- Research Article
- Open Access
A Level Set Filter for Speckle Reduction in SAR Images
© Hongga Li et al. 2010
- Received: 11 June 2009
- Accepted: 12 February 2010
- Published: 30 March 2010
Despite much effort and significant progress in recent years, speckle removal for Synthetic Aperture Radar (SAR) image still is a challenging problem in image processing. Unlike the traditional noise filters, which are mainly based on local neighborhood statistical average or frequencies transform, in this paper, we propose a speckle reduction method based on the theory of level set, one form of curvature flow propagation. Firstly, based on partial differential equation, the Lee filter can be cast as a formulation of anisotropic diffusion function; furthermore, we continued to deduce it into a level set formulation. Level set flow into the method allows the front interface to propagate naturally with topological changes, where the speed is proportional to the curvature of the intensity contours in an image. Hence, small speckle will disappear quickly, while large scale interfaces will be slow to evolve. Secondly, for preserving finer detailed structures in images when smoothing the speckle, the evolution is switched between minimum or maximum curvature speed depending on the scale of speckle. The proposed method has been illustrated by experiments on simulation image and ERS-2 SAR images under different circumstances. Its advantages over the traditional speckle reduction filter approaches have also been demonstrated.
- Synthetic Aperture Radar
- Minimum Mean Square Error
- Synthetic Aperture Radar Image
- Anisotropic Diffusion
- Speckle Noise
Due to some unique characteristics of Synthetic Aperture Radar (SAR), such as all-weather, capability of penetrating cloud cover, and special reflect spectrum for object, this gives it a considerable advantage over other infrared or optical sensors. Meanwhile, it accelerates SAR applications for environmental monitoring. However, the imagery produced by SAR systems is degraded by a high level of noise due to the inherent nature of radar backscatter. The largest source of noise in a SAR image is the speckle noise that is caused by the interaction of the coherent imaging radar beam and the relatively rough surface being imaged. Because of the speckle, a form of multiplicative, locally correlated noise, typical image processing techniques experience a great difficulty when applied to SAR imagery.
In order to tackle this problem for SAR image processing, many methods have been attempted. The most widely cited and applied in the SAR community are based on the minimum mean square error (MMSE) theory, and mainly focused on the geometric domain, including Lee , Enhanced Lee , Kuan , Frost , Enhanced Frost  filters. At the same time, more and more diffusion and random filters are put forwards to reduce SAR image speckle, such as anisotropic diffusion , nonlinear diffusion , speckle reducing anisotropic diffusion (SRAD) , detail preserving anisotropic diffusion (DPAD) , wavelet processing , Markov Random Fields , and hysteresis thresholding .
Most of above traditional speckle filters, which mainly rely on local adaptive statistical data related to the filtered pixel, have been proposed to reduce speckle, the main difficulty being to suppress speckle while preserving image information such as edges or textures . On one hand, because the results depend on the occurrence of the filter window over an area, the achievement of both speckle reduction and preservation of edge data is only possible when the filter window happens to cover an area that is uniform. On the other hand, traditional filters are not directional. Thus, more and more edge preserving techniques were developed and applied in to optimize MMSE speckle filtering . For example, ratio edge detector  and the evolution of the coefficient of variation  have been validated with the MMSE for detecting edge and maximum homogeneous area for speckle smoothing, respectively. For polarization SAR images, the classification map of scattering characteristics is used with speckle filtering of polarimetric synthetic aperture radar (POLSAR) filter  for speckle reduction, while perfectly preserving strong point target signatures, and retains edges, linear, and curved features. Meanwhile, for multitemporal SAR image filtering, a constant false-alarm rate (CFAR) edge detector and conventional statistical criteria are integrated for spatial and temporal information preservation . In addition, for the diffusion speckle removal approach reduces noise and texture but also preserves and enhances edges, the diffusion formulations such as anisotropic diffusion , nonlinear diffusion , SRAD , DPAD  are becoming an important way for edge-sensitive speckle smoothing.
