paper on electronics

Measurement 46 (2013) 2065–2072

Contents lists available at SciVerse ScienceDirect

Measurement

journal homepage: www.elsevier .com/ locate/measurement

Surface roughness classification using image processing

0263-2241/$ - see front matter � 2013 Elsevier Ltd. All rights reserved.http://dx.doi.org/10.1016/j.measurement.2013.03.014

⇑ Corresponding author. Tel.: +91 44 27474262; fax: +91 44 27474208.E-mail address: [email protected] (T. Jeyapoovan).

T. Jeyapoovan a,⇑, M. Murugan b

a Department of Mechanical Engineering, Hindustan Institute of Technology and Science, Chennai, Indiab Department of Mechanical Engineering, B.S. Abdur Rahman University, Chennai, India

a r t i c l e i n f o

Article history:Received 3 July 2012Received in revised form 28 January 2013Accepted 18 March 2013Available online 3 April 2013

Keywords:Image processingSurface roughnessEuclidean distanceHamming distanceSurface characterization

a b s t r a c t

Surface roughness is an important factor in determining the satisfactory functioning of themachined components. Conventionally the surface roughness measurement is done with astylus instrument. Since this measurement process is intrusive and is of contact type, it isnot suitable for online measurements. There is a growing need for a reliable, online andnon-contact method for surface measurements. Over the last few years, advances in imageprocessing techniques have provided a basis for developing image-based surface roughnessmeasuring techniques. Based upon the vision system, novel methods used for human iden-tification in biometrics are used in the present work for characterization of machined sur-faces. The Euclidean and Hamming distances of the surface images are used for surfacerecognition. Using a CCD camera and polychromatic light source, low-incident-angleimages of machined surfaces with different surface roughness values were captured. A sig-nal vector was generated from image pixel intensity and was processed using MATLABsoftware. A database of reference images with known surface roughness values was estab-lished. The Euclidean and Hamming distances between any new test surface and the refer-ence images in the database were used to predict the surface roughness of the test surface.

� 2013 Elsevier Ltd. All rights reserved.

1. Introduction

Machining of metal surfaces through processes such asplanning, milling, EDM or grinding produces a specific laypattern. For example, a milled surface has a typical regularand periodic lay pattern. Surface topography has two fun-damental features: the amplitude of peak or valley on thesurface, and the wavelength between the peaks or valleys.Surface measurements are generally expressed in terms ofprofile of a surface y(x) in two dimensions and are assumedas equivalent to three-dimensional expressions. Theaverage surface roughness parameter (Ra) represents theaverage deviation of a surface profile about a mean line.Ra is generally used for surface roughness measurementand characterization [1].

For many years, the stylus instrument has been widelyused for measuring surface roughness parameters with

high reliability. The vertical movement of tip of the stylusis measured for a predetermined length horizontally. Thestylus tip, however, could not reach into all the valleys ofthe surface, and thus acts as a low-pass filter of surfacedata. Therefore, the stylus tip radius has a limitation inthe measurements of fine surfaces. Also, the high fre-quency components of surface roughness are filtered bythe stylus tip, as well as any non-linear deformationin the surface cannot be adequately measured. In addition,the stylus tip may cause damage to and/or may getdamaged on contact with the surface being measured.Requiring considerable time for setting up the stylusinstrument before surface measurement is anotherlimitation.

The need for a high-speed, non-contact and highly reli-able surface measurement system is on the rise. Althoughmany techniques are available for surface roughness mea-surements, including the optical techniques, no techniquehas been established reliable and robust enough for shopfloor applications. The methods adopted and used in this

2066 T. Jeyapoovan, M. Murugan / Measurement 46 (2013) 2065–2072

paper were followed successfully in biometric recognition.An attempt is made in the present work to use a similartechnique for surface characterization of machined parts.The biometric recognition techniques are proven to berobust and reliable and are also found good for surfacecharacterization. This is one among the non-contact meth-ods using the surface images to measure the values of theEuclidean distance and Hamming distance of the referenceimages and a test surface image for comparison. The sur-face roughness measurements of reference surfaces weredone using stylus and the corresponding images werestored in the database. The test surface image was charac-terized based on value of Euclidean and Hamming dis-tance. The smaller this value, the greater is the matchingof the reference surface image with test surface ima-ge.Thus, the average surface roughness, Ra, of that refer-ence image can be attributed to the test image.

