A Data-Fusion Scheme for Quantitative Image ...

2554

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 57, NO. 11, NOVEMBER 2008

A Data-Fusion Scheme for Quantitative Image Analysis by Using Locally Weighted Regression and Dempster–Shafer Theory Zheng Liu, Senior Member, IEEE, David S. Forsyth, Member, IEEE, Mir-Saeed Safizadeh, and Abbas Fahr

Abstract—Dempster–Shafer (DS) theory provides a solution to fuse multisensor data, which are presented in a hypothesis space comprising mutually exclusive and exhaustive propositions and their unions. The fusion result is a description of the proposition with the values of support, plausibility, and uncertainty interval. However, in some applications, numerical values of a continuous function, instead of a Boolean value or a proposition, are expected. In this paper, a scheme based on DS reasoning and locally weighted regression is proposed to fuse the data obtained from the nondestructive inspections of aircraft lap joints for the estimation of the remaining thickness. The proposed approach uses a pairwise regression that is optimized by the DS method when multiple inputs are involved. The scheme is evaluated with the experiments on fusing conventional eddy current and pulsed eddy current data obtained from aircraft lap joint structures for the characterization of hidden corrosion. Index Terms—Classification, corrosion quantification, data fusion, Dempster–Shafer (DS) theory, local weighted regression.

I. I NTRODUCTION

T

HE LAP JOINT is one of the common elements across varied aircraft models. A typical lap joint structure consists of multiple sheets of overlapped and riveted aluminum. If the sealants or protection systems break down, corrosion may happen to the faying surface of these sheets. This type of corrosion is also known as hidden corrosion. According to Fahr et al. [1], the material loss by layer can serve as one of the corrosion metrics for structural integrity and life-prediction analysis of aircraft lap joint structures. This metric is obtained by estimating the remaining thickness of each layer. Multiple nondestructive inspection techniques were carried out to quantify the aircraft hidden corrosion [2], [3]. The purpose of corrosion quantification study is to generate a remaining-thickness Manuscript received November 9, 2007; revised March 11, 2008. First published May 28, 2008; current version published October 10, 2008. This work was supported by in part of the Defence Research and Development Canada, the Air Vehicles Research Section, and the National Research Council (NRC) Canada. Z. Liu is with the Institute for Research in Construction, National Research Council Canada, Ottawa, ON K1A 0R6, Canada (e-mail: zheng.liu@nrc-cnrc. gc.ca). D. S. Forsyth is with the Nondestructive Testing Information Analysis Center, Texas Research Institute Austin, Austin, TX 78746 USA. M.-S. Safizadeh is with Iran University of Science and Technology, Tehran 16846-13114, Iran. A. Fahr is with the Institute for Aerospace Research, National Research Council Canada, Ottawa, ON K1A 0R6, Canada. Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TIM.2008.924933

map of aircraft lap joints with improved accuracy from multimodal nondesctructive inspection (NDI) data [1], [4]. The Dempster–Shafer (DS) theory has been used to fuse the data from multiple sensing modalities in various applications. The uncertainty of the measurements is modeled with the probability mass function, which defines a value between zero and one, i.e., the basic probability assignment (BPA) to indicate the degree of support for a proposition [5], [6]. Both the flexibility and the difficulty consist of assigning the probability mass value to the propositions. In the DS theory, the frame of discernment θ consists of a finite set of propositions, which are mutually exclusive and exhaustive. Such a proposition is defined as a Boolean value or as a predefined class number to describe a certain condition or status. However, in practical applications, we may need to employ a numerical value of a continuous function to represent the final result. For example, the remaining thickness in the inspection of aircraft lap joints, which is represented by a numerical value, is employed to characterize the corrosion damage. Such measurement serves as one of the corrosion metrics for structural integrity and life-prediction analysis [1]. Conventionally, the corrosion quantification is achieved by comparison of the specific inspection results with a predetermined calibration curve. Only the first-layer corrosion can be characterized with the calibration method; however, error is introduced due to the superimposed corrosion between multiple layers. When the DS fusion approach is considered, the definition of probability mass function varies with the applications. Horn and Mayo [7] used both eddy current and ultrasonic testing to inspect pressure tubes in a nuclear power plant. The through-wall extent (depth) was determined from the multiple NDI measurement data. They generated the probability density function P (amplitude|depth) by fitting the NDI amplitude and flaw depth data with a normal distribution. When a DS-updating mechanism was applied, the probability mass function was defined as P (depth|amplitude), which was obtained by Bayes’ theorem. This generated a very similar result to that of Bayesian inference. Zhu et al. [8] used fuzzy c-means clustering to determine the mass function for each pixel of multiple magnetic resonance images of the human brain. The DS fusion rule was employedtosegment graymatter, whitematter, cerebrospinal and multiple sclerosis lesions, and background. Francois et al. [9] calculated the evidence mass according to the local and global amplitude distribution around each voxel from ultrasound and radiography measurements.

