You are on page 1of 5

MASTERS IN COMPUTER VISIONAssignment, Image Analysis Module, 2011

Wavelet Steerability & Riesz Transform with Application to Medical Image Analysis
Sharib Ali University of Burgundy, France, Le Creusot,
AbstractThe aim of this paper will be to make a literature review on design of steerable lters, and introduction to N th order extension of Riesz transform. It will provide a fundamental concept of the Simoncellis steerable pyramid whose construction was based on the lterbank design. The algorithm reviewed here allows the specication of wavelets with any order of steerability in any number of dimensions. Also, we will discuss dierent implementations of such steerable lters particular to that in medical image analysis. Index TermsHilbert Transform, multiresolution decomposition, Riesz transform, Steerable lters

I. Introduction For visual processing and perception, scale and directionality has key consideration. [4] Feature extraction has always been a measure challenge in the eld of image processing. Especially, when it comes to texture, feature extraction is ideally very dicult. Many work has been done by the researchers in image processing and applied mathematics for developing feature extraction algorithms. Mallat set the foundation of the multiresolution theory of the wavelet transform and involves in the continuously dened functions in L2 ( d ) [1]. This theory yields a one toone decomposition of a signal across scale; but unfortunately, the underlying bases (which are typically separable) have uneven angular responses and they are strongly biased towards the vertical and horizontal directions. Steerability is the ability to synthesize a lter of any orientation from a linear combination of lters at xed orientations (called basis lters). Such specically tunned to orientation lters were proposed by Freeman and Adelson. Later, Simoncelli and coworkers designed a complete family of directional multiscale transforms which is indexed by multiorientation with k bands. Their steerable lters were overcomplete (ie. redundant) by a factor of 4k/3 + 1 and has an important property of selfinverting. We will discuss these steerable pyramids and its applications in section II as it is the most fundamental concept of this paper. [3] Riesz transform is the multidimensional generalization of the Hilbert transform. Its introduction to signal processing is more recent. Felsberg usef the transform to dene the monogenic signals as a 2D generalization of the analytic signal[3]. We can generalize the key goals of the Riesz transform surveyed in this paper as A general framework for constructing families of steerable wavelet frames and pyramids of L2 ( d ) with arbitrary orders in any number of dimensions d 2.
Supervisor: Fred Truchetet

A continuousdomain formulation that formalizes and extends the technique proposed by Simoncelli et. al. Illustration of the method with the design of a novel family of multidimensional RieszLaplace wavelets that essentially behave like the N th order partial derivatives of an isotropic Gaussian kernel. We have divided this paper into dierent sections as: Section II will illustrate the basic framework of the steerable lters and construction of steerable pyramids proposed by Simoncelli et. al. Section III will demonstrate the Riesz transform and its higher order extensions. We have some applications of such steerable wavelet frame (Riesz transform) to medical image analysis in Section IV. Finally, in Section V, we will conclude the paper highlighting some major advantages and limitations.

II. Steerable Filters and Simocellis steerable pyramid Freeman & Adelson Steerable lters, 1991 In many vision and image processing applications such as texture analysis, edge detection, image data compression, motion analysis and image enhancement; oriented lters are used. Under adaptive control, it is required to examine the lter output at arbitrary orientation and phase. We must have an idea of how many lters are required and how to interpolate them between the responses. An approach developed by Freeman & Adelson enables us to realise such steerable lters which uses a class of lters having arbitrary orientation synthesized as a linear combination of a set of basis lters. consider the 2D, circularly symmetric Gaussian function, G, given by, G(x, y) = e(x
2

+y 2 )

(1)

where, scaling and normalization constants have been set to 1 for convenience. The rst derivative of a Gaussian, o G0 , is 1 G0 = 1
o
2 2 (x2 +y2 ) e = 2xe(x +y ) x

(2)

The rotated 90 degree of the same function is, G90 = 1


o
2 2 (x2 +y2 ) e = 2ye(x +y ) x

(3)

The arbitrary orientation can be synthesized by taking o o a linear combination of G0 and G90 [2]: 1 1 G = cos()G0 + sin()G90 1 1 1
o o o

(4)

MASTERS IN COMPUTER VISIONAssignment, Image Analysis Module, 2011

Fig. 1. (a).G0 , rst derivative w.r.t. x(horizontal) (b). G90 , 1 1 o rotated by 90o (c). G30 , linear combination of these two lters 1

Fig. 3. System diagram for steerable pyramid design: Boxes corresponding to 2D and 2U corresponds to downsampling and upsampling by a factor of 2. Only the lower bands are sampled

Fig. 4. Properties of the Steerable Pyramid relative to two other wellknown multi scale representation