In order to not only remove speckle but also keep edge, an alternative approach is proposed in this paper, it mainly relies on the theory of curvature anisotropic flow, that is, level set method. Firstly, the Lee filter can be transformed as a formulation of anisotropic diffusion function . Thus, after the analysis, the relationship between curvature flow and Lee speckle filter, we considered further and deduced the anisotropic diffusion filter into a formulation of level set. Therefore, the new technique is not only based on the minimum mean square error approach to filtering as the Lee filters, but also on anisotropic diffusion sensitive to edges. The paradigm of the level set is a numerical method for tracking the evolution of contours and surfaces. In this approach, an image is interpreted as a collection of iso-intensity contours. Similar to anisotropic diffusion, under a curvature-driven factor, smoothing takes place inside a region, but not across region boundaries. So the attractive quality of this approach is that sharp boundaries are preserved while speckles are smoothed. Secondly, in order to avoid long-term iteration in the manner of traditional anisotropic diffusion, the different speed of flow approach was introduced into the new technique as a selective smoothing switch according to effects by intensity and curvature gradient of neighbor pixels. An image is filtered as a set of iso-intensity contours which can be evolved. The spikes of noise correspond to high curvature objects, so these are easily reduced by flow propagation under curvature; meanwhile, oscillations in boundaries are smoothed out, all the while preserving sharp boundaries.
The rest of this paper is organized as follows. In Section 2, we first review the background of level set and the main SAR image noise removal algorithms. In Section 3, we propose new filter algorithm based on level set and Lee filter, including processing flowchart, filter characteristics, and the flow speeds switch control. Section 4 presents the results of experimental evaluation of simulation image and SAR images, respectively. Finally, Section 5 summarizes this proposed method.
During the last 20 years, many speckle filters for SAR images were developed and documented. Speckle noise is a common phenomenon in all coherent imaging systems like laser, acoustic, and SAR imagery. The source of this noise is attributed to random interference between the coherent returns issued from the numerous scatterers present on a surface, on the scale of a wavelength of the incident radar. Speckle noise gives the SAR image a grainy appearance and prevents target recognition and texture analysis efficiently. Obviously, speckle filtering turns out to be an important preprocessing step for some detection or classification optimization. Because speckle noise has the characteristics of a random multiplicative, thus traditional filtering will not remove it easily. In practice, the most widely used filters such as Lee (Enhanced Lee), Kuan and Frost filters are based on a Minimum Mean Square Error (MMSE) criterion. Thus, in this section, we briefly describe the representative MMSE method, that is, Lee filter and background of level set.
2.1. Lee Filter
where is the effective number of looks of the noisy image, and and are the intensity variance and mean over a homogeneous area of image, respectively. Obviously, in uniform areas, the value ofCsapproachesCuand nears to zero, leading to the same result as mean filter. At edges, the value ofCsbecomes larger, and nears to unity, resulting in little modification.
2.2. Level Set Method
Level set, introduced by Osher and Sethian , is an efficient numerical technique for interface propagation, which we have introduced in SAR images segmentation [17, 18]. A brief introduction of this method is given here. The detailed explanation can be found in Sethian .
The key of level set for image processing is to define reasonable scalar functionF. For SAR image, intensity model and curvature model are integrated to improve effectiveness. And in the previous image filter, scalar function is usually simplified to = /  or = .
Referringto the above Lee filter theory and level set numerical approach, in this section we derive a new level set method from traditional Lee filter for spackle noise reduction in SAR image.
3.1. The Level Set Formulation for Speckle Reduction
As we know, the Lee filter, or Kuan filter processes a current pixel based on its intensity and neighboring pixels inside a fixed square window. Based on Lee filter,Yu and Actom developed an anisotropic diffusion scheme.
3.1.1. From Lee Filter to Discrete Isotropic Diffusion Equation
Thus, traditional Lee filter firstly is represented as a discrete isotropic diffusion equation.
3.1.2. From Lee Filter to Discrete Anisotropic Diffusion Equation
where div is the divergence operator, and the gradient operator. The above modification of the Lee filter is a form of anisotropic diffusion with . The Lee filter processes a current pixel based on its intensity and intensities of neighboring pixels inside a fixed square window without mechanism to enhance edges or feature structures within a window.
The above anisotropic diffusion combines a normalized gradient magnitude operator and laplacian operator to act like edge detector for speckled imagery. The above equations from (10) to (15) are cited from the paper byYu and Acton  and the detailed deduction refers to their anisotropic diffusion scheme. In addition, Aja-Fernandez and Alberola-Lopez  modified the SRAD filter to rely on the Kuan filter rather than the Lee filter, and estimated the local statistics using a larger neighborhood than the 4 direct neighbors, showing that better results and better stability can be obtained using a 5 5 neighborhood.
3.1.3. From Lee Filter to Level Set Filter
As a result, we obtained a curvature-flow-driven level set filter from MMSE Lee filter technique in the manner of partial diffusion equation, in other words we deduced the level set numerical form of traditional Lee filter. Obviously, the interface represents appropriate finite difference operator for the spatial derivate, so calculation anisotropically takes place in different directions. The modification including directional sensitivity and filtering perpendicular to the edge direction would significantly improve reducing the speckle in the vicinity of edges and small features.