0018-9456/$25.00 © 2008 IEEE

Authorized licensed use limited to: IEEE Xplore. Downloaded on November 17, 2008 at 08:22 from IEEE Xplore. Restrictions apply.

LIU et al.: DATA-FUSION SCHEME FOR QUANTITATIVE IMAGE ANALYSIS BY USING LWR AND DS THEORY

Fig. 1.

2555

Lap joint from a service-retired Boeing 727 aircraft.

Gros et al. implemented the pixel-level fusion of eddy current testing (ET) images and the infrared image with the DS method in [10]. Different BPA values were tested and verified in the experiments. In this paper, a scheme to fuse conventional eddy current (ET) and pulsed eddy current (P-ET) data for quantifying the hidden corrosion in aircraft lap joints is proposed. Both techniques can provide information on the material loss [2]. The fusion of the two can achieve an improved estimation. The measurement data are first classified by a set of trained classifiers. The classification results are fused based on the DS fusion rule. The fusion results, which are represented by a thematic map, indicate the range of the remaining thickness. These outputs are further quantified with the locally weighted regression (LWR) method. The fusion result is assessed based on the X-ray “ground truth” data from the teardown inspection of the lap joints, and the effectiveness of the proposed approach is demonstrated. The rest of this paper is organized as follows: The description of the NDI methods and the lap joint used in the experiment is presented in Section II. The details about the proposed data fusion architecture are given in Section III. Section IV presents the experimental results. Discussion of the work and a conclusion can be found in Sections V and VI, respectively. II. N ONDESTRUCTIVE I NSPECTION OF A IRCRAFT L AP J OINTS A two-layer lap joint cut out from a service-tired Boeing 727 aircraft below the cargo floor (see Fig. 1) was inspected by multifrequency eddy current testing (at frequencies of 5.5, 8, 17, and 30 kHz) and pulsed eddy current testing. Eddy current inspections are commonly used during aircraft maintenance [11]. An alternating current can generate an induced magnetic field opposing the excitation field. The change of coil impedance reflects the electrical, magnetic, and geometrical properties of the specimen. The relation between the excitation frequency √ and the penetration depth is formulated as δ = 1/ πf μσ,

Fig. 2. NDI data and X-ray thickness map for the first layer. (a) ET 17 kHz. (b) ET 30 kHz. (c) P-ET LOI. (d) X-ray thickness map.

where δ is the penetrating depth at exciting frequency f , with the other two parameters μ and σ related to the magnetic permeability and electrical conductivity of the material under testing, respectively. In contrast to the conventional ET technique, the P-ET technique employs a broadband pulse, instead of a continuous wave, as the excitation signal [12], [13]. Therefore, the P-ET measurement may achieve more information related to depth. A time-domain feature named liftoff intersection (LOI), which is insensitive to liftoff variations, is used for analysis in this paper [12]. After nondestructive inspections, the lap joints were dissembled and cleaned of all corrosion products. The ground truth data or the remaining thickness map were obtained by using the digital X-ray mapping technique on each layer. The lap joint specimen was cut into seven sections labeled A–G. In the following experiment, data from section D are used for training, whereas those from section C are employed for testing. The lap joint is shown in Fig. 1. In this paper, we present the fusion result for the first layer only, because the calibration result for the first layer was available for comparison. The NDI measurements and the X-ray thickness map are given in Fig. 2. The magnitude of the

Authorized licensed use limited to: IEEE Xplore. Downloaded on November 17, 2008 at 08:22 from IEEE Xplore. Restrictions apply.

