Sei sulla pagina 1di 2

One simple improvement of the thermal impedance will be to bond these devices pside down.

Thus a flipchip bonding scheme is required to allow separate current injection into the gain and grating sections. Another option is a long grating section where the increased volume of the p n junction suppresses heating effects. The published results for buried heterostructure DBR lasers report CW tuning ranges of up to lOnm indicating that thermal problems do not dominate. Thus it remains to be Seen whether the thermal impedance in ridge waveguide DBR lasers can be reduced in a modified structure and hence achieve the wide tuning ranges reported for buried heterostructure DBR lasers.
Acknowledgments: The authors are grateful to S. Judge, T.


application of the filter itself to each colour band. Nevertheless such a technique adds new colours to the original image and this effect, in general, cannot be tolerated (for example when dealing with artwork pictures). This is why a generalisation of the usual scalar median lilter is proposed which works directly on colour vectors and gives better results in terms of spike noise suppression capabilities.

Fig. 1 Signal beforefiltering

Spooner, B. Quartermain, D. Ranasinghe, M. Harlow, I. Reid and W. Duncan for the fabrication of these lasers.

Given N vectors (say x1 ... xN). inside a rectangular window, the output of the vector median filter is defined by the following equations:
VM,{x, x2

n. SUNDARESAN I. D. HENNlNG British Telecom Research Laboratories, Martlesham Heath, Ipswich, IPS 7RE United Kingdom

24th July 1991

... XN}





.. XN}

1 KOCH, T. L., KOREN, U,,GNALL, R. P., BURRUS, c. A.,and MILLER, B. I.: Continuously tunable 1 . 5 ~ multiple-quantum-well GaInAs/ GaInAsP distributed-Bragg-reflector lasers, Electron. Lett., 1988, 24, pp. 1431-1433 2 h ~ u n ~ T *s., WTO, I., and KOBAYASHI, Tuning ranges for 1 . 5 ~ , K.: wavelength tunable DBR lasers, Electron. Lett., 1988, 24, pp. 511-519 3 SUNDAIWAN, H.,and FLETCHW, N. c.: Direct observation of shot noise in linewidth broadening of DBR lasers, Electron. Lett., 1990, 26, pp. 2004-2005

1 /IxYM- xillL 5 i=1 /Ixj - xillL i=


i = 1, 2 ... N



Indexing terms: Filters, Signal processing, Image processing
A new technique is described which couples median filtering and image deblurring techniques to filter noisy images without introducing defocusing side effects. To deal with colour images a vector median filtering procedure is proposed. Using this procedure a better edge preserving filter is obtained which does not introduce new colours. The deblurring operation is performed componentwise by fitting an ARMA model to the image. The AR part of the model estimates the image and the MA part estimates and blurring function. Finally the MA part is inverted and applied to remove the blur introduced by the median filter.

where V M , denotes the vector filtering operator. In practice, if a suitable L metric is stated, the output of the filter is the point in the window which minimises the sum of the L distances from the other N-1 points. If there are many points that satisfy eqn. 2 the output is determined according to their position in the window; a common solution is to favour choice of the central points. It is important to note that eqn. 2 depends on the metric used to calculate the distance between two points. The use of different metrics has two important effects: first a change in computational speed, and secondly a different final filtering quality. From the first point of view the squared Euclidean norm (L:)is the best one because a fast algorithm exists for its calculation. The L, norm is relatively fast because it uses only algebraic sums and absolute values. Finally the L, norm, which involves square root operations, is the slowest one. As regards the filter effectiveness, the previous ranking scale is reverted to, the L, norm being the best one and the L$ the worst. This is because the L$ norm is the closest to a simple average operation applied to the points in the window.

4. Ic

Fig. 2 Filtered signal Introduction: Colour image data bases are becoming more

and more important in many different application fields such

as in artwork acquisition and documentation, museum management, textile product data bases, high quality printing etc. A fundamental topic when dealing with colour image archives is fidelity to the original objects and scenes. An important point to this end is the development of suitable tools capable of filtering the images without blurring them and without changing their original chromatic contents. Classical noise reduction techniques are problematic in both these aspects because they smooth the original image and change its colour span. Even by using edge preserving filters, such as median filters, the final result is a badly smoothed image where fine spatial and chromatic details, which can be very useful in scientific analysis, are lost. For these reasons, in this letter a two step processing chain is presented which employs a vectorial extension of an edge preserving filter followed by a deblurring algorithm that compensates for the blur generated during the initial filtering step.
Vector median filtering: A straightforward generalisation of

Two major features of vector median filtering, as opposed to componentwise scalar filtering, must be emphasised: first it produces colour closed operations (i.e. no new colour is generated), and secondly it allows a better spike noise filtering. Colour closed operations are a direct consequence of the vector median filter definition, and its good filtering properties can be better understood considering the following example, that, for the sake of simplicity, refers to a 1-D signal in 2-D space. The signal consists of a step function with a noise spike in the first component of the vector. If the signal is filtered componentwise by means of a 5 x 5 pixel window, only the spike is shifted leading to an overlap between the original steps (Fig. 2). It is easy to see that this is not the case with an L, vector filter that completely eliminates the spike (Fig. 3).

