Available online at www.sciencedirect.

com

Procedia Engineering 35 (2012) 202 209

International Meeting of Electrical Engineering Research ENIINVIE-2012

A new recursive scheme of the unbiased FIR filter to image processing

L. J. Morales-Mendoza*, R. F. Vzquez-Bautista*, E. Morales-Mendoza*, Y. Shmaliy and H. Gamboa-Rosales
* Veracruz University, Poza Rica Ver. Mxico Guanajuato University, Salamanca Gto. Mxico, UAZ, Zacatecas Mxico

Abstract In this paper, we present a new computational scheme for the implementation of Unbiased Finite Impulse Response filter to image processing. The recursive scheme is focused both signal to noise ratio and root mean-square-error on filter improved performance. In this way, quantitative evaluations are computed to verify the performance carried out by the new recursive computational scheme under metrics. On the other hand, qualitative evaluations are showed for both cases: signals (one dimensional) and images (two dimensional). Finally, this efficient computational algorithm manifests an outstanding signal/image enhancing with strong metric convergence despite the computational complexity.

2011 Published by Elsevier Ltd. Selection and/or peer-review under responsibility of the Organizing Committee of the ENIINVIE-2012.
Keywords: unbiased FIR filter, image processing, recursive scheme

1. Introduction The problem of saving a sharp edge with a simultaneous image enhancing in the image, is typical case for image processing. An overall panorama of nonlinear filtering that follows the median strategy has been given by Pitas and Venetsanopoulos in [1], including the important modifications for a large class of

* Tel.:+52-782-8225878; fax: +52-782-8225878. javmorales@uv.mx.

1877-7058 2012 Published by Elsevier Ltd. doi:10.1016/j.proeng.2012.04.181

L. J. Morales-Mendoza et al. / Procedia Engineering 35 (2012) 202 209

203

nonlinear filters employing the order statistics. This type of nonlinear filters is applied in a variety of areas such as biomedical images processing [2], and remote sensing [3]. The structures with Finite Impulse Response (FIR) have inherent properties such as the Bounded Input/Bounded Output (BIBO) stability and round-off error that are desirables in almost all systems. A typical example of this system is the FIR-Median Hybrid (FMH) filter developed by Heinonen et. al. in [4] which predict the root signals from input signals based on Gaussian, double exponential, and uniform density functions. Otherwise, Shmaliy introduced, in [5], the Unbiased FIR (UFIR) filter which efficient algorithm is leaded by strong mathematical model from a modified polynomial filter to reducing the TimeInterval Error (TIE) for Global Position System (GPS) signals. Moreover, this filter was applied to processing 2-D signals (fetal ultrasound images) which enhance and preserve image characteristics through complex scenarios by Morales-Mendoza et. al. in [6] and [7]. An important step ahead has been made in [8], where the Unbiased FMH structures have been designed as a generic form of the p-lag. In the sequel, the FMH structures have extensively been investigated, developed, and used by many authors. The algorithm issues for the filter design have been discussed in [9] and [10]. Basically, hybrid FIR structures can be designed using different types of estimators. Among possible solutions, the polynomial estimators occupy a special place, since the polynomial models often well formalize a priori knowledge about different processes. Based on [8], the implementation of UFIR filters for image processing is relatively easy, but the estimated image characteristics are not ideal even when quantitative evaluation metrics so specified as presented in [11]. Nevertheless, the recursive schemes are an alternative way of implementing the computational algorithms for signal processing as shown in [12]. Therefore, in this paper we present a new computational scheme to implement the UFIR filters in image processing. 1.1. Polynomial Image Model A two-dimensional image is often represented as a kc kr matrix M = {i,j}. To provide two dimensional filtering, the matrix can be written in the form of a row-ordered vector or a column-ordered vector, respectively

x r = 1,1 xc
1,1

[ = [

1,k
c

k ,1
c

k ,k
c

k ,1

1,k

k ,k
c

], ].
T T

(1) (2)

The filtering procedure is then often applied twice, first to (1) and then to (2), or vice versa. An electronic image x1n can be substituted with the discrete time-invariant deterministic signal. Following [8], such a signal can be represented in state space by the state and observation equations as follows, respectively

x n = A N 1x n N +1 ,

(3) (4)

yn = Cx n + vn ,

where xn = [x1n x2n xKn] is the K 1 vector of the states, yn is the measurement representing the electronic image, vn is the measurement noise, the 1 K measurement matrix is C = [1 0 0 0], and the K K triangular matrix Ai is specified by

204

L. J. Morales-Mendoza et al. / Procedia Engineering 35 (2012) 202 209

1 i 0
i

1 2

( i )2
i
1 ... 0

... ... ... ... ...

A = 0 0 ... ... 0 0

( i )K 1 K 2 1 ( K 2 )! ( i ) K 3 1 ( K 3 )! ( i )
( K 1)!
1

(5)

... 1

Now, the noiseless model (3) projects ahead from n N + 1 to n using the degree Taylor polynomial which is represented as follows,

xn =

K 1 q =0

xl + q

q ( N 1)q
q!

