Objective Surface Evaluation of Fiber ReinforcedPolymer Composites

Stuart Palmer & Wayne Hall

Received: 18 June 2012 /Accepted: 27 August 2012 /Published online: 15 September 2012# Springer Science+Business Media B.V. 2012

Abstract The mechanical properties of advanced composites are essential for their struc-tural performance, but the surface finish on exterior composite panels is of critical impor-tance for customer satisfaction. This paper describes the application of wavelet textureanalysis (WTA) to the task of automatically classifying the surface finish properties of twofiber reinforced polymer (FRP) composite construction types (clear resin and gel-coat) intothree quality grades. Samples were imaged and wavelet multi-scale decomposition was usedto create a visual texture representation of the sample, capturing image features at differentscales and orientations. Principal components analysis was used to reduce the dimensionalityof the texture feature vector, permitting successful classification of the samples using onlythe first principal component. This work extends and further validates the feasibility of thisapproach as the basis for automated non-contact classification of composite surface finishusing image analysis.

Keywords Carbon fiber . Surface analysis . Image analysis . Wavelet texture analysis

1 Introduction

The mechanical properties of advanced composites are essential for their structural perfor-mance, but the surface finish on exterior composite panels is of critical importance forcustomer satisfaction [1]. Customers demand a flawless (Class A) surface finish, but this canbe difficult to achieve on composite surfaces. Dry spots can occur in wet lay-up processes,and the strong reinforcement fibers can ‘read through’ to the exterior surface, spoiling thecosmetic appearance [2], and surface imperfections can lower the gloss of the surface finish[3]. It is essential that composite manufacturers have reliable and repeatable methods for

Appl Compos Mater (2013) 20:627–637DOI 10.1007/s10443-012-9291-6

S. Palmer (*)Faculty of Science and Technology, Deakin University, Gheringhap Street, 3217 Geelong, Australiae-mail: spalm@deakin.edu.au

W. HallGriffith School of Engineering, Griffith University, Parklands Drive, 4222 Southport, Queensland,Australia

evaluating surface texture. To date, assessment of surface finish quality has tended to bebased simply on human visual observation. While this method has been found to deliverresults that are acceptable to customers, it is generally performed using several observers inorder to produce statistically meaningful results [4], and is therefore time-consuming and notdirectly adaptable to the automated manufacture of composite products [5]. Systems for theobjective assessment of surface quality do exist, but currently struggle to replicate the humanvisual assessment of surface finish [5], and commercially available systems are typicallyvery expensive.

It has been observed that many types of engineering surfaces contain textural features atmultiple scales [6], and may be fractal (self-similar at different scales) in nature [7, 8]. Whilethere exist a number of numerical methods for characterizing engineering surfaces, manyrequire that the distribution of surface features is stationary (i.e., the frequency content doesnot vary with location), an assumption that is often not valid [6]. It has been shown that thewavelet transform has the ability to effectively characterize surface profile data that containmulti-scale features and are non-stationary [6], and are fractal in nature [7]. For thecomprehensive characterization of surface features and texture, these inherent abilities ofthe wavelet transform place “it way ahead of other traditional methods” [9], and are why it is“generally considered to be state of the art in texture analysis” [10]. Wavelet analysis hasbeen applied to the characterization of material surface parameters. Data from 2D waveletanalysis were used as the basis for a successful empirical parametric mapping betweenmaterial surface images obtained via computer vision acquisition and standard surfaceroughness parameters obtained using conventional stylus measurement [11]. Wavelet anal-ysis has also been applied to the analysis of surface characteristics of reinforced polymercomposites [12–16].

It has also been observed that for resin transfer molded composite plates with surfacesthat have approximately similar quality, human visual observation generally outperformsobjective (mathematical) methods in the differentiation of sample surface quality – possiblybecause the standard surface roughness parameters commonly used may not provide anunambiguous indicator of surface quality that agrees with human visual assessment [5].Objective techniques for the characterization of surface quality of reinforced polymercomposites that can provide the same results as a human subjective evaluation are thereforehighly desirable. Physiological experiments have shown that the visual cortex appears toperform a 2D multi-scale decomposition of the visual field into a range of frequency bands/channels [17]. There is considerable similarity between the wavelet transform and biologicalvisual systems. This similarity has resulted in its use in biologically inspired computer visionsystems [18]. The 2D wavelet transform is a mathematically robust analysis tool for thecharacterization of material surface finish data in ways analogous to human visual processes,and offers practical and rigorous methods for the objective classification of surface quality.

