m comimpon andture aing cletho
The Euclidean distance is also used to represent the object similarity. The similarity canreach 87.5%, 62.5%, 75% and 87.5% respectively.
increasing. An object identication system  can also
are located in the corner place, and its security let peopleworried. If the monitored system can be stalled in toiletor outside locker room, when some specic strangers
base set  as an experimental basis, and thener as anproposed m
can reduce the waiting time of mobile devices whequery because of the less feature vector data and thple vector distance operation. They will help for theidentication according to these characteristics.
In the past, a lots of researchers majored on how tocombine the image features with the histogram to identifyimage features . However, different users have
http://dx.doi.org/10.1016/j.measurement.2014.01.0290263-2241/ 2014 Elsevier Ltd. All rights reserved.
Corresponding author at: Department of Civil Engineering, ShantouUniversity, Guangdong, PR China. Tel./fax: +86 75482904748.
E-mail address: email@example.com (L.-H. Juang).
Measurement 51 (2014) 100111
Contents lists available at ScienceDirect
journal homepage: www.elsevibe developed into the security monitor system due to to-day social security letting people panic, more public toilets
simple vector distance matching classiidentication source. We found that theuses aobjectethod
n userse sim-objectwith more convenient and friendly service. Think about it,if the service system can identify the object and its charac-teristics, then provides personal service for the differentobject, which can reduce the time when customer searchesfor products. In the other hand, it will also make the cus-tomer feel very intimate and accelerate the service for theirguests, nally it will make the companys performance
, but in the monitor system the face image has alow resolution or other factors lead to a lower recognitionrate. In this research, we will propose a preliminary designand experimental results of object recognition frommobiledevice that utilizes the texture and the color features witha simple vector distance matching classier to training andextract the characteristics. This paper uses the object data-Color featuresVector distanceMobile phone
Digital product development hason human life and convenience in inIf the identication system in a depstore entrance can accurately distingferent type products, we will be able 2014 Elsevier Ltd. All rights reserved.
many of benetsation technology.ent store or easybetween the dif-rovide customers
wander out, the identication system  will can earlynotify guards or the related people to handle it, so that canreduce to guard 24 h patrol and also has a protection onthe social security aspects.
Object identication is an important research at pres-ent. In the early, object identication studies were mostlyfocused on face or the prole feature such as genderKeywords:Object recognition
reach up to 100% of object identication rate when making a querying in a mobile phone.Object identication using mobile d
Li-Hong Juang a,b,, Ming-Ni Wu c, Zhi-Zhong WaDepartment of Civil Engineering, Shantou University, Guangdong, PR Chinab The Key Lab of Digital Signal and Image Processing of Guangdong Province,cDepartment of Information Management, National Taichung University of Te
a r t i c l e i n f o
Article history:Received 20 September 2013Received in revised form 17 December 2013Accepted 20 January 2014Available online 28 January 2014
a b s t r a c t
To detect object frois very difcult buta preliminary desigthat utilizes the textor distance matchthat the proposed mces
u University, Guangdong, PR Chinaogy, Taichung, Taiwan, ROC
plex background, illumination variations and texture by machinertant for adaptive information service. In this research, we presentexperimental results of object recognition from a mobile devicend the color features by image pre-processing with a simple vec-assier to train and extract the characteristics. The result showsd can adopt the few characteristic values and the accuracy can
er .com/ locate /measurement
different perspectives. In this research, our topic is to solvehow to nd out a specic object rapidly. In this research,we used two characteristics of feature extract and featurecomparison for image retrieval. In the feature extract part,we add the color feature information. Because there waslots of researchers in the past, they only adopted gray im-age or binary image, which cannot represent more detailinformation for image. Therefore, we rst transformedRGB color space to HSV color space, after this processing,we quantized it to 72 color numeric which can be used
right), shrink and rotation ones. Finally, we used Euclidean
tract its feature image. The third step bases on the previous
Fig. 1. HSV color attribute.
Fig. 2. The micro structure detection processing.
Fig. 3. The edge pixel similarity judgment.
