WAVELET FOR ULTRASONIC FLAW ENHANCEMENT AND IMAGE COMPRESSION W. Cheng, K. Tsukada, L.Q. Li, K. Hanasaki Department of Earth Resources Engineering, Graduate School of Engineering, Kyoto University, Kyoto, Japan, 606-8501 ABSTRACT. Ultrasonic imaging has been widely used in Non-destructive Testing (NDT) and medical application. However, the image is always degraded by blur and noise. Besides, the pressure on both storage and transmission gives rise to the need of image compression. We apply 2-D Discrete Wavelet Transform (DWT) to C-scan 2-D images to realize flaw enhancement and image compression, taking advantage of DWT scale and orientation selectivity. The Wavelet coefficient thresholding and scalar quantization are employed respectively. Furthermore, we realize the unification of flaw enhancement and image compression in one process. The reconstructed image from the compressed data gives a clearer interpretation of the flaws at a much smaller bit rate. INTRODUCTION Ultrasonic testing has been an important and one of the mostly used method in Non-destructive Testing and Evaluation (NDT&E). It utilizes some particular characteristics of the propagating stress waves [1], which broaden its application to a much larger scope, industry and medical field, etc. With its expanding usage, here rises the problems of getting more accurate flaw indication of the C-scan image, and of storing and transforming the image more efficiently as well. In this paper, we explore a method based on Discrete Wavelet Transform (DWT), to achieve flaw enhancement and image compression. Furthermore, we discuss the possibility and feasibility of solving the two problems in one process. The rest of this paper is outlined as follows. The main part is divided into section 2 and section 3, which discusses Wavelet in C-scan flaw enhancement and image compression respectively. A view looking into the combination of the two problems will be presented in section 4 before a conclusion is drawn in section 5. WAVELET IN C-SCAN IMAGE ENHANCEMENT Ultrasonic C-scan images are raster scanned displays which use gray-scale or pseudo-color to represent the reflected ultrasonic amplitude at each pixel in the image. The resolution of the ultrasonic C-scan image is often degraded by blur and noise. The size/shape or acoustic reflectivity of the flaw of low amplitude can be masked by noise. While blur does not greatly affect flaw detection, it does limit the ability to characterize CP657, Review of Quantitative Nondestructive Evaluation Vol. 22, ed. by D. O. Thompson and D. E. Chimenti © 2003 American Institute of Physics 0-7354-0117-9/03/S20.00 601 flaw size/shape, number, and spacing [2,3]. The solutions for these two problems at present include filtering (Wiener filter, etc), multi-sensor data fusion, and so on. Most of them depend on the setting of the experiment to acquire the data. Here we want to propose a post process method that mostly relies on the statistical properties of the data. Wavelet Method Multi-Resolution Analysis (MRA), as implied by its name, analyzes the signal at different frequencies with different time resolutions. MRA decomposes a given function/ into a series of spaces FmCL2(R), m • Z. These spaces describe successive approximation spaces of /with resolution 2~m [4]. A space of Wm orthogonal to Vm in space Vm-i is defined as details. , VmLWm (1) Then the original /can be represented by SmiM} (2) In wavelet transform, a scaling function 0 and a wavelet function If, are introduced, which r^(x) = 2-m/2r(2-mx-n)9 (m,fl)EZ 2 (3) the (^ w w ) w e z and (V m «)«ez constitute an orthonormal basis for Vm and Wm respectively. They are functions of the translation (time) parameter n and scale (frequency) parameter m [5]. The results by the multiplication of /with the two functions are defined as the approximation coefficient Am and the detail coefficient Dm. ^ =</,*.,.) (4) Dm=(f,Ym,,,) (5) The coefficients are an alternative mathematical representation of the signal / by which DWT easily restructures a signal statistically. The / can be losslessly represented by Equation (6), __ (6) 2-Dimensional DWT 2-dimensional DWT is based on one scaling function 0(t) and three wavelets W (where /c represents three spatial orientations: horizontal, vertical and diagonal ),which is the products of a one-dimensional 0ft) and Wft). From here we can see that the details extracted by the wavelets pick up the properties of the image in the three orientations. DWT decomposes an image into bands that vary in spatial frequency and orientation, which 602 0(x) <P(y)3 <P(x) W(y)3 V(x) <P(y)3 V(x) V(y) 3 <P(x) W(y)2 Y(x) <P(y)2 V(x) V(y)2 <P(x) V(y)l V(x) V(y)l Y(x) <P(y)l FIGURE 1. Two-dimensional DWT facilitates treating the image data differently according to different frequency and orientation. The 2-D DWT is shown in Fig. 1 [4]. C-scan Image Simulation Model We construct the C-scan simulation image model as having both blur and noise: g(x, y)=f(x, y)*h(x5 a, y, £)+n(x, y) (7) where g(x, y) is the received image function, f(x, y) is the