Previous work [19] established the feasibility of wavelet texture analysis (WTA) for thetask of automatically classifying the surface finish of carbon fiber reinforced polymer(CFRP) samples into two quality grades (‘good’ and ‘bad’) based on the detection of thepresence of surface dry patches. In response to suggestions from reviewers of the previouswork, the work documented here extends the previous work in a number of ways. Firstly,while there is a growing market for uncoated, visible composite components [20], themajority of composite applications incorporate a surface coating, so we have extended theinvestigation here to include both bare resin and gel-coat composite surface finish types.Secondly, while resin dry patches were explored previously, other types of surface defectscan lead to poor assessment of surface finish quality, so in this work we investigate theimpact of point surface defects – the ability to detect point defects suggests that line defects

628 Appl Compos Mater (2013) 20:627–637

can also be detected. Finally, we have extended the investigation to the classificationof three grades of surface finish quality. The work documented here extends theapplication of the WTA classification method and further validates the feasibility ofthis approach as the basis for automated non-contact classification of compositesurface finish using image analysis methods analogous to the functioning of the humanvision system.

2 Material and Methods

The CFRP panel comprised three layers of 200 g/m2 carbon fiber plain weave cloth (suppliedbyATLComposites – code ZP200) impregnated with epoxy resin (R180 epoxy resin and epoxyhardener H180 slow - supplied by Fiber Glass International, FGI). This visible plain weaveprovided an additional complication to the surface assessment process, but it is recognized thatmost external fiber reinforced polymer (FRP) composite surfaces are gel-coated or painted.Thus, representative gel-coated samples were also considered. The surface of the gel-coatedpanel was coated with three layers of Aquaguard Performance Gelcoat – supplied by FiberGlass International, FGI, code FM032502. Both the CFRP and gel-coat panels were manufac-tured on a pre-released flat glass mold surface. The gel-coat layers were generously applied toensure no fiber ‘read through’ and therefore the underlining reinforcement fibers was lessimportant for the gel-coat panel. For this reason, four layers of Chop Strand Mat (CSM) wereused as the reinforcement (450 g/m2 emulsion, FGI code F02254) in preference to carbon. Thematrix was Unsaturated Polyester (UPE) (Escon – FGI code F62300). Resin was introduced byhand using brushes and the panels were backed with a plywood base for flexural stiffness.Curing occurred under atmospheric conditions. Following curing, three sections (one for eachgrade) were cut from each of the panels.

To create three different surface finishes (grades 1–3), coarse sandpaper was usedto introduce a ‘random’ distribution of point defects. Each sample panel was posi-tioned surface down on the sandpaper and the plywood backing was tapped aspecified number of times with a mallet. The taps were of even force/pressure andwere distributed evenly over the back surface. Grade 1, 2 and 3 received 0, 10 and 30taps respectively. The sample panels were scanned at 600 pixels per inch (approxi-mately 236 pixels per cm) using a Hewlett-Packard HP3200C flatbed scanner to yieldhigh resolution 8 bit (256 grey scale) images. In addition, contrast enhancement wasapplied to the gel-coat sample images to highlight the surface defects which werealmost invisible to the naked eye. These high-resolution scans were then separatedinto 4 sections each, yielding 12 sample images for both clear resin and gel-coatconstruction, with 4 examples of each grade of surface finish. All numerical analysesdescribed hereafter were performed using the Matlab computing environment [21, 22].The wavelet analysis method is expedited by images that have linear dimensions of aninteger power of two. To this end, all 24 sample images were sized to be 1024 by1024 pixels for testing. Figure 1 shows typical clear resin sample test images for eachsurface finish grade produced in this manner. Figure 2 shows typical gel-coat sampletest images for each surface finish grade produced in this manner.