Fig. 4. The four kinds of judgme
L.-H. Juang et al. /Measurement 51 (2014) 100111 101two steps to separate them from their similarity variance.The fourth step makes characteristic comparisons fromthese variances. The fth step uses a simple vector distancematching classier to train and extract the characteristics,then recognize the queried object. They are explained asfollows:
2.1. HSV color space
This paper bases on HSV image as shown in Fig. 1 beingas object identication from the original data. In this imageprocess procedure, HSV color-level process was used toconvert a RGB color image to an intensity color-level im-age. Usually, intensity V, saturation S and hue H can be
nts for their focal points.distance to represent the object similarity. The accuracycan reach up to 100% for the above four deformation cases.The similarity can reach 87.5%, 62.5%, 75% and 87.5%respectively. Eventually, this research has a high accuracyfor different angles, sizes and directions. In the following,we will develop the procedure as follows.
2. The procedure of the image processing technique
This research uses some image processing techniquesfor object features extraction, and then bases on the tex-ture and the color features by using the micro structurefeature as an important characteristic reference for objectidentication. The procedure includes the ve steps. Therst step uses HSV color space to deal RGB transformation.The second step uses the micro structure technique to ex-for an object recognition on the mobile. In the feature com-parison part, in order to let every user can take photos byhis perspectives, we need rst to nd out the objects masscenter then transform the centroid into the polar coordi-nates. Using this method, it can solve the image rotationproblem in this object identication. In the experiment,we will transform the original photos into shift (left and
Fig. 5. The similar pixel merger.
102 L.-H. Juang et al. /Measurement 51 (2014) 100111used for RGB to HSV color-level converting which can beexpress as 
H 6 GBMAXMIN 60; if R MAX2 BRMAXMIN 60; if G MAX4 RGMAXMIN 60; if B MAX
Fig. 6. The system processing owchart.
Fig. 7. The process for the mS MAX MINMAX
where MAX =max(R,G,B) and MIN =min(R,G,B) representthe maximal value and the minimal value in the RGB colorspace respectively. The hue H value range is 0360, thesaturation S value range is 0100%, the lower value be-comes more gray level, and the intensity V value range isalso 0100%. In this research, we chose to use HSV colorspace in order to reduce the feature number and also re-duce the effects of chromatic aberration on image, there-fore H is divided into 8 sections, S has 3 sections and Vhas also 3 sections, then the total is 72 colors.
2.2. Micro structure
The researcher  proposed a capturing image methodbased on the micro structure characteristics, and his mainconcept is to use the relationships between texture to de-tect whether there are the same point within a particularregion and its process is shown in Fig. 2. The processingsteps are described as follows:
Step A: When the change for the axial X and Y of gradi-ent information is acquired, rst of all, we use Eqs. (4)(8) to calculate the angle h between the two vectorsa H00x ; S
and b H00y; S
jaj H00x 2 S00x 2 V 00x 2
jbj H00y 2
V 00y 2r
5erger of similar pixel.
will not affect on their comparison features. Fig. 3(a)shows the calculation results for this step scope.Step C: Here we judge these points with the same pixelvalue in the particular region and compare the sur-rounding points with the center point, and then leavethe same, otherwise delete them. For example, the pixelvalue of Fig. 3(b)s focal point is 2 and is surrounded bythe other eight pixels, and its above and its right arealso 2, therefore we only keep these three pixel dataas shown in Fig. 3(c).Step D: After step C processing, we reserve the outlinewhich their pixel values are same with the center pointas shown in Fig. 3(d). Meanwhile, we extend the com-parison rule to the other reference points to do thesame operation for the bigger image size as shown inFig. 4. Its image size is 6 6 pixels, and we cut it intothe four 3 3 Blocks as shown in Fig. 4(a). After usingstep B and step C operations to process Fig. 4(a), theinterval information block is formed as shown inFig. 5(a). Furthermore, we shift the four blocks inFig. 4(a) to the right, the down and the right down byone1X3 block location respectively as shown inFig. 4(b), (c) and (d)s block cutting position. Meanwhilestep B and step C operations are used again to obtainthe results as shown in Fig. 5(b), (c) and (d) respec-tively. Finally we merger these four cutting operationblocks to form the shaded pixel part, then acquire thecharacteristic pixel location as shown in Fig. 5(e),the characteristic pixel location will be recorded by
Fig. 8. The histogram statistics.
