GPR Objects Hyperbola Region Feature Extraction

Transcription

1 Advances in Computational Sciences and Technolog ISSN Volume 1, Number 5 (17) pp Research India Publications GPR Objects Hperbola Region Feature Etraction K. Rajiv Department of Information and Technolog, MRIET, Hderabad, India G. Ramesh Chandra Department of Computer Science and Engineering VNR Vignana Jothi Institute of Engg. & Tech., Hderabad, India B. Basaveswara Rao Computer Centre, Achara Nagajuna Universit, Guntur-551, A.P, India. Abstract The shape and feature etraction in Ground Penetrating Radar (GPR) data is the most addressed application in present underground object identification technique. The problem of automaticall detecting the shape and feature in GPR data is in analsing and fitting the hperbola features from GPR data objects. To address these, we propose a hperbole fitting method to etract the shape and features of snthetic and real GPR data objects. The proposed method is successfull implemented on different GPR data sets and comparativel proved better method for GPR data objects feature etraction. Our proposed method can be used in detecting the underground pipe identification technique. Kewords: GPR, corner region, feature etraction, hperbola recognition, 3D surface I. INTRODUCTION The concept of GPR is widel used in detection and mapping of data etracted from underground objects like pipes, cables, metals, and so on. The data etracted using

2 79 K. Rajiv, G. Ramesh Chandra & B. Basaveswara Rao GPR is represented in B-scan images and hperbolic curves. Due to the reflections from objects ling in underground creates a harmonic variations which receives at ground surface located GPR, displas the regions of objects with the effects of cross section, area, segments and impedance. The data collected is displaed in B-scan to analse the reflections in the shape of hperbolic curves. Contribution on this research work has developed different strategies to etract the features of GPR data. Under these contributions, [3] had proposed a Hough transformed based parameters detection of hperbola. As the parameters detection using Hough transform is time consuming, we need a method to detect the feature in GPR data as fast as it can. In [], a unimodal thresholding method is proposed based on totall different computation strategies, could etract ver low threshold level GPR data object features. In [1], a hperbolae fitting technique is proposed using column connection clustering (C3) algorithm, is applied to region of interest from complete GPR data set, which need further modifications. In our proposed method, we improved the computation b introducing Linear Variant Hough Transform technique, utilizing less time to compute the parameters of GPR data objects. We developed a threshold level hierarch to etract the features based on subject of interest. The selection of region is replaced with subject class. Based on class of feature interest the shape and feature etraction of GPR data object is performed in our proposed method. The rest of this paper is organised as follows. We demonstrate the proposed GPR Data feature etraction in section II, which is followed b a description of surface shape feature etraction process, corner region feature etraction process and the hperbola fitting method in GPR data process. In section III, the comparison of the eperimental results are shown, and finall, conclusions are drawn in Section IV. II. PROPOSED REAL TIME GPR DATA FEATURE EXTRACTION : A. Surface Shape Feature Etraction : We shall eplain and demonstrate the Ground Penetrating Radar (GPR) Data basic fundamental features etraction through figure 1, shown as feature class in red bo. Figure a, shows the GPR Data Set taken for analsis and figure b, shows the basic feature etracted automaticall from figure a, without shape information. The feature etraction shown in figure b, gives an approach to find the shapes of features in figure a. Figure c, shows the feature shape of figure a, which is a line shape feature etraction of figure a, which is a first-order line detection. Figure d, shows the second-order line detection, massivel removing the artifacts present in figure c.

3 GPR Objects Hperbola Region Feature Etraction 791 Figure a) Input GPR Data Set Figure b) Feature etracted without shape information of figure a. Figure c) First order line shape feature etraction of figure a. Figure d) Second order line shape feature etraction of figure a. Figure e) The low sharp and low curvature corner shapes of figure a. Figure f) The low sharp and high curvature corner shapes of figure a. Figure g) The high sharp and low curvature corner shapes of figure a.

