into account the noise statistics. Well-known examples of such
a

Correspondence to Lukin V.V.: e-mail lukin@xai.kharkov.ua tel./fax +38 057 3151186


Image Processing: Algorithms and Systems VII, edited by Jaakko T. Astola, Karen O. Egiazarian Nasser M. Nasrabadi, Syed A. Rizvi, Proc. of SPIE-IS&T Electronic Imaging, SPIE Vol. 7245, 724506 2009 SPIE-IS&T CCC code: 0277-786X/09/$18 doi: 10.1117/12.805298 SPIE-IS&T/ Vol. 7245 724506-1

filters are standard median, -trimmed mean, central weighted median, Wilcoxon filters, etc8. However, alongside with noise reduction, some of these filters can severely distort useful information contained in images8. Other filters such as center weighted median ones are able to preserve edges and details but they do not provide desired noise reduction. Thus, to get around aforementioned shortcomings some other techniques are to be designed. Many other filters that have demonstrated better performance than nonlinear non-adaptive filters require a priori knowledge of noise type and statistical characteristics. Among such filters there are wavelet and discrete cosine transform (DCT) based filters for additive noise case15, 16, different filters designed for coping with speckle noise17, 18, and other noise types19-21. To determine noise type from an image at hand, a method22 has been designed. However, this method provides discrimination of noise environments if one deals with a certain dependence of noise variance on local mean where it is supposed that this dependence holds for entire image. For such stationary dependence, methods for blind determination of statistical parameters of additive and multiplicative or additive and Poissonian noise have been developed recently 8,21. The corresponding methods of DCT based filtering have been proposed. Moreover, it has been shown that since DCT filtering is applied locally, in blocks of fixed or adaptively selected shape, it can be easily adapted to a priori known or pre-estimated noise statistics. This is a useful property of DCT based filtering for which in each block noise statistics can be assumed stationary. In opposite to noise characteristics considered in the papers8,21, here we deal with the situation for which aforementioned stationarity of noise dependence on local mean is not observed. For example, noise variance can change along rows or columns of an image, or both (examples will be given below for real life images). At the same time, we still assume local stationarity of noise meaning that, if a given image block is not too large, noise statistics within this block remains practically the same. Recently such assumption has been put into basis of designing a DCT based locally adaptive filter applied for processing of real and imaginary components of bispectrum23 for solving a problem of unknown signal shape (waveform) reconstruction. Recall that bispectrum is a 2-D complex function and in its real and imaginary components a strongly nonstationary noise is present. Thus, our idea is that the filter designed in the paper23 and its modifications can be also effectively applied to process different kinds of images for which non-stationary noise with a priori unknown local statistical properties is present. The paper is organized as following. In Section 2, we give a practical example of images for which the aforementioned properties of noise, i.e., its spatial non-stationarity and unpredictable change, hold. Section 3 deals with description of a locally adaptive DCT based filter and its preliminary testing for simulated images and noise. Examples of the designed filter applying to real life images are presented in Section 4. One drawback of the designed version is discussed and its modification intended on further improving the filters properties is proposed and tested. Finally, we draw conclusions and present possible directions of future work.

2. AN EXAMPLE OF SPATIALLY NON-STATIONARY NOISE FOR REAL LIFE DATA


To clarify basic properties of noise for considered situation let us give an example. A polarimetric radar data (called Rock, presented as real valued data arrays) have been obtained by a polarimetric coastal radar (these data have been provided by Dr. A.V. Popov, Dept 502 of National Aerospace University, Kharkov, Ukraine). The images (Fig. 1) mainly correspond to sea surface (basic part of images), one large rock (small rocky island placed in the image central part, very bright pixels) and several small ones (left lower corner, groups of bright pixels), and a shadowed zone behind the large rock. The horizontal axis corresponds to range and vertical axis relates to azimuth of the radar. Visual analysis of these images allows expecting that background intensity varies depending upon range. This has been partly confirmed by histograms (Fig.2) obtained for manually selected six homogeneous regions (marked by frames in images in Fig.1) for two different sectors and three different mean distances. As seen, the histogram shape changes with range. The distributions are asymmetric with respect to their means and there is a heavy tail to the right side from the distribution mode. Distribution modes are also different for different ranges. Obvious outliers are seldom (occur with quite small probabilities), but in the considered case their values differ from the mode values by approximately ten or more times.

