1994 restoration of noisy scanning tunneling microscope images


Published on

Published in: Education
  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

1994 restoration of noisy scanning tunneling microscope images

  1. 1. Restoration of noisy Scanning Tunneling Microscope images G.M.P. van Kempenl, P.M.L.O. Scholte2, I.T. Young’, F. Tuinstra2 1Pattern Recognition Group, Trystallography Group, Faculty of Applied Physics, Delft University of Technology, P.O. Box 5046,2600 GA Delft, The Netherlands Abstract the difference between the restored image j ( x , y ) and We have compared and improved several the true image f(x,y). The optimal linear restoration implementations of the Wiener filter to remove noise filter, given g(x,y), is then known as the Wiener effects from Scanning Tunneling Microscope images. smoothingfilter [3] , which is given by We have found that the implementation of Weisman et al. [61, using the noise model of Stoll et al. 121, F(u, V ) F * (U,v) provides the best pelformance on both simulated and w(u, v) = F(u, V ) F * (U,v )+ N(u, V ) N * (U,v) real STM images. 1 1: Introduction - - 4 1+- Snn (U’ Since its invention by Binnig and Rohrer in 1981 [l], Sf (44 the Scanning Tunneling Microscope (STM) has proved where F(u,v) and N(u,v) the Fourier transforms of are to be a very useful instrument for surface science. A respectively the true image and the noise. SfJ(u,v) STM images the height of a surface on atomic scale, by denotes the power spectrum of the true image, S,,(u,v) measuring the current of tunneling electrons between of the noise. The restored image is given by the tip of the microscope and the surface, when a voltage difference is applied between them. Variations (3) in the tunneling current are imaged by the STh4 as variations of the height of the surface. with ?(u,v) the Fourier transform of the restored Unfortunately, images made with a STh4 suffer from image and G(u,v) of the observed image. Since the various distortions. We have found that the internal power spectrum of the true image and the noise are in noise of the STM is the main source of distortion of general not known, one has to use an estimate or a STM images made in air. This distortion is model of them. We have compared two models of the characterised by long stripes in the scan direction, and noise power spectrum with real noise measurements, has been characterised as so-called I/f-noise [2].This and tested five estimations of the true power spectrum type of noise is strongly correlated and has a self-similar on simulated images. frequency distribution. The removal of this internal Z - g From experiments [ 2 ] , Stoll and Marti have found like noise from STM images is the main goal of the that the power spectrum of STh4 noise is determined by research presented here. IJ-like noise at low frequencies. They believe that this noise is caused by very slow unintentional motion of the 2: Image acquisition models applied to STM tip perpendicular to the scan direction and by the electronic components of the STM’s feedback system. To model the acquisition of a STM image, we have Based on these assumptions they model the noise power assumed that the internal STM noise n(x,y) is additive spectrum as, and non-correlated with the “true” image f(x, y) of the surface, (4) g(x,y)=f(x,y)+n(x,y), (1) [ + (vl?L,)ilf where g ( x , y ) is the observed STM image. We have used the mean square error as a criterion to minimise 134 1051-465U94 $04.00 0 1994 IEEE
  2. 2. where L is the number of points per scan line. They , Finally, we have tested a modified version of the argue that the effect of Z/f-like noise is much more Wiener filter proposed by Weisman et al. [6][5], pronounced in the v- than in the u-direction, which is in 1 contradiction with the model of Park and Quate [4], who w(u, v) = 9 (7) modelled the noise power spectrum as NUB.From our CL4 own noise measurements 151 we have found that the P model of Stoll and Marti describes real STM noise more [U2 + (v/2Lx,’]~ accurately than the model of Park and Quate. 1+ To restore STM images, Stoll and Marti have used the following implementation of the Wiener smoothing in which we have used the noise model of Stoll and filter, Marti instead of the model of Park and Quate. 1 w(u,v) = 9 3: Simulations with the Wiener filter CL4 1+ R In our tests with simulated and real STM images, we have mainly used graphite as a test surface. Mizes et in which the true power spectrum is modelled with a1.[7] show that STM images of graphite are well the constant lla. We have estimated this power approximated by a summation of three cosine waves spectrum, by using equation [ 11, as the difference oriented at 120’ angles. between the observed image and the noise model of Since a good physical model for the internal STM Stoll and M r i This leads to the following formulation at. noise is lacking, we have used a simple method to of the Wiener smoothing filter [5], simulate I/f-like noise. The I/’-noise was generated by c A inverse Fourier transforming a function with a amplitude and a uniformly distributed random phase (fig. 111, right). 