Fig. 1. A typical laser speckle image.

2. Literature

Optical methods [2] have more potential and numerouspossibilities for surface characterization. Some of the com-monly used optical methods for surface measurements areoptical microscopy, light scattering technique and use ofvision systems. In optical methods, a coherent or incoher-ent light is used and the scattering of light from the surfacecan be used for surface measurement and characterization.

2.1. Light scattering

Light scattering has been used in many studies of sur-face characterization of machined surfaces. Tian et al. [3]used plane-polarized light and a scatter light detector forsurface characterization. The angular-resolved scatter(ARS) and total integrated scatter (TIS) are the two meth-ods based on light scattering to measure surface rough-ness. In ARS method, the theoretical expression of ARSagainst surface roughness involves the state of polarizationof the incident light. In TIS method, the light that scattersinto a hemisphere from the surface being investigated iscollected and measured using a scatter light detector forsurface roughness measurement.

2.2. Laser speckle image

When a coherent beam of light (laser) is projected ontoa rough surface, a speckle image is obtained from themutual interference of scattered light because of the spa-tial fluctuations of the rough surface. A typical laserspeckle image [2] is shown in Fig. 1.

Several laser speckle techniques have been developedfor surface roughness evaluation. Light scattering is causedby roughness of the surface and so the speckle imagesobtained can be used to measure surface roughness [2].

Persson [4] used a speckle contrast technique to charac-terize surface roughness. The speckle pattern is producedby illuminating the rough surface with a He–Ne laser.Based on the contrast parameters of the speckle pattern,surface roughness measurements and characterization

are evaluated. The contrast parameters are obtained fromthe variation in the intensities of speckle image.

Fujii and Asakura [5] investigated surface roughnessfrom the statistical properties and distribution of speckleimage intensity. In the study, the standard deviations ofthe intensity fluctuations in the speckle patterns wereobserved to be having a linear relationship with surfaceroughness values.

2.3. Machine vision system

Machine vision-based techniques are suitable for onlineinspection of surfaces of machined parts and are safe onsurfaces being measured and the measuring system.

Kumar et al. [6] used a vision system to obtain surfaceimages and quantified the surface roughness using aregression analysis. The average gray value (Ga) of the sur-face image was calculated and calibrated with the respec-tive average surface roughness (Ra) of the surfacemeasured by the stylus.

Gadelmawla [7] and Tian and Lu [8] used the Gray LevelCo-occurrence Matrix (GLCM) method. This statisticalmethod considers the spatial relationship of pixels on thesurface image. Surface roughness is extracted by exploringthe relationships of average surface roughness (Ra) withthe features of GLCM of the surface image.

2.4. Advances in image processing techniques

Several techniques for the recognition of fingerprintsand iris for human identification are commerciallyavailable.

Ma et al. [9] used the local sharp variation points, repre-senting the appearance or vanishing of an important imagestructure, to characterize the iris. In their study, the proce-dure for feature extraction of iris consisted of two steps. Aset of 1D intensity signals were generated for characteriz-ing the most important information of the original 2Dimage. Then a matching scheme based on exclusive ORoperation was used to compute similarity between the

T. Jeyapoovan, M. Murugan / Measurement 46 (2013) 2065–2072 2067

images. The inner and outer boundaries of an iris were ta-ken as circles. The iris region was localized and exactparameters between the two circles were obtained usingedge detection algorithm and Hough transform. The inten-sity values of the captured images were normalized toovercome the effects of variations in lighting and differ-ences in distance between human eye and camera. Thenthe annular iris was unwrapped to a rectangular textureblock with a fixed size. In iris image processing, the centralidea is to capture the local sharp variations along the xdirection. Keeping with this, the authors averaged theintensity values of, say, 5 rows and generated a 1D signalrepresenting all the 5 rows. Thus, the whole 2D normalizedrectangular image was converted into a set of 1D intensitysignals using the following equation [9]:

Si ¼1M

XM

j¼1

Iði�1Þ�Mþj i ¼ 1;2;3; . . . N ð1Þ

where

I ¼

I1

..