(6)

By introduce the l-degree polynomial dependent filter gain hl(i, N), then the estimate of the electronic image xn is obtained by an averaging horizon of N points trough the convolution. In addition, represents the sampling time and q is the state of signal. 2. Unbiased FIR filter The unbiased FIR filter estimate is provided by the discrete convolution as follows

xn =

N 1 i =0

hl (i, N ) yni

(7)

where, hl(i, N) is the impulse response with order approximation l, the constant N is the averaging horizon of the samples and yn is the measurement signal. Moreover, the unbiased estimates gain hl(i, N) must accomplish the following conditions
N 1 i =0

hl (i, N ) = 1

N 1

and
i =0

hl (i, N )i u = 0 u [1, l].

(8)

The constraints (8) mean that the response of filter produces a unity response in the case of the input ramp signal and that the response at the zero frequency is equal to zero, respectively. It is known from optimal filtering theory that the order of the optimal (and so unbiased) filter is the same as the system. Thus, for the K state model, the gain can be represented with the l-degree polynomial such that

hl (i, N ) =

l j =0

a jl ( N )i j

(9)

where, ajl(N) are the polynomial coefficients and the degree l must be chosen such that l = K 1. The coefficients for the polynomial (9) have been found in [4]

a jl ( N ) = ( 1)

M ( j +1)1 ( N )

D( N )

(10)

The short (l + 1)(l + 1) symmetric matrix D(N) is given by

L. J. Morales-Mendoza et al. / Procedia Engineering 35 (2012) 202 209

205

D( N ) =

d 0 (N ) d1 ( N )

d1 ( N ) d 2 (N )

d l (N ) d l +1 ( N ) d 2l (N )

(11)

d l ( N ) d l +1 ( N )

where |D| and M(j+1)1 are the determinant and the minor of (10), respectively. The component of (11) can be determined using the Bernoulli polynomials Bn(x) as follows

d r ( N ) = 1 [Br +1 ( N ) Br +1 ] .
r +1
2.1. Low Degree Polynomial impulse response

(12)

In FIR filtering, an estimate is obtained via the discrete convolution applied to measurement. That can be done if to represent the state space model on an averaging interval of some N points. According to (9), the low-degree polynomial gains can thus be defined with

h1i ( p ) =

3 j =0

a ji ( p )i j .

(13)

Now, based on explained above and invoking previous filter design analysis for this research in [7] [10], the recursive scheme algorithm is derived to particular cases. These are oriented to cover an overwhelming majority of practical needs with relevant singular gains, i.e., uniform, linear, quadratic and cubic model as follows,

1 N 2(2 N 1) 6i h1 (i, N ) = N ( N + 1) h0 (i, N ) =

(14)

(15)

h2 (i, N ) = h3 (i, N ) =

3 3N 2 3N + 2 18(2 N 1)i + 30i 2 N ( N + 1)( N + 2)

(16)

8 2 N 3 3N 2 + 7 N 3 20 6 N 2 6 N + 5 i + 120(2 N 1)i 2 140i 3 N ( N + 1)( N + 2)(N + 3)

(17)

3. Recursive Scheme of UFIR Recursive forms processing algorithms are an alternative way to improve some features within a process by increasing the computational complexity required. This principle is used in both adaptive filtering and neural networks. Here, the problem is to enhance a signal or an image using the UFIR filters responses to preserve the edges contained in it, respectively. Each of the responses of the UFIR filters defined in (14) - (17) can be transformed recursively using the following form

206

L. J. Morales-Mendoza et al. / Procedia Engineering 35 (2012) 202 209

N ( k )1

xn =

i =0

hl (i, N (k )) y ni 0

if k [K min , K max ] otherwise

(18)

where, Kmin and Kmax are the range where the UFIR filter will perform the recursion. For the particular case, the UFIR filter with ramp response, we set (18) in (15) obtaining the next expression

h1 (i, N (k )) =
4. Simulation

2(2 N (k ) 1) 6i N (k )(N (k ) + 1)

(19)

In this section, we present some quantitative and qualitative evaluations for testing computational recursive UFIR filter. We employed two special cases, in the first case uses a signal of 512 samples and for the second case, an artificial image of kc = 512 by kr = 512 pixels. To achieve the numerical evaluation we apply two classical quantitative metrics: the SNR and RMSE metrics. The first and second metric are defined as follows,
kc kr

SNRdB = 10 log10

kc

i =1 j =1 kr

[x(i, j )]2
(20)
2

i =1 j =1

[x(i, j ) x(i, j )]

RMSE =
4.1. First Case: Signals

1 kc kr

kc

kr

i =1 j =1

[x(i, j ) x(i, j )]2

(21)

In this case, we proposed a square pulse as shown in Fig. 1 (upper). This signal is contaminated with white noise using a variance of 0.4. In Fig. 1 (lower) shows the noisy signal. The approximation order of FMH and recursive UFIR filters are lineal (i. e. ramp gain) and be defined in (15) and (19), respectively. For the simulations and qualitative/quantitative comparisons, we employed the FMH filter with N = 11 samples as a horizon. In the Fig. 2, we show the enhanced of the FMH filter (upper) and the recursive UFIR filter (lower), respectively. The quantitative evaluation is shown in the Table I.
Table I. Quantitative Evaluations SNR (dB) FMH filter h1(N) 16.7052 RMSE 0.0846

Recursive UFIR filter

h1(N)

17.5231

0.0770

L. J. Morales-Mendoza et al. / Procedia Engineering 35 (2012) 202 209

207

Fig. 1: Square pulse: (upper) original signal; (lower) noisy signal.