Areas of high curvature will diffuse faster than areas of low curvature. Hence, small jagged noise artifacts will disappear quickly, while large-scale interfaces will be slow to evolve. Moreover, the front of level set equals zero. Thus, the attractive quality of this approach is that sharp boundaries are preserved; smoothing takes place inside a region, but not across region boundaries, thereby preserving sharp boundaries between objects.
3.2. Curvature-Driven Model
Because the curvature plays important role in the proposed method, in this section, we detailed the numerical solution of curvature model by PDE. In order to obtain the stability of a boundary, an up-wind partial differential scheme is employed to compute curvature value, which relies on a one-sided derivative that looks in the up-wind direction of the moving wave front, and thus avoids the over-shooting associated with finite forward differences.
where computes the new value at , , and , ; thus information for the solution propagates from right to left. computes the new value at , , and , ; thus information for the solution propagates from left to right. computes the new value at , and , , thus information for the solution propagates from both sides.
With this equation, the direction and velocity of the speed dependent on the curvature can be easily derived.
Smoothing process anisotropically takes place in different directions. The speed function plays a central role in the level set as its value and normal direction determine the propagation velocity and direction of the front interface.
3.3. Min-Max Curvature Flow
According to function (16), unlike the common filters based on statistical average, the proposed approach relies on propagation diffusion iteration method. It should be noted that although the evolution at the boundary is slow, some diffusion still occurs. Thus, according to Grayson's theorem, continual application of this curvature flow scheme will eventually result in the removal of information as each contour shrinks to a point and disappears. So, we used the min-max flow  for preserving finer detailed structures in images from over-smoothing.
where is calculated as the average of pixel intensities along the direction perpendicular to the gradient at the extrema of the local neighborhood within the filter window . Comparison between the neighborhood average and the threshold is used to select the right speed function to use. Flow under min( ,0) would preserve some of the structure of the curve, while flow under max( ,0) completely diffuses away all of the information. This switch prevents the unwanted diffusion of the proposed curvature flow approach. In other words, a speed of will cause small dark regions in a predominantly light region to shrink. Conversely, a speed of will cause light regions in a predominantly dark region to shrink.
3.4. Filter Framework
Firstly, this filter requires four parameters used in the computation of the level set evolution, the number of iterations to be performed, the time step and the vicinity range, and the scale of the noise that one wants to remove.
Secondly, compute curvature-driven model and adaptive filter coefficient, respectively.
Thirdly, for preserving finer detailed structures in images, switch curvature flow according to min-max flow.
Continuously update pixel intensity by function (18).
This approach adjusts the value of each pixel based on the iteration.
In this section, we firstly test the proposed level set filter using simulated data, then, demonstrate the algorithm on land and ocean SAR data. In each experiment, we compare the results of the proposed level set filter with those of six existing schemes, that is, Lee filter, Enhanced Lee Filter, Kuan Filter, Frost Filter, Enhanced Frost Filter, Gamma Filter. We examine the mean preservation error, the standard deviation and error standard deviation reduction to quantify the performance of algorithms in different regions.
where is a result pixel after filter and is a supposed true pixel. Usually, STD represents information complexity while error standard deviation shows difference between result image after filter the degraded image and theoretically true image, that is, the initial image.
Furthermore, the profile graph was usually used to illustrate performance of noise smoothing and edge preservation . In this paper, profiles were select by arbitrary line in the image, where -axis represents line distance and -axis is the intensity of pixels.
It should be noted that although the evolution at the boundary is slow, some diffusion still may occur. In addition, considering more iterations will result in further smoothing and will linearly increase time consumption , thus the small number of iterations is selected to stop the computation of the level set evolution. In the following experiments, the filter was run with less than 5 iterations, 5 5 pixels vicinity range for intensity model.
4.1. Simulation Results
Firstly, the initial image is degraded by multiplicative and random noise in Figure 2(b). Then based on degraded image, result images computed by traditional methods such as Lee filter, Enhanced Lee Filter, Kuan Filter, Frost Filter, Enhanced Frost Filter, Gamma Filter and proposed level set filter are denoted in Figures 2(c), 2(d), 2(e), 2(f), 2(g), 2(h), and 2(i), respectively.
Mean, standard deviation and error standard deviation for regions A, B, and C.