L.-H. Juang et al. /Measurement 51 (2014) 100111 103ab H00xH00y S00xS00y V 00xV 00y 6
Fig. 9. The horizontal turn over.Cosda; b abjajjbj 7h Cos1da; b Cos1 abjaj jbj
where a H00x ; S00x ;V
and b H00y; S
are these pixel
value for hue H, saturation S and intensity V with dou-ble rotation at X and Y directions.Step B: When the included angle value is obtained, thenwe use it to calculate the property of edge image tocheck if they are similar. Meanwhile, we set the 30 asone unit and split into 6 intervals (the value is 05).In the case, a video with a deviation of shooting angle
Fig. 10. The vertical turn over.the nal merger pixel value.
Fig. 11. (a) The rst column pixel shift, and (b) the second-fth columnsforward shift.
Fig. 12. The trained image database.
Fig. 13. The test image database examples.
104 L.-H. Juang et al. /Measurement 51 (2014) 100111
2.3. Characteristic extraction
In this research, the characteristic extraction process isdivided into six major steps shown in Fig. 6, the followingis their procedure description:
Step A: A RGB color space image is converted by Eqs.(1)(3) into HSV color space image, then each pixel(R,G,B) in this image is mapped to (H,S,V).Step B: (H,S,V) of each pixel will be quantied into 72colors as the above description (H,S,V) are separatedinto three kind levels of 8, 3, and 3, and the quantitativeresults become (H0,S0,V0). Therefore it can reduce thetime complexity of image processing and improve thetoughness of color identication, Fig. 7(a) shows thequantitative results for this example.Step C: We convert (H0,S0,V0) by Eqs. (9)(11) into theplane coordinate and its conversion result is H00; S00;V 00.
H00 S cosH0 9
S00 S sinH0 10
V 00 V 0 11Step D: Using Sobel edge detection  for the pixel val-ues in the converted plane coordinates calculates the
d feature image example.
Table 1The original image querying accuracy.
Original image Accuracy (%)
Correct number 16 100Similar number Top 3 (10) 100
Table 2The left shifted image querying accuracy.
Original image Accuracy (%)
Correct number 16 100Similar number Top 3 (4) 100
Table 3The right shifted image querying accuracy.
Original image Accuracy (%)
Correct number 16 100Similar number Top 3 (5) 100
Table 4The shrunk image querying accuracy.
Original image Accuracy (%)
Correct number 16 100Similar number Top 3 (12) 100
Table 5The rotated image querying accuracy.
Original image Accuracy (%)
Correct number 128 100Similar number Top 3 (10) 100
L.-H. Juang et al. /Measurement 51 (2014) 100111 105horizontal as well as the vertical gradient value, thenuses the Eqs. (12), and (13) for the shielding edge detec-tion and the upper left corner image is the startingpoint, then from the pixels of left to right and top tobottom in the entire image will use the shield operationin order to acquire the direction change gradient  inX direction, H00x ; S00x ;V 00x, and Y direction, H00y; S00y;V 00y, and
Fig. 14. The extracte
106 L.-H. Juang et al. /Measurement 51 (2014) 100111then the variance between two vectors can be obtainedby the gradual change, meanwhile acquire its vectorangle for the next step.