Fig. 2: Square pulse processed: (upper) enhanced signal with FMH filter; (lower) enhanced signal with recursive UFIR filter.

4.2. Second case: images On the other hand, in the case of images (2D signals), to apply the recursive UFIR filters must be lexicographic form to process each of the columns/rows with the filter. The artificial image used to test the performance of recursive UFIR filter is shown in Fig 3(a). This image is the 2D transformation of the signal used in the first case (square pulse signal). The noises added to image are the white noise and speckled noise with a variance of 0.2 each one. The noisy image is shown in Fig. 3(b).

208

L. J. Morales-Mendoza et al. / Procedia Engineering 35 (2012) 202 209

For the simulations and qualitative/quantitative comparisons, we employed the FMH filter with N = 11 samples as a horizon. The image process by FMH filter with ramp gain is shown in Fig. 4(a). On the other hand, in the case of recursive UFIR filter we obtained as a result the image displayed in Fig. 4(b). The quantitative evaluation is present in the Table II. We can see that the recursive UFIR filter obtained the better performance that the FMH filter in the sense of the SNR and RMSE.

Fig. 3: (a) Original image; (b) Noisy image

Fig. 4: Enhanced image using a ramp response apply (a) FMH filter; (b) Recursive UFIR filter Table II. Quantitative Evaluations Gain Ramp FMH filter SNR (dB) 12.67 RMSE 0.102

Recursive UFIR filter

16.54

0.095

L. J. Morales-Mendoza et al. / Procedia Engineering 35 (2012) 202 209

209

5. Concluding Remarks This article presented a suitable way to derive UFIR filters focused in signal and image processing on enhancing. This was derived as a new recursive scheme algorithm, with some particular sub-models, that guarantees the SNR-and-RMSE-based outstanding computing approach (reported in Tables I and II). Furthermore, qualitative and quantitative computing comparison was realized with the recursive UFIR filter versus the FMH filter. The elemental drawback of this recursive algorithm is the high computational complexity required by the software of the algorithm. To future work, the computational complexity will be reduced via parallel computing of the UFIR filter. Acknowledgements The first author thanks to PROMEP and the FIEC Veracruz University for the whole research support. References
[1] [2] [3] [4] [5] [6] Pitas I. and Venetsanopoulos A. Nonlinear Digital Filters Principles and Applications. Boston: Kluwer academic Publisher; 1990. Najarian K. and Splinter R. Biomedical Signal and Image Processing. New York: Taylor & Francis, 2006. Jensen JR. Introductory Digital Image Processing A Remote Sensing Perspective. 2da Ed. New Jersey; Prentice Hall, 1996. Heinonen P. and Neuvo Y. FIR-median hybrid filters with predictive FIR substructures. IEEE Trans. on ASSP 1988; 36:892899. Shmaliy Y. An unbiased FIR filter for TIE model of a local clock in applications to GPS-based timekeeping, IEEE Trans. on UFFC 2006; 53:862-870. Morales-Mendoza LJ. Shmaliy Y. Ibarra-Manzano OG. Arceo-Miquel LJ. and Montiel-Rodrguez M. Moving average hybrid FIR filter in ultrasound image processing. In: Vzquez-Gonzlez JL, editor. Electronics, Communications and Computers, Puebla: IEEE Computers Society; 2008, p. 160-164. Morales-Mendoza LJ. and Shmaliy Y. Moving average hybrid filter to the enhanced ultrasound image processing. IEEE Trans. on AL 2010; 8: 9-16. Shmaliy Y. GPS-based optimal FIR Filtering of Clock Models. New York; NOVA Science Publishers, 2009. Shmaliy Y. and Morales-Mendoza LJ. FIR smoothing of discrete-time polynomials signals in state space. IEEE Trans. on SP 2010; 58: 2544-2555.

[7] [8] [9]

[10] Morales-Mendoza LJ. Shmaliy Y. and Ibarra-Manzano OG. Enhancing ultrasound images using hybrid FIR structures. In: Chen Y, editor. Image Processing, Croatia: INTECH; 2009, p. 287-310. [11] Morales-Mendoza LJ. Shmaliy Y. Vzquez-Bautista RF. Ramrez-Carpio R. and Justo-Martnez OM. An analysis of key points of p-lag median hybrid FIR filter in applications to image processing. In: Alejos-Palomares R, editor. Electronic, Communications and Computers, Puebla: IEEE Computer Society; 2011, p. 194-199. [12] Kalouptsidis N. and Theodoridis S. Adaptive System Identification and Signal Processing Algorithms. New York; Prentice Hall, 1993.