SPIE-IS&T/ Vol. 7245 724506-2

For each image homogeneous region we have determined the following parameters: minimal, maximal and mean mk values for each k-th fragment, k=1,,6; 2 variance k2 and relative variance rel k calculated as variance divided by squared mean; 2 inter-quantile estimate24 of relative variance iq k ; this estimate is practically insensitive to the presence of out-

liers, trend and other heterogeneities in data.

b Fig. 1. Data registered for HH (a) and VH (b) polarization of radar signal
700

500 450

600

400
500

350 300 250

400

300

200 150 100

200

100

50
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

0.5

1.5

2.5

3.5

4.5

a b Fig. 2. Sample histograms for the image in Fig. 1,b for small distance (a) and for middle distance (b) regions

SPIE-IS&T/ Vol. 7245 724506-3

The values mk increased by about 15 times with range increasing (for farer regions in comparison to closer regions), variances k2 increased by almost 20 times whilst relative variances and their inter-quantile estimates reduced by 1.41.6 times. This shows that noise is spatially non-stationary with dominant signal-dependent component. This 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 surface13, different mutual geometry of sea wave direction and radar azimuth, etc. And it is difficult to separate the influence of these factors. Besides, here we deal with atypical situation when noise variance has the tendency to increase when one moves along image rows, i.e., with growing distance to a sensed surface fragment. Note that if there are outliers (impulse noise) in images, it is commonly reasonable to first remove them and then to apply other filtering algorithms25. If impulse noise is not too intensive, the task of its preliminary removal can be effectively solved by applying the center weighted median filter26 with the 3x3 scanning window and the center element weight Wc=5. Keeping this in mind, below we will assume that impulse noise is absent or preliminarily removed at initial stages of image processing.

3. LOCAL ADAPTIVE DCT FILTER AND ITS PERFORMANCE ANALYSIS FOR SIMULATED IMAGES
For the very beginning, let us recall an operation principle of DCT based filters. In general, DCT based filtering (denoising) is carried out as follows16,18,27,28. For a given block, commonly of size 8x8, DCT is performed and one obtains spectral coefficients D(k , l , n, m) where indices k,l relate to DCT (spectral) coefficients and n,m denote a position (coordinates) of the left upper corner of an image block. The next step is to apply some kind of thresholding. One can use hard or soft thresholding. In the former variant (that we have exploited), the values D(k , l , n, m) whose absolute values are smaller than a local threshold T (n, m) are assigned zero values; other coefficients are remained unchanged. This way one gets DT (k , l , n, m) to which inverse DCT is applied. Thus, filtered values are obtained for all pixels that belong to a given block. This is a principal difference of DCT based denoising compared to conventional scanning window filters. Blocks can overlap (tessellate) or not. In the former case, one obtains several estimates of filtered image for each pixel. A simplest way to aggregate them is to average all estimates. The use of overlapping blocks improves DCT based filter performance16,27 similarly to spatially invariant wavelet denoising. In comparison to DCT based filtering performed in non-overlapping blocks, PSNR gain reaching 23 dB can be provided by DCT based processing in fully overlapping blocks16,27. Certainly, this requires more computations. DCT based filtering with fully overlapping blocks provides performance (in terms of MSE or PSNR) better or comparable to the best wavelet based denoising techniques16,18 for pure additive and pure multiplicative noise cases. But, probably, the main advantage of DCT based filtering is that, being carried out in block-wise or local manner, it can be easily adapted to statistical and spatial properties of noise. Commonly this is done by proper adjusting a threshold T (n, m) . As an example, consider the case of film grain noise for which a general observation model can be expressed as
tr tr I ij = I ij + ( I ij ) nij ,

(1)

tr where I ij , I ij and nij denote the noisy image sample (pixel) value, true image value, and signal independent noise com2 ponent that is characterized by the variance n , respectively, for the ij-th sample, is a parameter of film grain noise. Then for an nm-th block it is proposed to set the threshold as

T (n, m) = k n ( I (n, m)) ,

(2)