Detailed mathematical treatments of the wavelet transform are available elsewhere [23],and we provide additional background in our previous work [19]. Essentially, the twodimensional discrete wavelet transform (2DDWT) produces a nearly orthogonal decompo-sition of an image into sets of coefficients that separately represent the information in theoriginal image in three orientations (horizontal, vertical and diagonal) and different scales.

Appl Compos Mater (2013) 20:627–637 629

The scale is a characteristic dimension related to the wavelet basis selected for the decom-position. The DWT is an iterative decomposition, in which the scale doubles at each step,placing a limit on the number of levels of decomposition related to the wavelet basis and thesize of the images. Wavelet analysis requires the selection of a wavelet basis for thedecomposition. There are no definitive rules for selecting the ‘best’ wavelet for a particularanalysis application [24, 25]. A heuristic technique of analyzing sample data with a range ofcandidate wavelets and applying selection criteria to identify the optimal analysis wavelet isalso described [25, 26].

Wavelet texture analysis (WTA) creates a texture feature vector based on the waveletdetail coefficients (cD) from all decomposition levels and horizontal, vertical and diagonalorientations. This permits a rich representation of the texture in the image to be used as abasis for classification that includes features related to both scale and orientation. In thiscase, the elements of the texture feature vector are constructed from an energy measure ofeach set of wavelet detail coefficients. A range of energy measures are possible [10]; here weuse the square of the Frobenius norm of the wavelet detail coefficients, normalized by the

Fig. 1 Typical clear resin samples – (a) grade 1, (b) grade 2 and (c) grade 3

630 Appl Compos Mater (2013) 20:627–637

size of the coefficient set, as the energy measure. The construction of the texture featurevector for each sample image is given by Eq. (1):

Ejk¼ 1M�N cDk

jk k2

F�J�j�1;k¼h;v;dð Þ ð1Þ

where j is the wavelet analysis scale/level, J is the maximum analysis scale, k is the waveletdetail coefficient set orientation (horizontal, vertical or diagonal), andM×N is the size of thecoefficient set. Hence, the texture feature vector for each sample contains 3 J elements. Thesquare of the Frobenius norm of matrix A is defined as:

Ak k2F¼P

i;j

aijj j2 ð2Þ

The texture feature vectors for all of the test samples can be used in multivariate analysisto classify the samples. Principal components analysis (PCA) transforms a set of correlatedvariables into a smaller set of uncorrelated variables called ‘principal components’. PCA is

Fig. 2 Typical gel-coat samples – (a) grade 1, (b) grade 2 and (c) grade 3

Appl Compos Mater (2013) 20:627–637 631

often used as an initial step in multivariate analysis, and can help to assess the actualdimensionality of the data [27]. PCA uses linear matrix algebra to generate the principalcomponents from the original variables such that each principal component is a linearcombination of the original variables, and, taken together, all of the principal componentsform an orthogonal basis for the space of the original data, containing no redundantinformation [28]. The PCA transformation results in the first principal component havingthe highest variance possible (i.e., accounting for as much of the variability in the originaldata as possible). Each succeeding principal component subsumes as much of remainingvariance in the data as possible, under the condition that it is orthogonal (not correlated) withall the preceding principal components. This process generally results in a small number ofprincipal components embodying most of the information in the original variables, with arapid fall-off in importance beyond the first few principal components. In the applicationdescribed here, for both the clear resin and gel-coat samples sets, a number of candidatewavelet bases and wavelet decomposition levels were trialed, and the PCA results werevisualized by plotting the location of all samples on axes of the first two principal compo-nents normalized against their mean, as shown in Figs. 3 and 4. A combination of waveletbasis and decomposition level that yielded the good visual separation of the surface finishgrades was selected.