.

Ix

..

.

Ik

0BBBBBBB@

1CCCCCCCA¼ ðIT

1; � � � ITx ; � � � I

TkÞ

T ð2Þ

where I represents the normalized 2D image intensitymatrix; M is the total number of rows added to get eachrows in signal matrix Si; N is the total number of 1D signalsto be obtained from the image intensity matrix I; Ix repre-sents the gray values of xth row in image intensity matrix;T represents the transpose, each rows in the matrix I aretransposed and appended to get single column matrixand finally the transpose of it is obtained to get the linearrow matrix; and k represents the total number of rows inthe image intensity matrix I.

The 1D intensity signal is obtained from the imageintensity matrix I, and the expression (i–1)�M + j refersthe subscript in the image intensity matrix I, which areto be added and averaged to get the signal matrix S. Thesubscript i refers to the row numbers in the signal matrixS and j refers to the number of rows to be added to getthe signal vector. The number of columns in signal matrixS will be the same as the number of columns available inimage intensity matrix I. N is the total number of 1D sig-nals to be obtained from the image intensity matrix andthe number of columns in this is same as the number ofcolumns in the image intensity matrix I. The image signalmatrix S is a combination of successive horizontal scanlines that represent local variations in the surface alongthe horizontal direction. A set of such signals containsthe local sharp variations of the surface.

The recognition rate of the proposed method can beregulated by varying the parameter M. When the numberof rows merged, M, is smaller, more values in the signalvectors are produced for the same image intensity matrixI, and this will take more time for processing. For a largevalue for M, fewer values in the signal vectors are producedfor the same intensity matrix I, and it takes less processing

time but the recognition accuracy is low. In the presentwork, the optimum value of M = 3 and N = 1 was selectedby trial-and-error method.

Martin-Roche et al. [10] used the Euclidean distanceand the Hamming distance as feature metrics in iris recog-nition. The Euclidean distance is a distance between twopoints in a plane or space. For an n-dimensional space,the Euclidean distance (DE) is given by

DEðp;qÞ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiðp1�q1Þ

2þðp2�q2Þ2þ . . .þðpi�qiÞ

2þ . ..þðpN�qNÞ2

qð3Þ

DEðp; qÞ ¼

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXN

i¼1

ðpi � qiÞ2

vuut ð4Þ

where N is the dimension of the feature vector; pi is the ithcomponent of the feature vector and qi is the ith compo-nent of the template vector.

The Hamming distance represents the distance betweentwo items by the number of mismatches among their pairsof variables. It is commonly used for bitwise analysis ofimages in similarity measurements. It is also used to definea threshold. The pixel intensity values having absolute dif-ference below the threshold are taken as equal. The Ham-ming distance (DH) is given by

DHðp; qÞ ¼1N

XN

i¼1

pi – qi ð5Þ

The Hamming distance is simply the number of compo-nents that differ in value at corresponding areas of twoimages. Similar images of an object will have a less valueof Hamming distance.

3. Experimental setup and procedure

Six specimens were prepared with different surfaceroughness by milling. The feed was varied to achieve dif-ferent surface roughness values, but other cutting condi-tions, such as spindle speed, depth of cut and dry cutting,were kept constant. This is because although all the threecutting parameters, feed, spindle speed and depth of cut,can influence the surface roughness, the influence of feedis most significant. Specimens with different surfaceroughness were obtained using the feed rate of 900,1300, 1700, 2100, 2500, and 2900 mm/min at the spindlespeed of 1800 rpm and depth of cut of 0.2 mm. The averagesurface roughness (Ra) values of the specimens were mea-sured using a stylus instrument and recorded.

3.1. The Approach

A description of the approach of this study follows. First,a set of machined surfaces with known surface roughnessvalues were photographed and stored in database to actas reference images. The average surface roughness ofmachined surfaces was measured using a Taylor andHobson Talysurf surface roughness tester with stylus tipof radius 2 lm and cut-off value of 0.8 mm for a datalength of 5 mm. Care was taken to ensure that the surface

Fig. 2. Stages in the proposed method.