4 79 K. Rajiv, G. Ramesh Chandra & B. Basaveswara Rao Figure h) The high sharp and high curvature corner shapes of figure a. Figure i) Horizontal corner surface projection of figure a Figure j) Vertical corner surface projection of figure a Figure k) Shape feature etraction of figure a Figure l) 3D surface plot of figure a. Figure 1 : GPR Data Analsis : Line, Corner and Shape Feature etraction with 3D surface feature plot. Detection of feature corner points in figure a, is detecting where the bending of line shape features ver sharpl in the presence of feature curves, as for the figure a, is shown in figure e, f, g and h. Figure e shows the low sharp and low curvature corner shapes, figure f shows the low sharp and high curvature corner shapes, Figure g shows the high sharp and low curvature corner shapes and Figure h shows the high sharp and high curvature corner shapes. The corner feature detection shapes the corner surfaces projected in horizontal and vertical projections of figure a, is shown in figure i and j. All of the features from figure b to j, provides set of features for shape etraction of figure a, as shown in figure h. All these features to be collected and investigated together to find the localized and non-localized features of figure a, as an eample. B. Corner Region Feature Etraction : The properties of GPR Data features are measured to estimate the peak corners to local curvature corners of shape feature b eliminating the artifacts b performing region based analsis on GPR Data. Consider a GPR Data Set for localized GPR Data feature etraction, as in figure a of figure, shown as feature class in red bo. In Figure b of figure illustrate the peak and local curvature corners, is defined b considering a plane curve form.

5 GPR Objects Hperbola Region Feature Etraction 793 Figure a) Input GPR Data Set Figure b) After pre-processing :the peak and local curvature corners. Figure c) After pre-processing :the region based analsis Figure d) 3D surface plot of figure a Figure e) 3D surface plot of figure b Figure : GPR Data Analsis : Corner and Region based analsis with 3D surface feature plot.

6 794 K. Rajiv, G. Ramesh Chandra & B. Basaveswara Rao The corner is epressed as c ( ( C (, Eqn.(1) C describes the points in a plane curve. Here ( defines the end point in -direction, C corner position vector point magnitude in direction is represented b C=[1,], ( defines the end point in -direction, C corner position vector point magnitude in direction is represented b C=[,1] and t define an arbitrar position value. Figure b of figure illustrate the trace of the curve as t is related to curve space. At an point the U and U moves with the relation c... ( ( ( t ), Eqn.() in the direction angle 1 ( ( tan, Eqn.(3) ( The change in position of the curve corner is related with respect to arc length as, d (, where dl is the arc length, along the curve corner, here Ф is the angle of dl tangent of the curve corner. That is Ф= θ±3º,θ±45º, θ±6º and θ±9º, where θ is the gradient direction which represents normal direction to each point in a curve corner. The curve corners are represented as d( dl C k (, Eqn.(4) dl dt dl here relates to the curve arc length change with parameter t, given as dt therefore dl 1 dt (, Eqn.(5) ( C k.. ( ( ( ( ( 3/, Eqn.(6) [ ( ( ] The tangent of curve in figure b given b c(. c( (cos ( j sin( ), Eqn.(7) If this tangential curve equation is related to the curve arc length, the epression for tangential vector in plane curves is given b... d( c( c( ( sin( j cos ( )( ), Eqn.(8) dl

7 GPR Objects Hperbola Region Feature Etraction 795 In general, the tangential curve of the GPR Data curvature is computed b evaluating equation 8. In region based analsis, the description of features in object matching is considered, relating to feature size and rotation with change in illumination of feature in the following manner, ( c(,, C ( ) c(,, ))* P R(,, ) k, Eqn.(9) ( C ( ) here R(.) is region based function with respect to object matching parameter η and the object region scaling functions relates as, R(,, ) (,, Ck ( (,, ), Eqn.(1) here E is the object region scaling function and is total edge regions of the tangential curve which is illustrated in figure c of figure. C. Hperbola Fitting Method in GPR Data : This transformation technique locates the hperbola shapes in GPR data, it utilises the template matching process, based on similar object and target shape and edge detection technique. Hperbola shape detection is ver important in GPR Data analsis, the mapping of these shapes require a transformation technique. In this process, we map an ellipse with an hperbola b a similarit features, that is, ' sin( ) cos( ) F n, Eqn.(11) ' cos( ) sin( ) F n where (',') define the coordinates of the ellipse, σ represents the orientation, (F,F) is a scale factor and (n,n) is a translation factor. If we transform the definitions in equation 11, n a, F sin( ) a andf cos( ) b, Eqn.(1) and n b, F cos( ) a and F cos( ) b, Eqn.(13) then we have equation 11 of ellipse is formed into a a sin( ) b cos( ), Eqn.(14) and o b b sin( ) b cos( ), Eqn.(15) o These equations 14 and 15 corresponds to the hperbola representation containing a, b, a, b, a, b shape characterizing parameters and θ characterizing the shape of o o the hperbola. The orientation of these parameters are related as a tan( ), a a a andb b b, Eqn.(16) a where a and b are the aes of the hperbola. The location of center of the hperbola is given b a a sin( ) b cos( ), Eqn.(17) o k