Gx 1 0 12 0 21 0 1
Gy 1 2 10 0 01 2 1
Step E: We will detect and capture its characteristics byusing the above micro structure section as shown inFig. 5(e), which will combine with HSV color informa-tion in steps B to get the color value of the micro struc-ture for its image characteristics as shown in Fig. 7.Step F: In this step, rst we need to obtain the imagecharacteristic value from its histogram statistics whichare created by computing its frequency distribution ofthe elements in a vector input, its Matlab code is asfollows:
Y histu;n; 14where u is the input vector and n is the number of dis-crete bins. Fig. 7(c) is the results from the numerical cal-culation of the micro structure characteristic value forthe merger of similar pixel. After using the histogram
Fig. 15. The sorting results from the objectstatistics as shown in Fig. 8, then we can acquire itsimage eigenvector which is for converting the true colorimage into the indexed image as follows:
e0;e1; .. .;e710;0;0;0;1;1;0;0;0;0;1;1;0;2;0;1;0;1;1;0;0;1;1;1;1;3;1;0;0;0;0;1;0;0;0;1;1;0;0;1;0;0;0;0;0;0;0;1;2;4;0;0;2;1;0;1;0;0;0;1;0;0;0;0;0;0;0;0;1
2.4. Feature comparisons
In addition to the previous method for the feature com-parison using the original image characteristics, we alsoproposed the other three methods for the feature compar-ison including the horizontal ipping, vertical ipping andpixel shifting. For the horizontal and vertical ipping sec-tion, rst we need to nd its images symmetry axis, thenmake a mirror reversion as shown in Fig. 9(a) for the origi-nal image characteristics and in Fig. 9(b) for the result fromthe mirror reversion of perpendicular to the symmetryaxis. Similarly, Fig. 10(a) is for the original image charac-teristics and Fig. 10(b) is for the result from the mirrorreversion of horizontally to the symmetry axis. For the pix-el shifting section, as shown in Fig. 11(a), the rst columnpixels move to the last column pixels, the second columnpixels are forward to the remaining original columns asshown in Fig. 11(b) and so on. We will shift a bar in eachtime and a total of ve kinds of shiftings are used.
image querying based on Fig. 12(o).
L.-H. Juang et al. /Measurement 51 (2014) 100111 1073. Experiment test
In this research, we use an image database for theexperiment tests which have a total of 16 serial objectimages as shown in Fig. 12(a)(p) for the classicationtests, their image sizes are 342 256 pixels. The test imagedatabase examples include the shrunk, shifted and rotatedimages from the original image database as shown inFig. 13(a)(l). The rotated images include a total of eightrotated angles and the shrunk image size is a half of origi-nal size, 171 128 pixels, the others are same size in thistest image database. The test database has a total of 16shrunk images, a total of 32 shifted images including cen-ter-left and center-right, the original image database has16 images and together the rotated images consist a totalof 8 16 = 128 images, therefore the test image databasehas a total of 192 images. The training image databasehas a total of 12 distortion situation images as shown inFig. 13(a)(l). In this gure, (a) is an example of the originalimage, (b) is an example of the shrunk image, (c) is anexample of the left shifted image, (d) is an example ofthe right shifted image, and (e) to (l) are all examples ofthe rotated images for 0 (360), 45, 90, 135, 180,225, 270 and 315 respectively. In this experiment, weused a histogram statistics to analyze the extracted textureand the color information from these objects to test theiraccuracy. We adopted 6 characteristic values for their
Fig. 16. The sorting results from the left shiftedtexture features. We also adopted HSV processing of 72quantized colors for their color features. Then we usedEuclidean distance to calculate the similarity betweentwo objects as Eq. (15), nally we can present the mostsimilar objects according to the descending sorting.
Ep; q Xn
where p and q are for the database and the querying imagerespectively, i is the statistic number of histogram in eachsection and n is the total statistic numbers.
Fig. 14(a)(i) shows the results for the extracted featureimage example, we can see the objects maintaining theircolor and texture feature information without change.The acquired accuracy rate will be able to calculate thecorrection number and the similarity number as shownin Tables 15 and Figs. 1519 show the original imagequerying, the left shifted object image querying, the rightshifted object image querying, the shrunk object imagequerying, and the rotated object image querying results.We can see the accuracy rate is up to 100% for all of themin the rst sorting results, however the similarity numberis based on the second object sorting according to theabove proposed method. This experiment assumes that(b) and (p) in Fig. 12 are for a group, (e), (i) and (o) arefor another group, and the rest are bottles and steel cupsor mugs for a group. The major work in this experiment
object image querying based on Fig. 12(o).
Fig. 17. The sorting results from the right shifted object image querying based on Fig. 12(o).
Fig. 18. The sorting results from the shrunk object image querying based on Fig. 12(o).