Fig. 3. Experimental setup.

2068 T. Jeyapoovan, M. Murugan / Measurement 46 (2013) 2065–2072

roughness values of specimens covered a range of 0.8–2.6 lm, which is the common range of surface roughnessvalues obtained in milling operations.

Subsequently, the surface images of specimens with un-known surface roughness were captured, and the similar-ity/difference of surface test images was quantified using

Fig. 4. Reference images

two different metrics against the reference images, andthe surface roughness of the surfaces of specimens wasthus estimated.

Therefore, the proposed method has four stages (Fig. 2):image acquisition, establishing image database, featurecomparison and surface characterization.

3.2. Image acquisition

For image acquisition, a Basler PiA2400-12gm CCD cam-era fitted with a Zoom 6000 lens system having opticalmagnification up to 45.0X and the lighting system withtwo halogen bulbs of 1000 W were used. An adjustable

of six specimens.

Fig. 5. Computation of Euclidean distance between the test signal and thereference signal.

T. Jeyapoovan, M. Murugan / Measurement 46 (2013) 2065–2072 2069

table was fabricated to hold the specimen and to have thecamera at particular angles. The CCD camera could be ad-justed to any convenient angle to the specimen surfacewith the help of a protractor located at its center.

For the present work, images were obtained with theposition of the camera at 45� to the surface of the specimenwith constant settings. All images were obtained with thesame settings and position of the camera. Uniform illumi-nation of the surface was ensured by using a diffuse lightsource. Surface images were captured for each specimenfor different positions of the specimen. The image resolu-tion of 2058 � 2456 pixels was obtained for each of thesurface images. The arrangement of specimen, CCD cameraand lighting is shown in Fig. 3.

3.3. Image database

Six specimens were used in the study, and so the data-base contained six reference images. For each specimen,numerous images were captured; out of these, one wasstored in the database as reference image and six wereused as test images. The reference images in the databaseare shown in Fig. 4.

Soon after capturing, the images were normalized todeal with lighting variation or any fluctuations in imageacquisition that may affect the image processing. Normal-ization is a process to transform the image matrix to haveuniformity in image pixel intensity. The normalization ofintensity range of the image matrix is obtained as follows[12]:

gðx; yÞ ¼ f ðx; yÞ �minðf Þmaxðf Þ

� �� 255 ð6Þ

where g(x, y) is the normalized image intensity matrix, f(x,y) is the image intensity matrix, min(f) is the minimum va-lue in the image intensity matrix, and max(f) is the maxi-mum value in the image intensity matrix. All values ofg(x, y) are multiplied with the maximum gray scale valueof an 8-bit image, i.e., 255, to make pixel intensity in grayscale of the range between 0 and 255. An image signal of 3-pixel row band was obtained by averaging the pixel inten-sity of the respective columns.

The images were processed in MATLAB 7.1. A pro-gramme module was written to generate the image signalfrom the reference images stored in the database. Anothermodule was written to read the test image. The test imagesignal was generated similar to the reference images andwas used for further processing.

3.4. Feature comparison

For surface characterization, feature extraction andcomparison was performed using the metrics Euclideandistance and Hamming distance. These two metrics havebeen successfully used in iris recognition in human identi-fication. The Euclidean distance between two vectors, p (p1,p2, p3, . . .pn) and q (q1, q2, q3, . . .qn), or precisely the spatialdistance between two vectors p and q, was computed usingthe Eq. (3). It is the spatial distance between two vectors pand q, as well as the measure of dissimilarity between two

vectors p and q. Higher the value of Euclidean distance, thehigher is the dissimilarity. The procedure to compute theEuclidean distance (DE) is given in flowchart (Fig. 5), whichis based on circular shift-based matching method [9]. Thecircular shift-based matching eliminates the possibility ofa simple shift in the image that could drastically influencethe Euclidean distance.