original image. The h(x, a , y, j3 ) is the point spread function associated with the ultrasonic system, expressing how much the input value at position (x,y) influences the output value at position( a , j3 ). The n(x, y) is additive noise. The symbol * denotes linear convolution. The noise is assumed to be stationary and uncorrelated with the signal [2,5]. Wavelet Implementation in Image Enhancement Here 'enhancement' refers to both de-noising and de-blurring. The scheme of the implementation is outlined in Fig. 2. Here we take advantage of the DWT properties, namely, multi-resolution and orientation selectivity, to set soft thresholds A (the coefficients larger than A in absolute value are decreased by A if positive and increased by A if negative [6]) to the DWT coefficients. The criteria used to evaluate how the method works on eliminating blur and noise are SNR (Signal Noise Ratio = 101ogio[Var(signal)]/[Var(noise)]) and MSB (Mean Square Error between the image and the original image). The experiments show that the wavelet method can not only eliminate the blur and noise respectively, but also work well in cases that have both of them. Fig. 3 gives an example of the case. FIGURE 2. Wavelet in C-scan image enhancement 603 (a) (b) FIGURE 3. Wavelet in C-scan flaw enhancement, (a) flaw image with noise and blur. SNR=11.2216dB, MSE=0.6605, (b) the processed image of (a) by wavelet enhancement method, SNR=65.2826dB, MSE=0.5456 WAVELET IN IMAGE COMPRESSION Compression Process There are requirements for image compression and reasons why image can be compressed. First, there is redundancy in the digital images, which comes form statistical dependencies between pixels. Second, certain image degradations are not perceivable to human visual system [7]. In the specific case of ultrasonic C-scan image, the flaw edges are the most prominent feature that should be kept while some others can be omitted. Image data are transformed by wavelet to de-correlate the data as much as possible. Such transform is exactly invertible, therefore, lossless [6]. The edges are large coefficients of high frequency scales, which is only a small portion of the whole coefficient set. By emphasizing those coefficients that have much information about the flaw and by discarding those that don't have, compression is realized. The method used here is 'scalar quantization'. The compression scheme is outlined in Fig. 4. Selection of Wavelet Base When selecting a wavelet base to do image compression, some basic requirements of it should be considered. They include those mentioned in [4,8], such as avoiding distortion and artifacts in the reconstructed image, rapid convolution in DWT, smoothness of the reconstructed image, and so on. Unfortunately, not all of the requirements can be satisfied simultaneously. The highly important linear phase constraint corresponding to symmetrical wavelets can be maintained by relaxing the orthonormality and using biorthogonal bases [4]. The biorthogonal wavelets are more often used because of their relatively short compact support and best trade-off between regularity and vanishing moments [4,9]. By far, choosing wavelets for image com- Decompression Compressed data -* Decoder FIGURE 4. Wavelet method of compression 604 Q(x) xO xl x2 *M <x<Xi x3 y4 x4 x5 x6 x7 y3 (yi> x<xn yi FIGURE 5. A Non-uniform spatial quantizer pression is still a somewhat inexact science. Based on the overall properties and experiments, Bior4.4 (symmetric bi-orthogonal wavelet with 4 vanishing moment, linear phase) is selected. DWT Coefficients Quantization As shown in Fig. 4, the key part of compression is 'Quantization'. Quantization refers to the process of approximating the continuous set of original values in the image data with a finite (preferably small) discrete set of values. Besides, the precision of the floating point DWT output is also quantized by this step. Obviously, this is an irreversible, lossy part of the compression. For the sake of computational simplicity, Scalar Quantization, in which each input coefficient is treated separately in producing the output, is employed here. We carry it out by two approaches: spatial and spectral. Fig. 5 explains how a Non-Uniform spatial Scalar Quantizer with input range divided un-equally works. For common images, there is a method called Max Quantizer [4,10] to part the input quantization steps. It is on the basis of the probability distribution of the variable to be quantized. However, considering the flaw edge is mostly concerned in NDE, we base our method on the amplitude of the DWT coefficients. Large coefficients embody the flaw edges. To get a more concise view of this questioned parts, the steps between bigger coefficients are more intensive than those between smaller ones. Those whose values are among the smallest and thus set to 0 by quantization, is the largest in number. The spectral approach, what we call 'Weighting', puts weights to different frequency bands, shown in Fig. 6. Coefficients of some frequencies are emphasized by multiply a large 'weight'. They are mostly high frequency ones caused by flaw edges and those that human visual system is mostly sensitive to. Some are overlooked, and some others are given weight '0', which mostly are of low frequencies. The reconstructed image only uses the weighted coefficients of a particular frequency or frequencies [11]. In both approaches, by discarding these '0' pixels, compression can be realized. The Original DWT Coefficients W^D, W 2 *D 2 D2 Weighted Coefficients Weighting wm FIGURE 6. Weighting —-one Spectral Quantization method 605 FIGURE 7. Reconstructed image after compression of Fig.3(a) (PSNR=18.593, NCR=8bpp). (a) by Uniform Quantization, quantization step No.