108 L.-H. Juang et al. /Measurement 51 (2014) 100111
L.-H. Juang et al. /Measurement 51 (2014) 100111 109is on the similarity calculation for the rst two groups, wefound that the common textures are almost same but thecolors are different. In this condition, the group of (e), (i)and (o) in the ve cases test for the correction are all100%, and their similarity sorting is on the rst three rank-ings, therefore we dene the accuracy rate is 100% But thegroup of (b) and (p) has a lower accuracy rate and similar-ity, the similarity number in the parenthesis represents theworst sorting. For example, the shrunk images test, wefound that the two objects whose color information is verydifferent which will occur this problem, and the group of(e), (i) and (o) has a very close color in the middle largeblock and the remaining block still has some mixture
Fig. 19. The sorting results from the rotated ob
Fig. 20. The distance comparison.similar color, therefore its tolerance is acceptable. As shownin Tables 15, the similarity can reach 87.5%, 62.5%, 75%and 87.5% respectively. Fig. 20 shows the distance compar-ison result according to Euclidean distance statistics whichis based on Fig. 21 experiment test for the different anglesand the locations with the original image. The E, X, S, and Yrepresent the original image characteristics, horizontalrotation, pixel shifting and vertical rotation respectively.The experimental results show the accuracy rate is 58.7%by using the original feature image processing, howeverthe accuracy rate is up to 83.4% after the horizontal rota-tion, the pixel shifting and the vertical rotation processing.The results also show as long as the result appears on thetop six rankings, the judgment is properly, and otherwisethe judgment is failed. Summarily, when we use only onefeature conversion processing, the effect is the less signi-cant as Fig. 20 E index value shown, when we use the twofeatures conversion processing together, the effect is betteras Fig. 20 EX index value shown, Similarly, when we usethe four kinds of features conversion processing together,the effect is the best as Fig. 20 EXSY index value shown.Fig. 22 shows the object image querying results from amobile phone. The result is based on the four distancecalculation results and we took the top six rankings. Theresults have shown the proposed micro structure schemecan has better performance on mobile device. The reasonis the micro structure is based on the pixel point and smallimage size that are matched for the mobile device featurefor a small data processing quantity and a fast response.
ject image querying based on Fig. 12(o).
110 L.-H. Juang et al. /Measurement 51 (2014) 100111Meanwhile we surveyed the other methods , wefound our proposed micro structure method is best onaccuracy because we used the pixel point which is thesmallest unit.
In this research, we have proposed a realization of ob-ject detection from a mobile device using a simple vector
Fig. 21. The object test images for original image characteristic
Fig. 22. The object image querying result from mobile phone.distance matching. A lot of evaluations on the image data-base shows the encouraging results which accuracy rate isvery acceptable. The method can also reduce the waitingtime of mobile devices when users query because of theless feature vector data and the simple vector distanceoperation. In the future, we would like to build this systemon the real mobile device and take the shooting image di-rectly from its device lens, furthermore issues a query mes-sage from the back-end database, so we can implement areal-time identication function for living objects whichwill be best for the initiatively innovative services.
s, horizontal rotation, pixel shifting and vertical rotation.The authors deeply acknowledge the nancial supportfrom Shantou University, Guangdong, P.R. China underthe STU Scientic Research Foundation for Talents plan.
 http://en.wikipedia.org/wiki/Euclidean_distance. http://en.wikipedia.org/wiki/Content-based_image_retrieval. http://en.wikipedia.org/wiki/Sobel_operator. http://en.wikipedia.org/wiki/Gradient. http://en.wikipedia.org/wiki/HSL_and_HSV. C. Palm, Color texture classication by integrative co-occurrence
matrices, Pattern Recogn. 37 (5) (2004) 965976. G.H. Liu, L. Zhang, et al., Image retrieval based on multi-texton
histogram, Pattern Recogn. 43 (7) (2010) 23802389. G.H. Liu, J.Y. Yang, Image retrieval based on the texton co-occurrence
matrix, Pattern Recogn. 41 (12) (2008) 35213527. J. Luo, D. Crandall, Color object detection using spatial-color joint
probability functions, IEEE Trans. Image Process. 15 (6) (2006) 14431453.
 G.H. Liu et al., Image retrieval based on micro-structure descriptor,Pattern Recogn. 44 (9) (2011) 21232213.
 X. Xiao, Q. Luo, D. Hong, H. Fu, Browsing on small displays bytransforming web pages into hierarchically structured subpages, in:ACM Transactions on the Web, vol. 3, no. 1, 2009.
 E. Bruns, O. Bimber, Adaptive training of video sets for imagerecognition on mobile phones, Pers. Ubiquit. Comput. 13 (2) (2009)165178.