The principal components thus developed can be used as a new set of observationvariables for each sample image, but with reduced dimensionality, and hence, simplifiedsubsequent processing. The principal component scores were used as observation vectors indiscriminant analysis (DA) to classify the sample images. DA is a conventional probabilisticclassifier that allocates each observation to the class with which it has the highest posteriorprobability of membership. In the examples presented here, a simple linear classifier wasused. It presumes that the each element of observation vector has a normal/Gaussiandistribution. Based on ‘training’ with samples of each class of input type, linear partitionsare developed in the multi-dimensional space of the observation vector elements that canthen be used to automatically classify samples into input class types. In this work we areseeking to confirm that the samples can be viably separated by DA into discreet groupsbased on grade of surface finish. For each type of construction (clear resin and gel-coat), all12 samples were used in training, and the classification partitions generated by DA werevisualized to confirm the performance of the classification.

a bFig. 3 Plot of first two principalcomponents showing the locationof all clear resin samples

632 Appl Compos Mater (2013) 20:627–637

3 Results and Discussion

3.1 Clear Resin Samples

For the clear resin samples, good results were obtained using the Daubechies waveletwith seven vanishing moments (db7) and three levels of wavelet decomposition.Based on three levels of decomposition and three orientations, the wavelet transformproduced a texture feature vector for each sample containing nine elements. Thetexture feature vectors for all 12 clear resin sample images were computed andcombined as the input data for PCA. Following PCA transformation, it was foundthat the first principal component explained more than 96 % of the variance in theoriginal texture feature vector set, indicating that the PCA transformation had reducedthe dimensionality of the texture feature vectors from nine to effectively one, and thatthe first principal component alone would be sufficient for classification. Figure 3shows the location of all 12 clear resin samples, following PCA transformation,plotted on axes of the first two principal components. It is clear that the samplesseparate out into three distinct groupings associated with surface finish grade.

Given that virtually all of the variance in the original texture feature vector set isexplained by the first principal component (PC1), it was decided to use PC1 solely as thebasis for automatic classification by DA. Using the basic linear classification function inMatlab, the PC1 values for all 12 clear resin samples were used as training samples to deriveclassification points based on PC1. The DA classification points obtained are plotted as lines‘a’ (the boundary between grade 1 and grade 2) and ‘b’ (the boundary between grade 2 andgrade 3) on Fig. 3. It can be seen that the classification points successfully group all of the 12clear resin samples into their respective surface finish grades.

3.2 Gel-Coat Samples

For the gel-coat samples, good results were obtained using the Daubechies wavelet with fourvanishing moments (db4) and three levels of wavelet decomposition. The same analysisprocess that was applied to the clear resin samples was also used for the gel-coat samples. Itwas found that the first principal component explained more than 97 % of the variance in theoriginal texture feature vector set. As for the clear resin samples, the first principal

a bFig. 4 Plot of first two principalcomponents showing the locationof all gel-coat samples

Appl Compos Mater (2013) 20:627–637 633

component alone would be sufficient for classification. Figure 4 shows the location of all 12gel-coat samples, following PCA transformation, plotted on axes of the first two principalcomponents. Again, the samples separate out into three distinct groupings associated withsurface finish grade. As for the clear resin samples, the PC1 values were used for DA training,and the resultant classification points based on PC1 are plotted as lines ‘a’ (grade 1/2) and ‘b’(grade 2/3) on Fig. 4. Again, all 12 gel-coat samples were successfully classified into theirrespective surface finish grades.

3.3 General Discussion

The classification points ‘a’ and ‘b’ shown in Figs. 3 and 4 were derived by DA training, andcould be used as the basis the automatic grading classification of subsequent new imagedsamples from a similar population. The classification points could also be manually adjustedand/or further refined through the incorporation of additional grade examples in training. Asnoted above, the PCA process transforms a set of variables into a reduced number of‘principal components’, which are linear combinations of the original variables. Based onthe clear resin analysis, Table 1 shows the component loadings for each of the three waveletanalysis levels and each of the three analysis orientations within those levels, for thefirst principal component (PC1). PC1 is of interest because it was the basis for DAclassification.