=40, PSNR=14.455,NCR=0.0453bpp. (b) by Weighting, PSNR=15.371, NCR=0.0769bpp setting of the quantization step and weight in this process is trial and error. A mathematical model for the function of this setting with frequency, orientation, compression rate and display resolution will be a challenging task. The quality criteria are PeakSNR (taking the peak value of 255 of the original signal instead of its variance as the numerator in classical SNR) and NCR (nominal compression rate, the unit is 'bpp', bit per pixel)[4]. Some examples are displayed in Fig. 7. DISCUSSION ON TRADE-OFF AND UNIFICATION Thus far, we have explored the DWT ability to do ultrasonic C-scan flaw enhancement and image compression respectively. The ability is derived from the fact that the wavelet transform gives the image an alternative mathematical representation by DWT coefficients. Because they carry the information about time, frequency and orientation, these coefficients are easier to be sorted and handled than the original gray level value. In both the cases of enhancement and compression, one thing shared in common is that some coefficients are set to zero, though by different rules. However, it can be found that there is overlapping. More explicitly, some coefficients are small both in amplitude and frequency, which are set to zero in both cases. From this point, we try to explore how to realize the two tasks in one process. It is obvious that there is trade-off between image enhancement and compression. The threshold should be enlarged to get a higher compression ratio but at the expense of clarity. Compression will be sacrificed in increasing the quantization steps to get a clearer reconstructed image. Still, for the sake of image enhancement, the weighting should keep the low frequency though the compression ratio will be lowered. Fig. 8 shows the result of this process on the image given in Fig. 3(a). The effect of the enhancement is not as good as Fig. 3(b), and the compression is less optimal than Fig. 7. FIGURE 8. Unification and trade-off of compression and enhancement result of Fig. 3(a). PSNR=54.8173, NCR=0.1796bpp 606 But it realizes both of them in one process within the tolerance of NDE purpose. The data of the criteria is acceptable and satisfactory. CONCLUSION The experiments prove that by proper soft thresholding, DWT can enhance the Ultrasonic C-scan flaw image. The algorithms of non-uniform spatial quantization and weigh quantization perform well in the C-scan image compression. We realize the unification of flaw enhancement and image compression in one process. The reconstructed image from the compressed data gives a clearer interpretation of the flaws at a much smaller bit rate. REFERENCES 1. Bray, D.E., Stanley, R.K., Nondestructive Evaluation, a Tool in Design, Manufacturing and Service, McGraw-Hill Book Company, New York, 1989, pp.47 2. Howard, P.J., Gilmore, R.S., Ultrasonic C-scan Image for Hard-Alpha Flaw Detection and Characterization, Review of Progress in Quantitative Nondestructive Evaluation, Vol. 13A, edited by Thompson, D.O. and Chimenti, D.E., PLENUM PRESS, New York, 1994,pp.763-770 3. Hosur, M.V., Estimation of Impact-Induced Damage in CFRP laminates Through Ultrasonic Imaging, NDT&E International, Elsevier Science Ltd, Great Britain, Vol.31, No.5, 1998,pp.359-374 4. Barlaud, M., Advances in Image Communication 5 Wavelet in Image Communication, ELSEVIER, Amsterdam, 1994, pp.66-95, pp.165 5. Petrou, M., Bosdogianni, P., Image Processing, the Fundamentals, JOHN WILEY&SONS, LTD, Chichester, 1999, pp.6-7, pp. 155-157 6. Walnut, D.F., An Introduction to Wavelet Analysis, Birkhauser Boston 2002, pp.374, pp.371, pp.387 7. Ramstad, T.A., Aase, S.O. and Hus (£> y, J.H., Subband Compression of Images: Principles and Examples, ELSEVIER, Amsterdam, 1995, pp.5 8. Rioul, O., On the Choice of' Wavelet' Filters for Still Image Compression, IEEE International Conference on Acoustics, Speech and Signal Processing, Image and Multidimensional Signal Processing, Volume V of V, 1993,IEEE, New York, 1993, pp.V-550-V-553 9. Hallat, S., A Wavelet Tour of Signal Processing, Academic Press Limited, San Diego, 1998, Capt 10 10. Gersho,A., Gray, R.M., Vector Quantization and Signal Compression, Kluwer Academic Publishers, Dordrecht ,1992 11. Polikar, R., The Wavelet Tutorial, http:// www.public.iastate.edu 607
© Copyright 2025 Paperzz