 X. Shu, X.-J. Wu, A novel contour descriptor for 2D shape matchingand its application to image retrieval, Image Vis. Comput. 29 (2011)286294.
 G.H. Liu, Z.Y. Li, L. Zhang, Y. Xu, Image retrieval based on micro-structure descriptor, Pattern Recogn. 44 (9) (2011) 21232133.
 H. Qi, K. Li, Y. Shen, W. Qu, An effective solution for trademark imageretrieval by combining shape description and feature matching,Pattern Recogn. 43 (2010) 20172027.
 C.C. Lai, Y.C. Chen, A user-oriented image retrieval system based oninteractive genetic algorithm, IEEE Trans. Instrum. Meas. 60 (10)(2011) 33183325.
 D. Cerra, M. Datcu, A fast compression-based similarity measurewith applications to content-based image retrieval, J. Vis. Commun.Image Represent. 23 (2) (2012) 293302.
 D.G. Lowe, Object recognition from local scale-invariant features, in:International Conference on Computer Vision, vol. 2, 1999, pp.11501157.
 D.G. Lowe, Distinctive image features from scale-invariantkeypoints, Int. J. Comput. Vision 60 (2) (2004) 91110.
 Robert M. Haralick, K. Shanmugam, Textural features for imageclassication, IEEE Trans. Syst. Man Cybern. 6 (3) (1973) 610621.
 B. Julesz, Textons, the elements of texture perception and theirinteractions, Nature 290 (5802) (1981) 9197.
 B. Julesz, Texton gradients: the texton theory revisited, Biol. Cybern.54 (1986) 245251.
 R.C. Gonzalez, R.E. Woods, Digital Image Processing, third ed.,Prentice-Hall Inc., 2007.
 A.L. Amet, A. Ertuzun, A. Ercil, An efcient method for texture defectdetection: sub-band domain co-occurrence matrices, Image Vis.Comput. 18 (2000) 543553.
 M.J. Swain, D.H. Ballard, Color indexing, Int. J. Comput. Vision 7 (1)(1991) 1132.
 R. Brunelli, O. Mich, Histograms analysis for image retrieval, PatternRecogn. 34 (8) (2001) 16251637.
 R. Min, H.D. Cheng, Effective image retrieval using dominant colordescriptor and fuzzy support vector machine, Pattern Recogn. 42 (1)(2009) 147157.
 Y.K. Chan, C.Y. Chen, Image retrieval system based on color-complexity and color-spatial features, J. Syst. Softw. 71 (2004) 6570.
 T. Hurtut, Y. Gousseau, F. Schmitt, Adaptive image retrieval based onthe spatial organization of colors, Comput. Vis. Image Underst. 112(2008) 101113.
 H. Nezamabadi-pour, E. Kabir, Image retrieval using histograms ofuni-color and bi-color blocks and directional changes in intensitygradient, Pattern Recogn. Lett. 25 (14) (2004) 15471557.
 P.W. Huang, S.K. Dai, Image retrieval by texture similarity, PatternRecogn. 36 (3) (2003) 665679.
 R.M. Haralick, K. Shanmugam, I. Dinstein, Texture features for imageclassication, IEEE Trans. Syst. Man Cybern. 3 (6) (1973) 610621.
 N. Jhanwar, S. Chaudhuri, G. Seetharaman, B. Zavidovique, Contentbased image retrieval using motif co-occurrence matrix, Image Vis.Comput. 22 (14) (2004) 12111220.
 B.M. Mehtre, M.S. Kankanhalli, W.F. Lee, Shape measures for contentbased image retrieval: a comparison, Inf. Process. Manage. 33 (3)(1997) 319337.
 C.H. Wei, Y. Li, W.Y. Chau, C.T. Li, Trademark image retrieval usingsynthetic features for describing global shape and interior structure,Pattern Recogn. 42 (3) (2009) 386394.
L.-H. Juang et al. /Measurement 51 (2014) 100111 111
Object identification using mobile devices1 Introduction2 The procedure of the image processing technique2.1 HSV color space2.2 Micro structure2.3 Characteristic extraction2.4 Feature comparisons
3 Experiment test4 ConclusionsAcknowledgementReferences