Fig. 2. Design of Steerable lters with 2nd order Gaussian

Since, G0 and G90 span the set of G lters, we call 1 1 1 o them basis lters for G . The cos() and sin() terms are 1 the corresponding interpolating functions for these basis lters. Designing Steerable Filters [2] Those functions which have small number of basis lters are practically important. Let us design a steerable quadrature pair based on the frequency response of the second derivative of a Gaussian, G2 given be, G = k1 ()G0 + k2 ()G60 + k3 ()G120 2 2 2 2

representations. Figure 5 shows the implementation output of such multi scale and multi oriented steerable lter pyramids on a synthetic image. We can clearly see that at the nal level, the image is much more enhanced with clear detection of edges. Many successful applications of such steerable pyramids can be listed Contour detection. Image ltering and denoising. Orientation analysis. Texture analysis and synthesis. However, limitations of such approach could be outlined as They are purely discrete framework. They do not extend to dimensions higher than two. III. Riesz transform and its higher order extensions The major goal of this section is to lay the basic mathematical formulations and understanding before going to the higher order Riesz transforms. So, we give brief explanations on the Hilbert transform (most fundamental 1D transform for building the Riesz operators), then we will introduce a short introduction to complex wavelet trans-

(5)

,where, kj ()is the interpolation function with basis lters 120 G0 , G60 and G2 . Then, the steerable lter architec2 2 ture can be shown in Figure 2. Most steerable lters do not have xy separable as it takes high computational time. However, we can have such xy separable basis but we are not discussing here as the motive is to just give the reader a brief idea on steerable lter design. Simoncellis steerable pyramids, 1995 [3] Simocelli proposed an architecture for ecient and accurate linear decomposition of an image into scale and orientation subbands. A recursive system shown in Figure. 3 constrains the tuning of the scale. The orientation tuning is constrained by steerability property. A selfinverting (i.e. matrix corresponding to the inverse transformation is same as the transpose of the forward transformation matrix) and aliasingfree properties are achieved. The nal transform is 4k/3 overcompleter, k being the orientation bands. The table shown in Figure 4 gives a clear idea of advantage of such steerable pyramids over other multiscale

Fig. 5. A 3level steerable pyramid applied on an image (three bandpass image at each scale and a nal lowpass lter)

Sharib Ali: WAVELET

C.1 Invariance The Riesz transform is translation and scale invariant. It is also rotationinvariant. xo
Fig. 6. Analytic signal
d

, Rf (. xo )(x) = Rf (.)(x xo )
+

(10) (11)

a C.2 Steerability

. x , Rf ( )(x) = Rf (.)( ) a a

forms (which will formulate the real positive considerations for getting more locally oriented Riesz transform) and nally we will discuss Riesz transform both as 2D and higher dimensions. A. Hilbert transform The Hilbert transform is a 1D linear, shiftinvariant operator which is dened as a function that maps all the cosines into sine functions, without aecting the amplitude of the signal. Thus, it behaves like an allpass lter with the following transfer function: Hf (x) = (h f )(x) jsgn()f (x) = j Its impulse response is given by: h(x) = 1 x
F

The Riesz transform is steerable of which the component lters are 90o rotated versions of each other. Inother words, Riesz transform are directional version of the Hilbert transform.
d

Hu f (x) =
n=1

un Rn f (x) =< u, Rf (x) >

(12)

C.3 InnerProduct Preservation The Riesz transform satises the Parsevallike identity giving a direct consequence to the energy conservation property.
d

f (x) ||

(6) Rf

2 Ld 2

=
n=1

Rn f

2 L2

2 L2

(13)

C.4 Orthogonality (7) Let f(x) L2 ( d )be a realvalued function whose Fourier transform f satises the coordinate wise symmetry (1 , ...., d )|2 = |f (1 , ...., d )|2 then, relations: |f Rn f, f
L2

Equation 7 represents the decaying feature which means that it is a nonlocal operator. B. CWT An analytical signal can be represented as the sum of the signal and its Hilbert transform adn can be given by Equation 8. Inother words, such signals are constructed from real valued signal and its Hilbert transform. fana (x) = f (x) + jHf (x) (8)

= 0, f orn = 1, ...., d
L2

(14) (15)

Rn f, Rm f

= 0, m = n

We can represent such analytic signal after Fourier analysis as in Figure 6 C. Riesz Transform Riesz transform is a multidimensional extension of the Hilbert transform. The spacedomain representation of dchannel lterbank is given by: R1 f (x) . F w d Rf (x) = . j w f (x), f Lp ( ) (9) Rd f (x) The impulse responses of the Riesz transform are all anti symmetric: n x = hn (x) = hn (x) and essentially de1 caying given by , where d is the number of dimension. ||x||d So, though Riesz transform are nonlocal operators but at higher order they become more local. Properties of Riesz Transform [5]

HigherOrder Riesz transform [5] Higherorder Riesz transform has improved angular selectivity. To explain such steerable behaviour, we relate R(N ) to the N fold version of the directional Hilbert transform Hu along the direction specied by the unit vector u. In Fourier domain we can rewrite the equation 16 as
d

Hu f () =
n=1

ui

ji f ()

(16)

The block diagram representing such higherorder riesz transform is shown in Figure 7 IV. Riesz transform applied to medical image analysis Lung texture analysis [7] In medical image interpretation, the rst step is to detect the abnormal image patters which is related to visual perception. Such visual perception greatly relies on the texture properties which gives a characterization of biomedical tissue. Like, healthy and pathological lung parenchyma in highresolution computed tomography (HRCT ) from patients with interstitial lung diseases (ILD) can only be described in terms of texture properties.