2556

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 57, NO. 11, NOVEMBER 2008

Fig. 3. General data-fusion procedure. TABLE I DEFINITION OF INFORMATION CLASSES IN TERMS OF MATERIAL LOSS

imaginary part of ET data at 17 and 30 kHz and the magnitude of P-ET LOI are used to characterize the first-layer thickness. The unit of the ET and P-ET measurements is voltage, whereas the X-ray map is represented in terms of thickness (in inches).

III. D ATA -F USION S CHEME A. Procedure The flowchart in Fig. 3 depicts the procedure of the proposed data-fusion scheme. The estimation is carried out based on the previous available data, which are named the training data. First, the number of clusters for the training data is found by using fuzzy c-means clustering and cluster-validation indexes [14], [15]. The nearest mean classifiers are assigned to the ET and P-ET data according to the cross-validation test results. The labeled training data obtained from data clustering are used to train the specific classifiers for the ET and P-ET inputs. The classified results determine the “data class,” whereas the labeled X-ray data set defines the “information class.” The X-ray data are labeled according to the thickness range, as given in Table I. The probability mass function is defined based on the relation between these two types of classes. When a prediction is carried out on the vector of new NDI measurements, the trained classifiers are applied to assign the inputs a predefined data class number. The probability

mass value is derived from the probability model established during the training process. The DS combination rule fuses the probability values, and a decision is made on the maximum mass output. The input vector is assigned a “corrosion type” Ci (j = 1, 2, . . . , N ). The regression is implemented with only the corresponding data points defined by information class Ci . In other words, only the X-ray and NDI data that correspond to the specified information class Ci are involved in the regression process. Thus, a continuous estimation, together with a belief value, can be obtained. The information class Ci defined in this application is disjointed. In other words, there is no intersection between the classes. Therefore, the simplest form of the DS method is employed. Bayesian inference could probably be another option, but we do not know P (Cj ), which is prior knowledge of the corrosion damages. An average value for P (Cj ) does not always assure a good result. We will not have such a problem with DS fusion once the mass function is properly defined. B. Definition of Probability Mass Function and DS Fusion Assuming that Ci (i = 1, 2, . . . , N ) represents the infor mation class (corrosion types) and x is the vector of the measurement values, the mass value can be defined as the probability of being a certain information class based on statistical information from available training data sets, i.e., ms (Ci ) =  P (Ci | x ) [16]. Herein, s indicates the different data sources

Authorized licensed use limited to: IEEE Xplore. Downloaded on November 17, 2008 at 08:22 from IEEE Xplore. Restrictions apply.

LIU et al.: DATA-FUSION SCHEME FOR QUANTITATIVE IMAGE ANALYSIS BY USING LWR AND DS THEORY

2557

the mass value is updated. The final decision is made according to the maximum belief, i.e., the maximum mass output. Therefore, the fusion result is represented by a thematic map at this stage. C. LWR Recalling the requirements of characterizing corrosion damage, we need to estimate the remaining thickness of a lap joint. Thus, LWR is employed to estimate the value of interest using information pertaining only to a neighborhood of the input  query. In our case, the query point is vector x q , which is a new NDI measurement. The fusion result, i.e., the information class number, determines which data set is used to get the LWR result of this query vector. The LWR method is a form of lazy learning or memory-based learning, which defers the processing of training data until a query needs to be answered [17]. The training data are stored in memory. Relevant data are found from the database when a query is presented. LWR uses locally weighted training to average, interpolate, extrapolate, or combine training data [18]. Supposing that variable y represents the material thickness, we have vector x ∈ m for NDI data and need to find a  mapping function f : m →  with the training data set {( x p , yp )}np=1 . Here, n is the number of corresponding training data. Thus 

y p = f ( x p ) + εp

Fig. 4. Probability model of NDI data. (a) Pulsed eddy current and (b) eddy current (17 and 30 kHz). 

(s = 1, 2, . . . , S). Now, input x is mapped to data class dj (j = 1, 2, . . . , M ) by a classification operation. Thus, the BPA is defined as ms (Ci ) = P (Ci |dj )Ps (dj |Ci ).