Fig. 3 &filtered signal
r -


median filters to the colour image case is the componentwise


loth October 1991 Vol. 27 No. 21

Deblurring algorithm: It is very difficult to extend deblurring

procedures directly to a vector space; this is why a componentwise ARMA deblurring method has been adopted to compensate for the smoothing. The aim of the deblurring step is to estimate the point spread function (PSF) of the unknown degrading system. The image is described through a nonsymmetric half plane (NSHP) model

has been described. Such a result has been achieved by chaining a vector median filter and an ARMA deblurring technique. Vector median filters allow good results in terms of spike noise suppression and ensure colour closed operations.

s(m, n) =

(k. 1 ) E C

dk, M m - k, n - 0 + w(m, n )


where 0 II < M ) u ( - M 5 k IO, C = {(l I I , k M 1 5 l I ) } and w(m, n) is a zero mean white Gaussian noise M with variance U The observed image is a distorted version of : . Hm, n). If we model the distortion as an FIR noncausal system, the resulting image will be
r(m, n) =

h(i, j ) s ( m - i, n

11, j ) E C

- j ) + v(m, n )


,Fig. 5 Vector median filtered images before (left) and after (right) The deblurring part of the algorithm can be applied to the luminance component of the image if colour closed operations are required. In this Letter the blurring function has been assumed to be positive and symmetric to achieve four quadrant factorisation. When this hypothesis does not hold, only an approximation to the correct PSF can be obtained and this reduces the final quality of the deblurred image.

where C = { - H 5 i 5 H , - H I j I H} and v(m, n) is additive noise. If we assume that the additive noise is negligible, by substituting eqn. 3 into eqn. 4 we obtain

r(m, n) =

(k,I )

1 +

c(k, 0r(m - k , n - r)

h(i, j ) w ( m - i, n

-j )


(i. j ) E C

The observed image may then be modelled by an ARMA model, where the AR part is based on the image model parameters and the MA part is based on the blur parameters. The parameters and the order of the ARMA model must be determined. A least squares estimate of the parameters is achieved by searching the minimum of the quantity Zw(m, n). When the function H(o,, 2 )is decomposed into a four quadw rant fact~risation~~ previous minimum can be achieved the by means of a recursive procedure. This is possible if H(o,, w 2 ) 2 0 (e.g. a Gaussian blurring function). Once the parameters of the MA part have been determined, the deblurring step is completed by using such parameters in inverse filtering the observed image.
Experimental results: To evaluate the algorithm previously

15th July 1991


Uniuersita di Firenze, Dipartimento di Ingegneria Elettronica, Via S. Marta, 3-50139 Firenze, Italy


described (a vector median filter followed by an ARMA deblurring technique), some experiments have been performed by using art images. When dealing with artwork images it is mandatory to filter the acquisition noise without blurring the image themselves and without generating new colours; hence the proposed processing chain is particularly suitable.

HAAVISTO, W O , Y.: Vector median filters, P., and Proc. IEEE, 78, (4), April 1990 TEKALP, A. M., KAUFMAN, H., and WOODS, w.: J. Identification of image and blur parameters for the restoration of noncausal blurs, IEEE Trans., 1986, A S P - 3 4 , (4), pp. 963972 EKSTROM, M. P., and WOODS, Two-dimensional I. P.: spectral factorization with applications in recursive digital filtering, IEEE Trans., April 1976, A S P - 2 4 , pp. 115-128 ROGERS, D. F.: Procedural elements for computer graphics (McGraw Hill 1985).


Indexing terms: Gallium arsenide, Field-effect transistors, Semiconductor devices and materials
A new effect in planar GaAs MESFETs, whereby a sharp increase in optical gain at the transistor edges occurs, is reported for the first time. This gain effect only appears when

Fig. 4 Original image Fig. 4 shows the original image before noise filtering. By applying a 3 x 3 vector median filter the result shown in the upper left part of fig. 5 is obtained. Evidently some noise is removed and some blur is generated. By componentwise applying the ARMA deblurring algorithm outlined previously, the result shown in the upper right part of fig. 5 is obtained, where the blur effect is almost completely eliminated. Similar results have been obtained by using a 5 x 5 vector median filter (lower part of Fig. 5).
Conclusions: A technique which allows the filtering of noise in colour images with a significant reduction of the image blur

a large resistor is inserted in series with the gate, to produce the conditions for photovoltaic gate biasing. The mechanism for increased gain at the edges is suggested to be due to carrier photogeneration in the substrate that is subsequently collected by the gate. Application in the area of X-Y addressable transistor array imagers is proposed. The photoresponse of the GaAs MESFET has received much attention due to the potential application of the GaAs MESFET in high speed optoelectronic communications, OEICs and optical tuning of microwave devices. Various optical gain mechanisms have been reported, including photovoltaic gate biasing. This effect occurs when the gate photocurrent flows through an external series gate resistor R,, thus increasing the gate voltage and hence drain current. To produce a significant increase in drain current, a large R, 10th October 1991 Vol. 27 No. 21