SPIE-IS&T/ Vol. 7245 724506-4

where I (n, m) is the estimate of the local mean for the block and k denotes the DCT based filter parameter commonly set equal to 2.6. As seen from (2), the product ( I (n, m)) , in fact, is an estimate of noise standard deviation (n, m)
n

in a given block. For =0, i.e., for pure additive noise one gets exactly n and the threshold is fixed and equal to k n . Similarly, for pure multiplicative case (n, m) = ( I (n, m))2 where in this case is the multiplicative noise standard
n n

deviation. As I (n, m) it is possible to use the DCT spectral coefficient D(0, 0, n, m) although there are also alternative variants like using a block sample median as I (n, m) . As seen, for setting a local threshold one needs some local estimate of noise variance or standard deviation (n, m) . In 30 31,32 general, (n, m) can be estimated in spatial or in spectral (orthogonal transform) domain . Conventional estimate of local variance30 is sensitive to image content in a block or in a scanning window. This property is useful in hardswitching locally adaptive filter design14,30 but is a drawback for the considered situation of unknown noise variance. The reason will be considered below. Because of this, let us first concentrate on estimation of noise local standard devia tion in spectral domain. For a given block we propose to estimate (n, m) as
(n, m) = 1.483med { D (k , l , m, n) , k = 0,..., 7; l = 0,..., 7 except k = 0 l = 0} , (3)

where med{} denotes a median for a data sample. This estimate is similar to median of absolute deviations (MAD) used as a robust data scale estimator8 but here, in (3) noise standard deviation (scale) estimate (n, m) is obtained in spectral domain. Although the used estimator is robust, image local content (image details, texture, edges) present in a given block can influence an estimate (n, m) . Besides, this estimate is obtained in blocks of rather small size (8x8 pixels) and this limited size of data size influence estimation accuracy even for blocks that belong to image homogeneous regions. Let us consider these aspects more in detail for a particular example. Assume that a test image, e.g., the standard test image Barbara is corrupted by zero mean additive noise with a constant variance 2 = 100 . The noisy image is shown in Fig. 3,a. The values (n, m) are visualized in Fig. 3,b (larger values are lighter). As seen, the values (n, m) for blocks that correspond to image homogeneous regions vary a little indicat ing the noisiness of the estimate (n, m) . But they are mainly rather small (darker) and the values (n, m) for the blocks that correspond to image heterogeneities like edges, textures and details are commonly slightly larger due to influence of image content. After such a brief analysis of the behavior of (n, m) , let us carry several experiments with different types of noise. If one knows that noise is additive with fixed 2 , it is possible to apply the standard DCT based filter16 with fixed T (n, m) = 2.6 . Such a filter produces an image with PSNR=34.4 dB or, equivalently, output MSE
tr MSEout = (1/( IJ 1)) ( I ijf I ij ) 2 = 24.0 where I and J denote an image size, I ijf is the jj-th pixel value for the i =1 j =1 I J

processed (filtered) image. Assume now that we do not know the noise type and variance and let us apply the proposed DCT filter with adaptively set threshold T (n, m) = 2.6 (n, m) . It produces PSNR=32.8 dB and MSEout 34.0 . The main reason why this time PSNR is smaller (and, respectively, MSE is larger) is the following. It is known14, that the DCT filter with fixed T (n, m) = 2.6 is not very good in the sense of edge/detail preservation and if a threshold increases their edge/detail preservation ability becomes poorer. Besides, for texture preservation the best choice for the threshold14 is not T (n, m) = 2.6 but T (n, m) 2.2 . Since for edge/detail/texture neighborhoods the estimates (n, m) are larger than , then actually the adaptive filter uses T (n, m) 3 for the corresponding blocks and this additionally worsens edge/detail/texture preservation and, thus, leads to reduction of PSNR.

SPIE-IS&T/ Vol. 7245 724506-5

a b Fig. 3. Noisy test image Barbara (a), visualized ( n, m) (its values have been multiplied by 7) (b)

Consider now a more complicated example. Let this time we have the standard test image Lena corrupted by Poisson tr noise for which noise variance for a given ij-th pixel is equal to I ij . The input MSE is equal to
tr MSEinp = (1/( IJ 1)) ( I ij I ij ) 2 = 123.9 . The optimal way of DCT based filtering of an image in this case is seti =1 j =1 I J

ting T (n, m) = k I (n, m) . This filter produces MSEout = 21.0 . In turn, the adaptive DCT based filter that exploits the threshold setting as T (n, m) = 2.6 (n, m) produces MSEout = 24.9 , i.e. only slightly larger. These two examples show that, in general, performance of the proposed locally adaptive DCT filter is good enough but its edge/detail/texture preservation ability is worth improving.