It can be seen that the main/largest loading, and therefore main discriminating ability,occurs at the third analysis scale. A physical interpretation of this result can be observed byselectively reconstructing a sample image based on partial sets of wavelet decompositioncoefficients. Figure 5(a) shows the original grade 3 sample from Fig. 1. Figure 5(b) showsthe image reconstruction obtained from the inverse 2DDWT using all decompositioncoefficients except those from level 3. Figure 5(c) shows the image reconstruction obtainedusing only the coefficients from level 3. Figure 5(d) is the inverse/negative of Fig. 5(c) tohighlight the information in the level 3 reconstructed image. While the PC1 loadings givenin Table 1 show that the level 3 wavelet coefficients do not uniquely define the informationin the original image related to the surface finish grade, Fig. 5(c/d) does capture the pointdefects applied to the test sample, and the visual appearance of the defects in Fig. 5(b) ismuch reduced.

In the case of visual assessment of the quality of composite surface finish, the ability ofWTA to characterize surface texture in a way analogous to the human visual system holdsgreat promise. The essential method of orthogonal decomposition of an image into separate

Table 1 Wavelet analysis com-ponent loadings for first principalcomponent for the clear resinsamples

Level Orientation PC1 Loading

1 Horizontal 0.0133

1 Vertical 0.0010

1 Diagonal 0.0027

2 Horizontal 0.1279

2 Vertical 0.0738

2 Diagonal 0.0367

3 Horizontal 0.7301

3 Vertical 0.5743

3 Diagonal 0.3374

634 Appl Compos Mater (2013) 20:627–637

planes containing the information relating to different scales and orientations present in theoriginal image has the ability to isolate different types of visible surface flaws. This means, asdemonstrated here, that it is largely agnostic to the underlying composite construction (i.e.,woven or non-woven), and to whether the underlying reinforcement matrix is visible or not.

4 Conclusions

This paper describes the application of wavelet texture analysis (WTA) to the task ofautomatically classifying the surface finish properties of two FRP composite constructiontypes (clear resin and gel-coat) into three quality grades. Samples of both types of FRPconstruction were created and three grades of surface finish were applied. Followingscanning of the samples, wavelet multi-scale decomposition was used to create a visualtexture representation of the sample, capturing image features at different scales andorientations. Principal components analysis was used to reduce the dimensionality of the

Fig. 5 Selective reconstruction of clear resin grade 3 sample – (a) original sample, (b) all wavelet coefficientsexcept level 3, (c) level 3 coefficients only and (d) inverse of (c)

Appl Compos Mater (2013) 20:627–637 635

texture feature vector from nine elements, permitting successful classification of the samplesusing only the first principal component. This work extends and further validates thefeasibility of this approach as the basis for automated non-contact classification of compositesurface finish using image analysis methods analogous to the functioning of the humanvision system.

References

1. Palardy, G., Hubert, P., Haider, M., Lessard, L.: Optimization of RTM processing parameters for Class Asurface finish. Compos. B Eng. 39(7–8), 1280–1286 (2008). doi:10.1016/j.compositesb.2007.12.003

2. Wood, K.: Taking subjectivity out of class A surface evaluation. Compos. Tech. 14(4), 67–70 (2008)3. Das, P.D., Ravine, J., Pung, C.P., Vaidya, U.: Enhancing the Surface Finish of LFT Compression Molded

Parts via Application of Surface Film. Paper presented at the Composites & Polycon 2007, Tampa, FL,17–19 October

4. Schubel, P.J., Warrior, N.A., Kendall, K.N., Rudd, C.D.: Characterisation of thermoset laminates forcosmetic automotive applications: Part I - Surface characterisation. Compos. Appl. Sci. Manuf. 37(10),1734–1746 (2006). doi:10.1016/j.compositesa.2005.09.013

5. Hu, S., Hoa, S.V., Ganesan, R.: A new evaluation method for surface finish of compositeautomotive panels using waveform analysis. J. Compos. Mater. 43(8), 843–875 (2009). doi:10.1177/0021998309102038

6. Podsiadlo, P., Stachowiak, G.: Multi-scale representation of tribological surfaces. Proc IME J J Eng Tribol216(6), 463–479 (2002)

7. Goa, Y., Lu, S., Tse, S., O, K.: A wavelet–fractal-based approach for composite characterisation ofengineering surfaces. Int. J. Adv. Manuf. Tech. 20(12), 925–930 (2002)