0 otherwise wiener3 test1 - - periods --o- 4.0 12.0 20.0 28.0 + 36.0 sim. graphite + 440 1 . 1 . 1 ’ I I 2.0 4.0 6.0 8.0 10.0 signal to noise ratio wiener3 (eg 6 ) Figure 1: The left picture shows the normalised mse between the true image and restored image of our Wiener implementation (eq. 6) as function of the signal to noise ratio between the true image and the noise for different numbers of the period of the graphite lattice, with a constant p of 1.50. The right image shows the filter performance of our Wiener implementation,with the true image (upper left), the noise realisation (upper right), the distorted image (down left) and the filter output (down right). The signal to noise ratio is 1.O, p 1.0 and the number of lattice periods is 8.0. The normalised mse of the filtered image is 0.095. 135
  3. 3. STM image wiener3 I e g . 6 1 wiener5 Ieg. 7 ) Figure 2: Wiener filtering of a real STM image. A STM image of graphite is shown, wiener3 denoted our implementation, wiener5 the modified filter of Weisman et al. with the filter parameter a of 25.0. We have tested five Wiener filter implementations: and have fit the amplitude of the noise on the spectra of the ideal Wiener filter (eq. [2]), the ideal filter with the real STM images. noise model of Stoll et Marti (to test this noise model), The results of the two Wiener filter implementations the implementation of Stoll and Marti, our differ from what we would have expected from the implementation and the modified Weisman filter. simulations. The modified filter of Weisman et al. has The performance of these Wiener filters have been performed significantly better then our implementation tested as function of the signal-to-noise (SNR) ratio (fig. 2). We have found that this difference in between the true image and the noise. The S N R ratio is performance can be explained by the deviations of the determined by the noise amplitude A , the p of Z- @ STM-noise from the Z/f-noise model of Stoll. These noise, and the period P of the graphite image. We have deviations of real STM-noise are, apparently, larger varied A , p and P so that they represent the range of than our simulated Z/f-noise. these variables in real STM images. The performance of the Wiener filters have been References determined by calculating a normalised mean square error, which is the mean square error between the [l] G. Binnig, H. Rohrer, Ch. Gerber and E. Weibel, “Surface restored and the true image, divided by the average studies by Scanning Tunneling Microscopy,” Physical Review Letters 49(1) (1982) pp.57-60. power of the true image. [2] E.P. Stoll and 0. Marti, “Restoration of scanning tunneling From our tests [ 5 ] with the ideal Wiener filter, we microscope data blurred by limited resolution, and conclude that the additive I/f-noise is well suppressed in hampered by l/f-like noise,” Surface Science 181 (1987) the simulated graphite images. The second filter gives pp.222-229. only a slightly worse performance than the first one. [3] A.K. Jain “Fundamentals of digital image processing”, Thus, the parametric model of Stoll for the power Prentice-Hall, Englewood Cliffs (1989). [4] Sang-il Park and C.F. Quate, “Digital filtering of scanning spectrum which is used in this filter, is a good model to tunneling microscope images,” Journal of Applied Physics describe the noise power spectrum. The performance of 62(1) (1987) pp.312-314. our Wiener implementation and the modified filter of [5] G.M.P. van Kempen, “Restoration of Scanning Tunneling Weisman et al. give good results, reducing the amount Microscope Images,” Master thesis, Delft University of of noise from 1.0% down to 0.1% of the original Technology, August 1993. amount of noise (fig 1, left). The filter of Stoll et al. [6] A.D. Weisman, E.R. Dougherty, H.A. Mizes and R.J. only reduces this amount by half. Dwayne Miller, “Nonlinear digital filtering of scanning- probe-microscopy images by morphological pseudoconvolutions,” Journal of Applied Physics 7 l(4) 5: Restoration of real STM images (1992). [7] H.A. Mizes, Sang-il Park and W.A. Harrison, “Multiple-tip We have applied our Wiener filter and the modified interpretation of anomalous scanning-tunneling-microscopy Weisman implementation of the Wiener filter to real images of layered materials,” Physical Review B 36(8) STM images. We have found that the stationary tip (1987) pp.4491-4494. [8] M. Aguilar, E. Anguiano, A. Diaspro and M. Pancorbo, images describe the noise in the real images “Digital filters to restore information from fast scanning qualitatively well, but that they underestimate the total tunneling microscopy images,” Journal of Microscopy power of the noise. We have therefore assumed that the 165(2) (1992) pp.311 - 324. shape of the noise has been conserved in real images, [9] K. Besocke, “An easily operable scanning tunneling microscope,” Surface Science 181 (1987) pp.145-153. 136