4. EXAMPLES OF APPLICATION TO REAL LIFE IMAGES AND MODIFICATIONS


First of all, let us give two examples of applying the adaptive DCT based filter described above to real life data. Fig. 4 presents the output images for the polarimetric radar data presented in Fig. 1. Note that in this case a two stage procedure has been used. At the first stage, we have applied the center weighted median filter 26 with the 3x3 scanning window and the center element weight Wc=5. This has been done in order to remove outliers (impulse noise). At the second stage, the adaptive DCT based filter has been used. As seen from comparison of images in Figures 1 and 4, noise is well suppressed whilst useful information in images is preserved well. Consider now two sub-band images formed by the hyperspectral imager AVIRIS33. Signal-to-noise ratio in different subbands of AVIRIS hyperspectral data as well as in other hyperspectral data varies in wide limits12,34. Due to this, in some sub-band images of hyperspectral data noise is well seen and it is desirable to suppress it. However, there are no established and backgrounded models of noise in hyperspectral images35. The only valid assumptions are that there are additive and signal-dependent components. Thus, this is just the case to apply the designed locally adaptive DCT based filter. The original noisy image in the 108-th sub-band of the image Moffett Filed is depicted in Fig. 5,a. The processed image is demonstrated in Fig. 5,b. Essential suppression of noise is observed whilst useful information is preserved well enough.

SPIE-IS&T/ Vol. 7245 724506-6

b Fig. 4. Output images for HH (a) and VH (b) polarization of radar signal

As it was mentioned in the previous Section, the edge/detail preservation of the proposed filter is worth additional improving. Let us analyze behavior of conventional estimate of local variance14,30. For a given ij-th pixel, it is commonly

2 calculated as c ij = (1/(QS 1))

i + ( Q 1) / 2

q = i ( Q 1) / 2 s = j ( S 1) / 2

j + ( S 1) / 2

( I qs I ij ) 2 where I ij is the local mean for the scanning window of

size Q x S centered on the ij-th pixel. Since here we deal with blocks, then similarly an estimate c2nm can be calculated for an nm-th block. The only difference is that commonly Q and S are odd14,30 and for blocks used in DCT based filtering the expression slightly changes and can be rewritten as

c2 (n, m) = (1/(QS 1))

n + Q 1 m + S 1 q =n


m+S

s=m

( I qs I qs )2 , Q = S = 8 .

(4)

There is also slightly another way to compute local variance estimate by using smaller block size

c2 (n, m) = (1/(QS 1))

n+Q

q = n +1 s = m +1

( I qs I qs ) 2 , Q = S = 6 .

(5)

The obtained map of the estimates c (n, m) for Q=S=6 is visualized in Fig. 6,a for the test image Barbara corrupted by zero mean additive noise with a constant variance 2 = 100 . Comparing the visualized data in Figures 3,b and 6,a, it is seen that the estimate c (n, m) is much more sensitive to heterogeneities of image present in a block than the estimate (n, m) . This property (difference) can be exploited.

SPIE-IS&T/ Vol. 7245 724506-7

a b Fig. 5. The original image in the 108-th sub-band of the AVIRIS data Moffett Field (a) and the obtained output image (b)

Consider the ratio R(n, m) = c (n, m) / (n, m) . Its example for the same test image and 2 = 100 is visualized in Fig. 6,b (in fact, we represent the image obtained as Rvis (n, m) = [40 R (n, m)], n = 1,..., I 7, m = 1,..., J 7 where [] means rounding to the nearest positive integer not larger than 255). Joint visual analysis of the test image Barbara (Fig. 3,a) and the ratio image (Fig. 6,b) shows that quite large ratios (essentially larger than 1 indicated by brighter color pixels) are observed for blocks located in the image heterogeneous regions.

r
a

Fig. 6. The visualized map of the estimates c ( n, m) (multiplied by 3.5) for Q=S=6 (a) and the ratio image Rvis (n, m) = [40 R(n, m)], n = 1,..., I 7, m = 1,..., J 7 (b) There are several ways to exploit the aforementioned property. The simplest one considered below is to adaptively set the threshold as