For comparison of images, the start point of the test andreference image signals may not begin from an identicallocation. In shift-based matching, the Euclidean distanceof test and reference image signals was calculated in aloop. The loop will retain the same start point for referenceimage (p) signal for i = 1 to N and the start point for test im-age (q) signal for j = 1 to N by default and the Euclidean dis-tance was found. Then the start point for test image signalwas shifted for j = 2 to N and ended at 1, and the Euclideandistance was found; again the start point was shifted forj = 3 to N and ended at 2, and the Euclidean distance wasfound. This shifting was continued for one complete circu-lar shift and the Euclidean distance was found. The lowestof the Euclidean distance of the reference image with the

Fig. 6. (A) Euclidean distances of test images and (B) Euclidean distances for six test images and six reference images.

2070 T. Jeyapoovan, M. Murugan / Measurement 46 (2013) 2065–2072

test image was assumed as the matching of the referenceimage with the test image and was considered for surfacecharacterization.

For the second metric (Hamming distance), assumingthat the feature components follow a Gaussian distribu-tion, the standard deviation of the test image signal wasobtained. In feature comparison, the number of compo-nents of the test image feature vector falling outside thearea defined by the reference image feature vector wascounted [10] using the condition ðjðyi � piÞjÞ > pv

i , whereyi is the ith component of the reference image feature vec-tor; pi is the ith component of the test image feature vec-tor; pv

i is the standard deviation of the test image featurevector. By using the count of the number of mismatches,the Hamming distance of reference image signals and testimage signal was obtained. The lowest of the Hammingdistance was the matching of the reference image withthe test image and was used to characterize the testsurface.

3.5. Surface characterization

The MATLAB Image Processing Toolbox was used forfeature comparison and surface characterization. A pro-gramme was written with four modules: (a) to create adatabase of reference images of specimens; (b) to generatereference image signals; (c) to generate test image signaland (d) to evaluate the Euclidean distance and Hammingdistance.

Euclidean distances for the reference image signals withthe test image signal were obtained and tabulated. A low

value of Euclidean distance indicates the matching of thattest image signal with the reference image signal.

Hamming distances were calculated by comparing thereference image signals and test image signal as explainedpreviously. Hamming distances for all the reference imageswith the test image were obtained and tabulated. The leastvalue of Hamming distance shows a perfect matching ofthe reference image with the test image.

4. Results and discussions

Test image signals were compared with reference imagesignals and the Euclidean distances between the test imagesignals and reference image signals were computed andtabulated for six specimens. Fig. 6A shows the Euclideandistances of six test images with respect to referenceimages.

From the figure, it is seen that the Euclidean distancebetween test image T1 and reference images 2–6 is sub-stantially higher than the Euclidean distance between testimage 1 and reference image 1. The same trend is observedfor other test images with reference images of differentsurface roughness values.

Fig. 6B shows the Euclidean distances of six test surfaceimages with the six reference images. A low value ofEuclidean distance means a test image matches closelywith a reference image. The reference image having theleast Euclidean distance with the test image was assumedto be matching and was used for surface characterization.Thus, the average surface roughness, Ra, of that reference

Fig. 7. (A) Hamming distances of test images and (B) Hamming distances for six test images and six reference images.

T. Jeyapoovan, M. Murugan / Measurement 46 (2013) 2065–2072 2071

image can be attributed to the test image. The accuracy ofthis method can be significantly improved when a largedatabase of reference images covering the entire range ofsurface roughness values is established.

The Hamming distances for the test images with thereference images were obtained. Fig. 7A shows the Ham-ming distances of six test images with the referenceimages. The least value of Hamming distance means a testimage matches closely with a reference image. It is ob-served from the figure that the value of Hamming distancebetween test image T1 and database images 2–6 is muchhigher than the Hamming distance between test image 1and reference image 1. Hamming distance of the test imageand reference image was found to be very low for thematching reference images (below 0.30 in all cases). In trailtesting, when the test image was one of the referenceimages, the Hamming distance was zero, and was alsofound to be close to zero in few test images for perfectmatching with reference images.

Fig. 7B shows the Hamming distances for six test sur-face images with six reference images. The test image hav-ing the least Hamming distance with the reference imagewas used for surface characterization. Thus, the averagesurface roughness, Ra, of that reference image could beattributed to the test image. A large database of referenceimages covering the entire range of surface roughness val-ues are needed to improve the accuracy of this method.