Error std: 0.000
Error std: 0.000
Error std: 0.000
Degraded image by speckle
Error std: 1694.889
Error std: 576.947
Error std: 864.951
Error std: 796.122
Error std: 110.071
Error std: 285.072
Enhanced Lee Filter
Error std: 725.142
Error std: 85.426
Error std: 283.281
Error std: 756.300
Error std: 104.209
Error std: 266.664
Error std: 737.993
Error std: 104.189
Error std: 253.703
Enhanced Frost Filter
Error std: 705.102
Error std: 76.273
Error std: 268.414
Error std: 777.002
Error std: 113.264
Error std: 326.011
Level set Filter
Error std: 420.243
Error std: 323.605
Error std: 376.056
Comparing the traditional filter results of region A in Figure 2, it is clearly that noise in the margin is removed perfectly by proposed level set filter and the edges of image and annotation are preserved effectively (Figure 2(i) versus Figures 2(c), 2(d), 2(e), 2(f), 2(g), 2(h)).
In addition, because the level set filter not only preserves edges but also enhances edges by inhibiting diffusion across edges and allowing diffusion on either side of edge. So, standard deviation will become large in Table 1. Meanwhile, it appears that some edges are enhanced, that is, the difference of intensity value on either side of edge becomes large (Figure 3(i)). At a word, black will become more black while white more white on either side of edge.
4.2. ERS-2 SAR Image Results
ERS-2 SAR (Precision Image product with number of looks 3) PRI image with 559 649 pixels on the Shangdong peninsula landscape in Figure 4. In addition, profile of partial pixels intensity along vertical line D is plotted in Figure 5.
ERS-2 SAR (Precision Image product with number of looks 3) PRI image with 295 360 pixels at the South China Sea seascape in Figure 6. In addition, profile of partial pixels intensity along vertical line E is plotted in Figure 7.
In Figure 4, vertical line D is cross some croplands and forms these show comb structure in Figure 5. Figures 4(b), 4(c), 4(d), 4(e), 4(f), and 4(g) demonstrate that the above traditional filters cannot preserve the narrow strips perfectly due to only inhibiting smoothing near edges and not supporting direction filter. In contrast, proposed level set filter (Figure 5(h)) keeps up the large strips and smoothing the small strips.
In Figure 6, besides high noise coming from the inherent nature of radar backscatter, due to the dampening effects of Bragg wave on the sea generated by low backscatter cross section of the surface, SAR images usually have a high noise. Figure 6 shows that the result of proposed filter is near to the result of traditional methods such as Lee filter, Enhanced Lee Filter, Kuan Filter, Frost Filter, Enhanced Frost Filter and Gamma Filte (Figure 6(h) versus Figures 6(b), 6(c), 6(d), 6(e), 6(f), and 6(g)).
In this paper, a filter, which is based on level set, has been explored for speckle removal in SAR images. As we know, Level set is an efficient numerical technique while Lee filter is a classical speckle noise filter. What is new here is the correspondence between these two types of frameworks, moreover, for speckle filtering purposes. Unlike other existing speckle reducing filters, it is based on the theory of flow propagation and Lee filter, which uses intensity and curvature as the front advancing impetus to modify noise by differential equation iterations. According to theoretical analysis, imagery experiment including simulation and landscape, and seascape of ERS-2 SAR images, experimental results reveal that the proposed algorithm provides superior performance in comparison to the conventional speckle filters, such as Lee filter, Enhanced Lee Filter, Kuan Filter, Frost Filter, Enhanced Frost Filter and Gamma Filter, in terms of both smoothing speckle inner or outer uniform regions and preserving (sometimes enhancing) edges and detail features. We hope that the method described in this paper provides a basis for further exploration of level set in remotely sensed image processing.
This research was funded by Ministry of Science and Technology of PRC under Contract no. 2006AA12Z116, and their support is gratefully acknowledged.