2.6 (n, m), if R(n, m) < TR T ( n, m ) = . kdet (n, m), if R(n, m) TR

(6)

where TR is a preset threshold of the ratio R(n,m), kdet < 2.6 is a factor that determines the hard threshold for DCT coefficients in edge/detail neighborhoods and in textural regions. In fact, the DCT based filter described by (3), (5), (6)

SPIE-IS&T/ Vol. 7245 724506-8

can be referred to the class of locally adaptive hard-switching filters14,29,30. At the same time, DCT filters themselves can be considered as locally adaptive ones that adapt to image content in blocks by thresholding DCT coefficients and simultaneously to noise by setting a proper threshold. Thus, a task is to set the parameters TR and kdet of the modified method in an optimal or some intuitive manner. To solve a first part of the task, consider distributions of the ratio R(n,m) which is a random variable. If the estimates c (n, m) and (n, m) are obtained in image homogeneous regions, 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 neighbor hood of unity. Besides, as it follows from ratio map in Fig. 6,b, c (n, m) is commonly larger than (n, m) in image heterogeneous regions. Therefore, the distribution of the ratio R(n,m) can have a heavy right-hand tail. These assumptions have been tested for the already used test images and noise models. The histogram of the obtained values R(n,m) for the test image Barbara (pure additive Gaussian noise, 2 = 100 ) is presented in Fig. 7,a. Analogous histogram for the image Lena corrupted by Poisson noise is given in Fig. 7,b. As seen, both distributions really have modes in the neighborhoods of unity and they possess heavy right-hand tail. It can be supposed that by setting TR 1.35 one is able to discriminate blocks that correspond to homogeneous and heterogeneous regions (similar procedure has been used in hard switching locally adaptive filter design30). The simulation data presented below in Table 1 for the test image Barbara corrupted by additive Gaussian noise (several values of the threshold have been tested and the results obtained for three of them are presented) show that TR = 1.35 is a good choice.
7 6 5 4 3 2 1 0 0 2 4 6 8
6 4 2 0 0 2 4 6 8

x 10

12 10 8

x 10

Fig. 7. Histograms of the ratio R(n,m) for the test images: Barbara corrupted by zero mean pure additive Gaussian noise with

2 = 100 (a) and Lena corrupted by Poisson noise (b)


The second question deals with setting kdet . For answering it, we have carried out simulations for the same test images and noise models. The obtained results are presented in Table 1. We present the values of MSEout for several kdet . As it follows from analysis of the obtained results, it is reasonable to set kdet 1.7 that produces the neighborhood of minimal

MSEout for both test images and noise models. Reduction of MSEout by approximately 1.4 times in comparison to the version of the locally adaptive DCT based filter (see Section 3) is achieved for the test image Barbara. Improvement for the test image Lena is smaller. Anyway, the values of MSEout are in both case almost the same as were observed for the DCT filter properly adjusted to given noise models.

SPIE-IS&T/ Vol. 7245 724506-9

Table 1. The values of MSEout for several kdet for the test images kdet Image Barbara TR =1.35 Image Barbara TR =1.32 Image Barbara TR =1.38 Image Lena TR =1.35 2.6 34.0 34.05 33.80 24.26 2.3 30.28 30.72 30.76 23.05 2.0 27.89 28.20 28.47 22.09 1.7 26.55 26.84 27.13 21.54 1.4 26.64 26.98 27.16 21.70 1.1 28.71 29.20 29.04 22.86

Fig. 8. Original one-channel image of AVIRIS hyperspectral data (a), the output of DCT filter (b), the ratio image (c) and the output of the hard switching DCT based filter (d) Let us give also one more real life image processing example. The 224-th sub-band image of the Lunar Lake AVIRIS image is presented in Fig. 8,a. Noise present in this image is well seen. The output image for the DCT based filter de-

SPIE-IS&T/ Vol. 7245 724506-10

scribed in Section 3 is shown in Fig. 8,b. Noise is well suppressed although sharp edges are slightly smeared. The ratio image Rvis (n, m) = [40 R (n, m)], n = 1,..., I 7, m = 1,..., J 7 is visualized in Fig. 8,c. As seen, the most sharp edges and details are marked by lighter color pixels in this map. Due to this, they are better preserved in the output image obtained by the locally adaptive DCT based filter described above in this Section (see Fig. 8,d, TR =1.35, kdet = 1.7 ).