(1)

(3)

where εp is a random variable such that E[εp ] = 0 and  E[εp εq ] = 0 ∀p = q. Given query vector x q , to obtain the regression applicable to this query vector, the following cost function must be minimized:     n  2  | xp − xq |  f ( x p , β) − yp K J= (4) h p=1 where K(·) is a Gaussian kernel function for that weighs each output data sample. The weight function is determined by width  parameter h and the distance of the input value x p from query  vector x q , i.e.,       −( x p − x q )2 | xp − xq | 2 h . (5) K =e h

The first item P (Ci |dj ) gives the probability of a pixel belonging to class Ci when the corresponding NDI data are classified as dj . The second value Ps (dj |Ci ) is a measure of the capability of each data source to discriminate information class Ci . These two values can be derived from the NDI data models, which reveal the relation between data class and information class, as shown in Fig. 4. According to the DS theory, theBPA values must be normalized to meet the requirement i ms (Ci ) = 1 before the updating operation is applied. Following

A larger h gives a “smoother” weighting function. In our experiment, h was assigned a fixed value of 0.6. The solution to the preceding cost function (4) is [7]

mET,P−ET (Ck ) = mET ⊕ mP−ET (Ck )  Ci ∩Cj =Ck mET (Ci )mP−ET (Cj )  = 1 − Ci ∩Cj =φ mET (Ci )mP−ET (Cj )

Matrix X consists of data samples, with [ x p , 1]T in the pth row, and y = [y1 , y2 , . . . , yn ]T is the variable corresponding to the remaining thickness. The estimation is then given by

(2)

βˆ = (Z Z)−1 Z v

(6)

where Z = WX, and v = Wy. Wis a diagonal matrix,    with the ith diagonal element wii = K(| x i , x q |, x q |/h. 

  yˆ( x p ) = βˆ x p .

Authorized licensed use limited to: IEEE Xplore. Downloaded on November 17, 2008 at 08:22 from IEEE Xplore. Restrictions apply.

(7)

2558

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 57, NO. 11, NOVEMBER 2008

The fused classification results determine the data used for the LWR process. Therefore, the fusion result can further be quantified with (7). A numeric value can be obtained from the regression.

D. Fusion Result Evaluation To objectively assess the fusion results, a straightforward way is to compare with the X-ray reference. There are a number of metrics for image comparison when a perfect reference is available. The commonly used approaches include root-meansquare error (RMSE), peak signal-to-noise ratio (PSNR), and correlation (CORR). The advantage of these methods is the simplicity and computational efficiency. More sophisticated methods include difference entropy (DE) and mutual information (MI) [19]. The DE between two images reflects the difference between the average amount of information they contained. The MI between the input and reference images is defined on the normalized joint gray-level histogram and normalized marginal histogram of the two images. In this paper, we employed the five aforementioned metrics for the fusion result assessment.

Fig. 5. Classification results with the nearest mean classifier. (a) ET data. (b) P-ET data.

IV. E XPERIMENTAL R ESULT Data sets from the inspection of a lap joint from a serviceretired Boeing 727 aircraft shown in Fig. 1 are used in the experiment. Data from two sections of a lap joint D and C, as shown in Fig. 1, are used for training and testing, respectively. The ET data obtained at 17 and 30 kHz, together with P-ET LOI, are used for the first-layer thickness estimation. However, the NDI data, which are presented as 2-D images, are of different resolution. Before using them, we need to register them [20] so that the corresponding pixels in different NDI images are associated with the same physical point on the specimen. Eventually, the size of the images for sections D and C are 114 × 366 and 114 × 499, respectively. The number of points for each corrosion type (information class) used for training is listed in Table I. As the training data come from a lap joint section, the data points for each classes are different. To take full advantage of the available data, we used all the data points in the training process. The ET and P-ET data are clustered by applying the fuzzy c-means clustering algorithm. The initial cluster number is set from 2 to 14. The clustering partition index, separation index, Xie and Beni’s index, and Dunn’ s index are considered to select a proper cluster number for NDI data [14], [15]. As a consequence, the ET and P-ET data are clustered into seven and eight groups, respectively. The X-ray data are segmented into seven parts based on the percentage of material loss. The clustered/labeled NDI data are used to train classifiers. To find an efficient classifier, the cross-validation test is applied to several candidate classifiers. The one with the smallest error is selected. In the experiment, the nearest mean classifier is chosen to classify the ET and P-ET data. According to Section III-B, the probability functions are defined for the ET and P-ET data, respectively, as shown in Fig. 4.