8. Landini, G.: Fractals in microscopy. J. Microsc. 241(1), 1–8 (2011). doi:10.1111/j.1365-2818.2010.03454.x9. Pal, S., Ghatak, S.K., De, S., DasGupta, S.: Evaluation of surface roughness of a plasma treated polymeric

membrane by wavelet analysis and quantification of its enhanced performance. Appl. Surf. Sci. 255(5,Part 1), 2504–2511 (2008)

10. Bharati, M.H., MacGregor, J.F.: Image texture analysis: Methods and comparisons. Chemometr. Intell.Lab. Syst. 72(1), 57–71 (2004)

11. Chang, S.I., Ravathur, J.S.: Computer vision based non-contact surface roughness assessment usingwavelet transform and response surface methodology. Qual. Eng. 17(3), 435–451 (2005)

12. Gerasimov, V.V., Khandetsky, V.S., Gnoevoy, S.N.: Research of probability characteristics in defectdetection of composite materials using wavelet transform. Paper presented at the 8th InternationalConference of the Slovenian Society for Non-Destructive Testing, Portorož, Slovenia

13. Sarathi, R., Sahu, R.K., Tanaka, T.: Understanding the hydrophobic characteristics of epoxy nanocompo-sites using wavelets and fractal technique. IEEE Trans. Dielectr. Electr. Insul. 15(1), 178–186 (2008)

14. Loutas, T.H., Kostopoulos, V.: Health monitoring of carbon/carbon, woven reinforced composites:Damage assessment by using advanced signal processing techniques. Part II: Acousto–ultrasonicsmonitoring of damage development. Compos. Sci. Technol. 69(2), 273–283 (2009). doi:10.1016/j.compscitech.2008.09.042

15. Bussiba, A., Kupiec, M., Ifergane, S., Piat, R., Böhlke, T.: Damage evolution and fracture events sequencein various composites by acoustic emission technique. Compos. Sci. Technol. 68(5), 1144–1155 (2008).doi:10.1016/j.compscitech.2007.08.032

16. Hu, N., Shimomukai, T., Yan, C., Fukunaga, H.: Identification of delamination position in cross-plylaminated composite beams using S0 Lamb mode. Compos. Sci. Technol. 68(6), 1548–1554 (2008).doi:10.1016/j.compscitech.2007.10.015

17. Daugman, J.G.: Two-dimensional spectral analysis of cortical receptive field profiles. Vision Res. 20(10),847–856 (1980)

18. Lindblad, T., Kinser, J.M.: Inherent features of wavelets and pulse coupled networks. IEEE Trans. NeuralNetw. 10(3), 607–614 (1999)

19. Palmer, S., Hall, W.: Surface evaluation of carbon fibre composites using wavelet texture analysis.Compos. B Eng. 43(2), 621–626 (2012). doi:10.1016/j.compositesb.2011.08.024

20. Roberts, M.: Carbon composite process provides superb surface finish. Reinforc. Plast. 54(2), 29–32(2010). doi:10.1016/s0034-3617(10)70062-x

21. The MathWorks Inc.: Matlab Version 7.11.0.584 (R2010b). In. The MathWorks Inc., (2010)22. The MathWorks Inc.: Matlab Wavelet Toolbox Version 4.6. In. The MathWorks Inc., (2010)

636 Appl Compos Mater (2013) 20:627–637

23. Mallat, S.: A Wavelet Tour of Signal Processing. Academic, San Diego (1998)24. Hubbard, B.B.: The World According to Wavelets. A K Peters, Ltd., Wellesley, MA (1996)25. Percival, D.B., Walden, A.T.: Wavelet Methods for Time Series Analysis. Cambridge University Press,

Cambridge, UK (2000)26. Czarkowski, D., Domijan, A.: Wavelet analysis of electric adjustable speed drive waveforms. American

Society of Heating, Refrigeration and Air Conditioning Engineers Transactions 104(1A), 559–569 (1998)27. Johnson, D.E.: Applied multivariate methods for data analysts. Duxbury Press, Pacific Grove, California

(1998)28. Krzanowski, W.J.: Principles of Multivariate Analysis: A User's Perspective. Clarendon, Oxford (1990)

Appl Compos Mater (2013) 20:627–637 637