5. CONCLUSIONS AND FUTURE WORK


Two DCT based filtering techniques for suppressing non-stationary noise have been proposed and studied. It has been demonstrated that the first version based on local estimation of noise variance in blocks and setting the corresponding threshold proportionally to the obtained estimate of noise standard deviation performs rather well. However, it can be further improved by means of local (in-block) analysis the ratio R(n,m) and adjusting two thresholds depending upon image content in a block (decision does a given correspond to a homogeneous or heterogeneous region). This allows decreasing MSEout and improving edge/detail/texture preservation in processed images. The recommendations concerning selection of the parameters kdet and TR for the proposed modified technique are given. The designed filters have been applied to real life images and have demonstrated good results. Our future work will deal with applying the designed filters to multichannel (multispectral and hyperspectral) image filtering and compression.

REFERENCES
1. 2. Barton D.K., Radar System Analysis and Modeling, Artech House, USA, 2005. G.P. Kulemin, A.A. Zelensky, J.T. Astola, V.V. Lukin, K.O. Egiazarian, A.A. Kurekin, N.N. Ponomarenko, S.K. Abramov, O.V. Tsymbal, Y.A. Goroshko, and Y.V. Tarnavsky, 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. C. Lopez-Martinez, E. Pottier, On the Extension of Multidimensional Speckle Noise Model from Single-Look to Multilook SAR Imagery, IEEE Transactions on Geoscience and Remote Sensing, Vol. 45, No 2, 2007, pp. 305320. Hyperspectral Data Exploitation: Theory and Applications, Edited by Chein-I Chang, Wiley-Interscience, 2007. Lukin V., Ponomarenko N., Bunaeva I., M. A. Matar, Post-processing of multilook and sequentially formed images in radar and ultrasonic coherent systems, CD-ROM Proceedings of the 46th IEEE International MidWest Symp. on Circuits and Systems, Cairo, Egypt, Dec. 2003 - Jan. 2004, 7 p. R. Lukac, K.N. Plataniotis, B. Smolka, and A.N. Venetsanopoulos, cDNA Microarray Image Processing Using Fuzzy Vector Filtering Framework, Journal of Fuzzy Sets and Systems: Special Issue on Fuzzy Sets and Systems in Bioinformatics, 152, no. 1, pp. 17-35, May 2005. Nonlinear Signal and Image Processing: Theory, Methods, and Applications (Electrical Engineering & Applied Signal Processing Series), Ed. by K. Barner and G. Arce, CRC Press, 2003, 560 p. J. Astola, P. Kuosmanen, Fundamentals of nonlinear digital filtering, CRC Press LLC, Boca Raton, USA, 1997. V.V. Lukin, N.N. Ponomarenko, S.K. Abramov, B. Vozel, K. Chehdi, J. Astola, Filtering of radar images based on blind evaluation of noise characteristics, Proceedings of Image and Signal Processing for Remote Sensing XIV, Cardiff, UK, Sept 2008, SPIE Vol. 7109, 12 p. C. Oliver, S. Quegan, Understanding Synthetic Aperture Radar Images, SciTech Publishing, 2004. Touzi R., A Review of Speckle Filtering in the Context of Estimation Theory, IEEE Transactions on Geoscience and Remote Sensing, Vol. 40, No 11, pp. 2392-2404, 2002. 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., March 2005. Kulemin G.P., Millimeter-Wave Radar Targets and Clutter, Artech House, London-Boston, 342 p., 2003. Tsymbal O.V., Lukin V.V., Ponomarenko N.N., Zelensky A.A., Egiazarian K.O., Astola J.T., Three-state Locally Adaptive Texture Preserving Filter for Radar and Optical Image Processing, EURASIP Journal on Applied Signal Processing, No 8, pp. 1185-1204, May 2005. S. Mallat, A Wavelet tour of signal processing, Academic Press, San Diego, 1998.

3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15.