Fig. 6.

(a) DS fusion result and (b) X-ray thematic map.

The classification results of the ET scan (17 and 30 kHz) and the P-ET LOI scan are shown in Fig. 5(a) and (b), respectively. Fig. 6(b) indicates the segmented X-ray thickness map. The corresponding definition of classes is given in Table I. The fused result of Fig. 5(a) and (b) is presented in Fig. 6(a) after a morphological processing. It is noticed that the DS fusion result is far from its X-ray reference. This is partly due to the accuracy of probability models, which are based on the available training data. The regression result and X-ray thickness map are shown in the top and bottom panels of Fig. 7. The fusion result is compared with the X-ray thickness reference in terms of a number of image comparison metrics listed in Table II, i.e., RMSE, CORR, PSNR, DE, and MI. The DS-based fusion, followed by the LWR process, achieves the best result. The output of the DS fusion shown in Fig. 8 indicates the degree to which we can trust the result. V. D ISCUSSION The new scheme presented in this paper achieves a better result than that reported in [4] in terms of the selected metrics.

Authorized licensed use limited to: IEEE Xplore. Downloaded on November 17, 2008 at 08:22 from IEEE Xplore. Restrictions apply.

LIU et al.: DATA-FUSION SCHEME FOR QUANTITATIVE IMAGE ANALYSIS BY USING LWR AND DS THEORY

2559

of the proposed fusion algorithm can further be improved in two ways: One is to improve the accuracy of the classifier; the other way is to build a better probability model. However, it does not mean that the performance increases with the accumulation of data, because the quality of the data is not assured. Therefore, the study on selecting and validating data to improve the fusion analysis through the use of the accumulated historical data would be an interesting topic for future work.

VI. C ONCLUSION

Fig. 7.

(a) LWR final result. (b) X-ray thickness map. TABLE II EVALUATION OF THE FUSION RESULT

In this paper, a data-fusion scheme based on the DS theory and LWR is proposed. The measurement value is optimally classified by fusing the classification results with the DS combination mechanism. The final numeric estimation of the remaining thickness of aircraft lap joints is achieved by applying LWR of the preclassified results. The data-fusion method achieves better estimation of lap joint thickness than the calibration method. For our future work, the use of multiple P-ET slices around the LOI point is considered. The data-fusion algorithm may be applied to these P-ET images for characterizing deeplevel corrosion. This will also help identify the contribution of the eddy current technique for the thickness estimation. The other work includes developing a method to validate measurement data and updating the models and classifiers for an improved estimation and characterization. R EFERENCES

Fig. 8.

Map of support for the fusion result.

Obviously, the fusion algorithm developed here can also be applied to the deep-layer thickness estimation. In that case, ET data acquired at lower frequencies should be used. The capability of the P-ET technique for revealing and discriminating deeper layer corrosion has not been fully explored in this paper. Only the LOI scan, which is a time-domain feature, is employed for the analysis. The paper by Lefebvre and Dubois [21] indicates that the LOI is not fixed and changes with the presence of different corrosion damages in terms of material loss. Thus, using single-point LOI scan might not be an optimal solution. In the proposed scheme, the classification error of the selected classifiers is not considered in the probability model. This value indicates how much we can rely on the results of a specific classifier and the probability of a measurement belonging to a certain class. On one hand, an iterative classification scheme might be helpful in reducing the classification error; on the other hand, this error can be built into the probability models. The classification-based approach presented in [22] provides a promising solution to identify corrosion at different layers, because identifying the location of corrosion is a typical classification problem. The classification-based approach leaves space for further improvement when new data are available. Currently, the quantification analysis relies on artificially prepared “damage” on a calibration specimen. The performance