Daugman [11] conducted tests on few millions of irisimages and found that when testing the iris images ofthe same person, the Hamming distances were less than0.32 for human identification. For the value of Hamming

distance less than or equal to 0.32, it was reported thatthe possibility of false acceptance was 1 in 151,000 andof false rejection was 1 in 128,000 in the human identifica-tion systems.

In the present work, the values of Hamming distances inall identification were less than 0.30 for correct matching,and the Hamming distances of most of the other imageswere greater than 0.30. The same trend was observed forall other test images with the reference images of differentsurface roughness values.

Considering the consistency in the recognition of testimages, it is concluded that Euclidean distance and Ham-ming distance are probable metrics for in-process surfacecharacterization. The value of Hamming distance wasnearer to zero for perfect matching when compared tothe matching using the Euclidean distances. Thus, surfacecharacterization using Hamming distance has larger scopethan using the Euclidean distance.

5. Conclusion

This work attempts to evolve a vision system-based im-age processing technique for non-contact evaluation ofsurface roughness. A set of two-dimensional images ofmilled surfaces was obtained, and from this a set of refer-ence images was converted into one-dimensional refer-ence signals. For any image with unknown surfaceroughness (test image), the test signal was arrived at byadopting a similar procedure. The Hamming distance be-tween the reference signals and test signal was used toestimate the surface roughness of test image.

2072 T. Jeyapoovan, M. Murugan / Measurement 46 (2013) 2065–2072

Using Euclidean distance and Hamming distance, thematching of test images with reference images providedexcellent results. It was observed that the Euclidean andHamming distances were very low for surfaces with simi-lar surface roughness values. Therefore, this technique isideal for online surface characterization of machined sur-faces. The use of larger database of reference images andexploration of this technique for other machining pro-cesses, such as planning, EDM and ground surfaces, arepotential areas for future research.

References

[1] D.J. Whitehouse, Surface metrology, Measurement Science andTechnology 8 (1997) 955–972.

[2] T.V. Vorburger, E.C. Teague, Optical techniques for on-linemeasurement of surface topography, Precision Engineering 3(1981) 61–83.

[3] G.Y. Tian, R.S. Lu, D. Gledhill, Surface measurement using activevision and light scattering, Optics and Lasers in Engineering 45(2007) 131–139.

[4] U. Persson, Real time measurement of surface roughness on groundsurfaces using speckle-contrast technique, Optics and Lasers inEngineering 17 (1992) 61–67.

[5] H. Fujii, T. Asakura, Effect of surface roughness on the statisticaldistribution of image speckle intensity, Optics Communications 11(1974) 35–38.

[6] R. Kumar, P. Kulashekar, B. Dhanasekar, B. Ramamoorthy,Application of digital image magnification for surface roughnessevaluation usingmachine vision, International Journal of MachineTools & Manufacture 45 (2005) 228–234.

[7] E.S. Gadelmawla, A vision system for surface roughnesscharacterization using the gray level co-occurrence matrix, NDT &E International 37 (2004) 577–588.

[8] G.Y. Tian, R.S. Lu, Hybrid vision system for online measurement ofsurface roughness, Journal of the Optical Society of America 23(2006) 3072–3079.

[9] L. Ma, T. Tan, D. Zhang, Efficient iris recognition by characterizing keylocal variations, IEEE Transactions on Image Processing 13 (2004)739–750.

[10] D. Martin-Roche, C. Sanchez-Avila, R. Sanchez-Reilo, Iris recognitionfor biometric identification using dyadic wavelet transformzero-crossing, in: Proceedings of the 35th IEEE InternationalCarnahan Conference on Security Technology. London, UK, 2001,pp. 272–77.

[11] J.G. Daugman, High confidence visual recognition of personsby a test of statistical independence, IEEE Transactions onPattern Analysis and Machine Intelligence 15 (11) (1993) 1148–1161.

[12] R.C. Gonzalez, R.E. Woods, Digital Image Processing, Prentice Hall,2007.