SPIE-IS&T/ Vol. 7245 724506-11

16. V.V. Lukin, R. Oktem, N. Ponomarenko, K. Egiazarian, Image filtering based on discrete cosine transform, Telecommunications and Radio Engineering, 66, No 18, pp. 1685-1701, 2007. 17. M.C. Motwani, M.C. Gadiya, R.C. Motwani, F. C. Harris, Survey of Image Denoising Techniques, Proceedings of GSP 2004, Santa Clara, CA, September 27-30, 2004. 18. 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, Article ID 42472, 10 p., 2007. 19. L. Klaine, B. Vozel, K. Chehdi, An Integro Differential Method for Adaptive Filtering of Additive or Multiplicative Noise, Proc. of ICASSP, pp. 1001-1004, 2005. 20. F. Argenti, G. Torricelli, L. Alparone, Signal dependent noise removal in the undecimated wavelet domain, Proc. ICASSP, pp. 3293-3296, 2002. 21. A. Foi, Pointwise Shape-Adaptive DCT Image Filtering and Signal-Dependent Noise Estimation: Thesis for the degree of Doctor of Technology, Tampere University of Technology, Tampere, Finland, Dec. 2007. 22. Carton-Vandecandelaere, M.-P., Vozel, B., Klaine, L., Chehdi, K., Application to Multispectral Images of a Blind Identification System for Blur, Additive, Multiplicative and Impulse Noises. Proceedings of EUSIPCO, V. III (2002) 283-286. 23. Fevralev D.V., Lukin V.V., Totsky A.V., Egiazarian K., Astola J., Combined bispectrum filtering technique for signal shape estimation with DCT based adaptive filter, Proceedings of the 2006 International TICSP Workshop on Spectral Methods and Multirate Signal Processing, SMMSP2006, Florence, Italy, Sept 2006, pp. 133-140. 24. S. Abramov, V. Lukin, A. Zelensky, J. Astola, Use of inter-quantile distance estimation in image processing, Proc. SPIE Conf. on Mathematics of Data/Image Pattern Recognition, Compression, and Encryption with Applications IX, SPIE Vol. 6315, San Diego, USA, 2006, 12 p. 25. Lukin V.V., Koivisto P.T., Ponomarenko N.N., Abramov S.K., Astola J.T., Two-stage Methods for Mixed Noise Removal, CD-ROM Proceedings of EURASIP Workshop on Nonlinear Signal and Image Processing (NSIP), Japan, 6 p., May 2005. 26. Sun T., Gabbouj M., Neuvo Y. Center weighted median filters: some properties and their applications in image processing, Signal Processing, Vol. 35, No 3, pp. 213-229, 1994. 27. 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. 28. N.N. Ponomarenko, V.V. Lukin, A.A. Zelensky, P.T. Koivisto, K.O. Egiazarian, 3D DCT Based Filtering of Color and Multichannel Images, Telecommunications and Radio Engineering, No 15, pp. 1369-1392, 2008. 29. V. V. Lukin, D. V. Fevralev, S. K. Abramov, S. Peltonen, J. Astola, Adaptive DCT-based 1-D filtering of Poisson and mixed Poisson and impulsive noise, CD ROM Proceedings of LNLA, Switzerland, 8 p., August 2008. 30. Melnik V.P., Lukin V.V., Zelensky A.A., Astola J.T., Kuosmanen P., Local Activity Indicators: Analysis and Application to Hard-Switching Adaptive Filtering of Images, Optical Engineering Journal, 40, No 8, pp. 1441-1455, 2001. 31. L. Sendur, I.W. Selesnick, Bivariate shrinkage with local variance estimation, IEEE Signal Processing Letters, 9, No 12, pp. 438-441, 2002. 32. N.N. Ponomarenko, V.V. Lukin, S.K. Abramov, K.O. Egiazarian, J.T. Astola, 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, SPIE 5014, Santa Clara (CA, USA), pp. 178-189, 2003. 33. AVIRIS Home page, http://aviris.jpl.nasa.gov/, Accessed Jan 6, 2007. 34. N. Ponomarenko, V. Lukin, M. Zriakhov, A. Kaarna, Preliminary automatic analysis of characteristics of hyperspectral AVIRIS images, Proceedings of MMET, Ukraine, pp. 158-160, 2006. 35. E. Christophe, D. Leger, and C. Mailhes, Quality criteria benchmark for hyperspectral imagery, IEEE Transactions on Geoscience and Remote Sensing, 43, No 9, pp. 2103-2114, Sept 2005.

SPIE-IS&T/ Vol. 7245 724506-12

You might also like