[1] A. Fahr, D. S. Forsyth, and C. E. Chapman, “Survey of nondestructive evaluation (NDE) techniques for corrosion in aging aircraft,” Nat. Res. Council Canada, Winnipeg, MB, Tech. Rep. LTR-ST-2238, Oct. 1999. [2] D. S. Forsyth and J. P. Komorowski, Applications of NDT Data Fusion. Norwell, MA: Kluwer Academic, 2001, ch. NDT Data Fusion for Improved Corrosion Detection, pp. 205–225. [3] Z. Liu, D. S. Forsyth, M. S. Safizadeh, M. Genest, A. Fahr, and A. Marincak, “Fusion of visual and eddy current inspection results for the evaluation of corrosion damage in aircraft lap joints,” in Proc. SPIE, San Diego, CA, Mar. 6–11, 2005, vol. 5768. [4] Z. Liu, D. S. Forsyth, B. A. Lepine, S. Safizadeh, and A. Fahr, “Quantifying aircraft hidden corrosion by using multi-modal NDI,” in Review of Progress in Quantitative NDE, D. Thompson and D. Chimenti, Eds. Green Bay, WI: Amer. Inst. Phys., 2003, pp. 1355–1362. [5] D. L. Hall, Mathematical Techniques in Multisensor Data Fusion. Norwood, MA: Artech House, 1992. [6] R. Hamid, “An experimental data fusion model for multisensor system,” Ph.D. dissertation, New Mexico State Univ., Las Cruces, NM, 1989. [7] D. Horn and W. R. Mayo, “NDE reliability gains from combining eddycurrent and ultrasonic testing,” NDT E Int., vol. 33, no. 6, pp. 351–362, Sep. 2000. [8] Y. M. Zhu, O. Dupuis, V. Kaftandjian, D. Babot, and M. Rombaut, “Automatic determination of mass functions in Dempster-Shafer theory using fuzzy c-means and spatial neighborhood information for image segmentation,” Opt. Eng., vol. 41, no. 4, pp. 760–770, Apr. 2002. [9] N. Francois, “A new advanced multitechnique data fusion algorithm for ndt,” in Proc. 15th World Conf. NDT, Rome, Italy, Oct. 2000. [10] X. E. Gros, J. Bousigue, and K. Takahashi, “NDT data fusion at pixel level,” NDT E Int., vol. 32, no. 5, pp. 283–292, Jul. 1999. [11] F. Turner, Advanced Manual for Eddy Current Test Method. Gatineau, QC, Canada: Can. Gen. Stand. Board, 1986. [12] B. A. Lepine, J. S. R. Giguere, D. S. Forsyth, J. M. S. Dubois, and A. Chahbaz, “Applying pulsed eddy current NDI to the aircraft hidden corrosion problem,” in Proc. 5th NASA/FAA/DoD Conf. Aging Aircraft, Orlando, FL, Sep. 2001. [13] B. A. Lepine, J. S. R. Giguere, D. S. Forsyth, J. M. S. Dubois, and A. Chahbaz, “Interpretation of pulsed eddy current signals for locating and quantifying metal loss in thin skin lap splice,” in Review of Progress

Authorized licensed use limited to: IEEE Xplore. Downloaded on November 17, 2008 at 08:22 from IEEE Xplore. Restrictions apply.

2560

[14]

[15] [16] [17] [18] [19] [20] [21] [22]