- Lee J-S: Speckle suppression and analysis for synthetic aperture radar images. Optical Engineering 1986, 25(5):636-643.View ArticleGoogle Scholar
- Lopes A, Nezry E, Touzi R, Laur H: Maximum a posteriori filtering and first order texture models in SAR images. Proceedings of the International Geoscience and Remote Sensing Symposium (IGARSS '90), 1990, College Park, Md, USA 2409-2412.View ArticleGoogle Scholar
- Kuan DT, Sawchuk AA, Strand TC, Chavel P: Adaptive restoration of images with speckle. IEEE Transactions on Acoustics, Speech, and Signal Processing 1987, 35(3):373-383. 10.1109/TASSP.1987.1165131View ArticleGoogle Scholar
- Frost VS, Stiles J, Shanmugan K, Holtzmann J: A model for radar images and its application to adaptive digital filtering of multiplicative noise. IEEE Transactions on Pattern Analysis and Machine Intelligence 1982, 4: 157-166.View ArticleGoogle Scholar
- Perona P, Malik J: Scale-space and edge detection using anisotropic diffusion. IEEE Transactions on Pattern Analysis and Machine Intelligence 1990, 12(7):629-639. 10.1109/34.56205View ArticleGoogle Scholar
- Alvarez L, Mazorra L: Signal and image restoration using shock filters and anisotropic diffusion. SIAM Journal on Numerical Analysis 1994, 31(2):590-605. 10.1137/0731032MathSciNetView ArticleMATHGoogle Scholar
- Yu Y, Acton ST: Speckle reducing anisotropic diffusion. IEEE Transactions on Image Processing 2002, 11(11):1260-1270. 10.1109/TIP.2002.804276MathSciNetView ArticleGoogle Scholar
- Aja-Fernandez S, Alberola-Lopez C: On the estimation of the coefficient of variation for anisotropic diffusion speckle filtering. IEEE Transactions on Image Processing 2006, 15(9):2694-2701.View ArticleGoogle Scholar
- Achim A, Tsakalides P, Bezerianos A: SAR image denoising via Bayesian wavelet shrinkage based on heavy-tailed modeling. IEEE Transactions on Geoscience and Remote Sensing 2003, 41(8):1773-1784. 10.1109/TGRS.2003.813488View ArticleGoogle Scholar
- Sveinsson JR, Benediktsson JA: Almost translation invariant wavelet transformations for speckle reduction of SAR images. IEEE Transactions on Geoscience and Remote Sensing 2003, 41(10):2404-2408. 10.1109/TGRS.2003.817844View ArticleGoogle Scholar
- Kanaa TFN, Tonye E, Mercier G, et al.: Detection of oil slick signatures in SAR images by fusion of hysteresis thresholding responses. Proceedings of the International Geoscience and Remote Sensing Symposium (IGARSS '03), July 2003, Toulouse, France 2750-2752.Google Scholar
- Wu Y, Maître H: Smoothing speckled synthetic aperture radar images by using maximum homogeneous region filters. Optical Engineering 1992, 31(8):1785-1792. 10.1117/12.59897View ArticleGoogle Scholar
- Touzi R: A review of speckle filtering in the context of estimation theory. IEEE Transactions on Geoscience and Remote Sensing 2002, 40(11):2392-2404. 10.1109/TGRS.2002.803727View ArticleMathSciNetGoogle Scholar
- Lee J-S, Mitchell RG, Dale LS, Pottier E, Ferro-Famil L: Scattering-model-based speckle filtering of polarimetric SAR data. IEEE Transactions on Geoscience and Remote Sensing 2006, 44(1):176-187.View ArticleGoogle Scholar
- Trouvé E, Chambenoit Y, Classeau N, Bolon P: Statistical and operational performance assessment of multitemporal SAR image filtering. IEEE Transactions on Geoscience and Remote Sensing 2003, 41(11):2519-2530. 10.1109/TGRS.2003.817270View ArticleGoogle Scholar
- Osher S, Sethian JA: Fronts propagating with curvature-dependent speed: algorithms based on Hamilton-Jacobi formulations. Journal of Computational Physics 1988, 79(1):12-49. 10.1016/0021-9991(88)90002-2MathSciNetView ArticleMATHGoogle Scholar
- Huang B, Li H, Huang X: A level set method for oil slick segmentation in SAR images. International Journal of Remote Sensing 2005, 26(6):1145-1156. 10.1080/01431160512331326747View ArticleGoogle Scholar
- Huang X, Huang B, Li H: A fast level set method for synthetic aperture radar ocean image segmentation. Sensors 2009, 9(2):814-829. 10.3390/s90200814View ArticleGoogle Scholar
- Sethian JA: Level Set Methods and Fast Marching Methods: Evolving Interfaces in Computational Geometry, Fluid Mechanics, Computer Vision and Materials Science. Cambridge University Press, Cambridge, UK; 1999.MATHGoogle Scholar
- Rudin L, Osher S, Fatemi E: Nonlinear total variation based noise removal algorithm. In Modelisation Matematiques pour le Traitement D'images. INRIA; 1992:149-179.Google Scholar
- Sapiro G, Tannenbaum A: Image smoothing based on affine invariant flow. In Proceedings of the Conference on Information Sciences and Systems, 1993. Johns Hopkins University;Google Scholar
- Whitaker RT, Xue X: Variable-conductance, level-set curvature for image denoising. In Proceedings of the IEEE International Conference on Image Processing, October 2001, Thessalonik, Greece. IEEE; 142-145.Google Scholar
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.