MASTERS IN COMPUTER VISIONAssignment, Image Analysis Module, 2011

Fig. 7. Steerable lterbank implementation of the N th order directional Hilbert transform in multiple dimensions

Fig. 10. Toprow: 3D original uoresecence image of collagen laments and its isosurface representation; Bottomrow: Anisotropic feature enhancement image and its isosurface representation(using 3D riesz transform)

Fig. 8. Lung textures showing healthy and dierent ILD

In literature, there are many comprehensive automated characterization of the local orientations requiring the multiscale and multiorientation analysis with higher angular precision. To compare textures with distinct local dominant orientations dom determined, the Riesz templates are aligned based on the local prevailing orientation of: () dom (xp ) =argmax ((h1 g) f )(xp ) (17)
[0,]

3D wavelet frame has the key property of steerability. The link between the Riesz transform and the directional Hilbet transform to dene a multiresolution monogenic signal analysis in 3D has been exploited and has been used in biological imaging with the enhancement of anisotropic structures in 3D uorescence microscopy. As shown in Figure 10, this technique is able to smooth the image while preserving the directional features, which gives good result to the isosurface visualization in 3D of the laments. V. Conclusion We have investigated the basics of steerable ters, steerable pyramids and then connected them to better understand the most important property of Riesz transform which is Steerability . We have seen 2D and ND Riesz transform and have known how the increase in high dimension makes localizes the riesz transform. Then, we gave some examples available in literatures regarding the 2D and 3D implementation of Riesz transform in particular to medical image analysis. Not much has been done in this eld though it is the very aective in enhancing the image quality in Ndimensional could be taken as one of the major limitation of Riesz transform. References
[1] S. G. Mallat, A theory of multiresolution signal decomposition: The wavelet representation, IEEE Trans. Pattern Anal. Mach, Intell., vol.11, no. 7, pp. 674693, Jul. 1989. [2] W. T. Freeman and E. H. Adelson, The design and use of steerable lters, IEEE Trans. Pattern Anal. Mach, Intell., vol.13, no. 9, pp. 891906, Sep. 1991. [3] E. Simoncelli and H. Farid, Steerable wedge lters for local orientation analysis, IEEE Trans. Image Process., vol.5, no. 9, pp. 13771382, Sep. 1996. [4] M. Jacob and M. Unser, Design of steerable lters for feature detection using Cannylike criteria, IEEE Trans. Pattern Anal. Mach, Intell., vol.26, no. 8, pp. 10071019, Aug. 2004. [5] Michael Unser and Dimitri Van De Ville, Wavelet steerability and the higherorder Riesz transform, IEEE Transactions on Image Processing, vol.19, no. 3, pp. 636652, March 2010. [6] N. Chenouard and M. Unser Steerable Pyramids and Tight Wavelet Frames in L2 ( d ), Proceeding of the Eighth IEEE International Symposium on Biomedical Imaging: From Nano to Macro (ISBI11), 30th April2011

Due to the steerability property of Riesz lterbanks, dom (xp ) and the orientation of the templates are obtained analytically. The method was evaluated on 13808 32x32 overlapping blocks from annotated 85 cases with leave onepatientout crossvalidation and compared with op 3 timized GLCM having orientation angles 0, , , , 4 2 4 distances d = 1 : 5 and greylevels l = 8, 16, 32. Optimized support vector machines were used for classication. The results are shown in Figure 9 Analysis of the result:The classication performance allows a global accuracy of 6.1% compared to GLCMs. Microscopy Image Analysis and Processing [8]

Fig. 9. Classspecic accuracies of best performing setups

Sharib Ali: WAVELET

[7] A. Depeursinge, F. R. Antonio, D. V. D. Ville, Lung Texture Classication Using LocallyOriented Riesz Components, Medical Image Computing and ComputerAssisted Intervention MICCAI 2011, vol. 6893, pp. 231238, Springer, 2011. [8] N, Chenouard and M. Unser 3D Steerable wavelets and monogenic analysis for bioimaging, Proceedings of the Eighth IEEE International Symposium on Biomedical Imaging: From Nano to Macro (ISBI11), 30th April, 2011.,

You might also like