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 57, NO. 11, NOVEMBER 2008

in NDE, vol. 21, D. Thompson and D. Chimenti, Eds. Melville, NY: Amer. Inst. Phys., 2002. A. M. Bensaid, L. O. Hall, J. C. Bezdek, L. P. Clarke, M. L. Silbiger, J. A. Arrington, and R. F. Murtagh, “Validity-guided (re)clustering with apllications to image segmentation,” IEEE Trans. Fuzzy Syst., vol. 4, no. 2, pp. 112–123, May 1996. X. L. Xie and G. A. Beni, “Validity measure for fuzzy clustering,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 13, no. 8, pp. 841–847, Aug. 1991. D. S. Forsyth, Z. Liu, J. P. Komorowski, and D. Peeler, “An application of NDI data fusion to aging aircraft structures,” in Proc. 6th Joint FAA/DoD/NASA Conf. Aging Aircraft, San Francisco, CA, Sep. 2002. C. G. Atkeson, A. W. Moore, and S. Schaal, “Locally weighted learning,” Artif. Intell. Rev., vol. 11, pp. 11–73, Apr. 1997. M. Birattari, G. Bontempi, and H. Bersini, “Lazy learning meets the recursive least squares algorithm,” in Proc. Conf. Adv. Neural Inf. Process. Sys. II, 1999, pp. 375–381. Z. Xue and R. S. Blum, “Concealed weapon detection using color image fusion,” in Proc. 6th Int. Conf. Inf. Fusion, Queensland, Australia, Jul. 2003. Z. Liu and D. S. Forsyth, “Registration of multi-modal NDI images for aging aircraft,” Res. Nondestruct. Eval., vol. 15, no. 1, pp. 1–17, Jan./Mar. 2004. J. H. V. Lefebvre and J. M. S. Dubois, “Lift-off point of intercept (LOI) behavior,” in Review Progress Quantitative NDE, vol. 24, D. Thompson and D. Chimenti, Eds. Melville, New York: Amer. Inst. Phys., 2004. M. S. Safizadeh, Z. Liu, D. S. Forsyth, and A. Fahr, “Automatic classification and characterization of hidden corrosion using pulsed eddy current data,” in Proc. 16th World Conf. NDT, Montreal, QC, Canada, Aug. 2004.

Zheng Liu (M’03–SM’05) received the B.E. degree in mechanical and automation engineering from Beijing Institute of Chemical Fibre Technology, Beijing, China, in 1991, the M.E. degree in automatic instrumentation engineering from Beijing University of Chemical Technology, in 1996, and the Ph.D. degree in engineering from Kyoto University, Kyoto, Japan, in 2000. From 2000 to 2001, he was a Research Fellow with the Control and Instrumentation Division, School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore. He then joined the Institute for Aerospace Research (IAR), National Research Council Canada, Ottawa, ON, Canada, as a Government Laboratory Visiting Fellow selected by the Natural Sciences and Engineering Research Council. After being with IAR for five years, he joined the Institute for Research in Construction, National Research Council Canada, where he is currently a Research Officer. His research interests include image/data fusion, computer vision, pattern recognition, senor/sensor network, and nondestructive inspection.

David S. Forsyth (S’92–M’92) received the B.E. degree in engineering physics and the M.E. degree in electrical engineering from the University of Saskatchewan, Saskatoon, SK, Canada. He has worked on nondestructive testing (NDT) for more than 15 years, first with the National Research Council Canada, Ottawa, ON, Canada, and, since 2005, with the Nondestructive Testing Information Analysis Center, Texas Research Institute Austin, Austin, TX. His research interests are the use of advanced signal/image processing, including data fusion to improve the reliability of nondestructive testing, and the measurement of the reliability of NDT.

Mir-Saeed Safizadeh received the B.Sc. degree in mechanical engineering and the M.Sc. degree from Iran University of Science and Technology, Tehran, Iran, in 1988 and 1991, respectively, and the Ph.D. degree in machine diagnosis in time–frequency plane from the Ecole Polytechnique of Montreal, Montreal, QC, Canada, in 1999. From 1999 to 2005, he was with the Institute for Aerospace Research, National Research Council Canada, Ottawa, ON, Canada. Since 2005, he has been a Professor with Iran University of Science and Technology. He is currently actively involved in the field of nondestructive evaluation and diagnostic of rotating machinery.

Abbas Fahr received the B.Sc. degree in physics from Mashad University, Mashad, Iran, in 1971, the M.Eng. degree in nondestructive testing of materials from Brunel University, London, U.K., in 1975, and the Ph.D. degree in materials science from Imperial College of London University, London, U.K., in 1979. He is the Group Leader of nondestructive evaluation and a Senior Research Officer with the Institute for Aerospace Research, National Research Council Canada, Ottawa, ON, Canada.

Authorized licensed use limited to: IEEE Xplore. Downloaded on November 17, 2008 at 08:22 from IEEE Xplore. Restrictions apply.