8 796 K. Rajiv, G. Ramesh Chandra & B. Basaveswara Rao and b o b sin( ) b cos( ), Eqn.(18) Figure 3 shows the GPR Data hperbola etraction process. Figure a shows the GPR Data Set taken for etraction of hperbola from the proposed transformation technique. B change in the position of center of hperbola with respect to scaling factor with the interval range of [.1 :.1 :.9], are shown in figures [b, c, d, e, f, g, h, i, j]. These figures illustrate the etraction of hperbola feature of feature class and also its entire region classes. Figure k shows the feature class segment to derive the hperbola shape under its region of interest. To determine the feature class segment, translation factor is varied within the interval range of [...6], are shown in figures [l, m, n]. Figure n clearl show the parabola feature etraction of feature class subject under consideration. We can observe that there is more than one hperbola to be located in figure a and figure k. This gives rise to the other higher and lower hperbola feature etraction in feature classes of interest, to discover the structures produced particular factors combination. Figure a) Input GPR Data Set Figure b) Scaling factor =.1, of figure a. Figure c) Scaling factor =., of figure a. Figure d) Scaling factor =.3, of figure a. Figure e) Scaling factor =.4, of figure a. Figure f) Scaling factor =.5, of figure a.

9 GPR Objects Hperbola Region Feature Etraction 797 Figure g) Scaling factor =.6, of figure a. Figure h) Scaling factor =.7, of figure a. Figure i) Scaling factor =.8, of figure a. Figure j) Scaling factor =.9, of figure a. Figure k) Feature Class Segment of figure a Figure l) Translation factor =., of figure k. Figure m) Translation factor =.4, of figure k. Figure n) Translation factor =.6,with Hperbola feature etraction, of feature class segment of figure k. Figure o) 3D surface plot of figure a Figure p) 3D surface plot of figure k Figure 3: GPR Data Analsis : Hperbola feature etraction process with 3D surface feature plot.

10 798 K. Rajiv, G. Ramesh Chandra & B. Basaveswara Rao In our proposed method, we discuss onl about hperbola with smmetr ais parallel to -ais in the coordinate sstem. Accordingl the hperbola is defined b equation : H( ) a 1 b 1 c d, Eqn.(19) where a, b, c and d are hperbola point curve search values. Fitting of hperbola for the given curve search is denoted b : F( ) i 1 i 1 i 1, for i=1...m feature constants, Eqn.() The fitting of hperbola equation is given b the value of a, b, c, ). B ( d carring geometr hperbola fitting process, the iterative process is given b H( ) a b 1 c d a 1 b 1 c c c c c c The orthogonal distance for proposed method hperbola is epressed b where O i d d 1, c Eqn.(1) min[( a a( )) ( b b( )) ], Eqn.() ( )) ( )) a andb are the corresponding closest point of determine these values, analse the m O i i d a andb point, to arg min, Eqn.(3) equation (3), analsed results the closest points of hperbola to real, the minimum distance between the points is given b d a z z i b i ( i ) i, Eqn.(4) dz zi z ai bi ( ) zi z equation(4), is used to find the closest point on the current hperbola. The shape of the hperbola indicated b two directive perpendicular straight line. The realization of these projected straight lines is given b epressions dright dstraight dstraight dright dstraight dright, Eqn.(5) dz dz d and where d left dz d straight straight dstraight dleft dstraight dleft dstraight d dstraight, Eqn.(6) dz straight d straight d straight d,, d,, right d right left d, are directive straight, right and left d left left respectivel, distances between right and left points of hperbola directive straight lines. If left and right directive distances are smmetric to straight line, then a region in right and left ais is picket to initialise the hperbola surface points. If those are asmmetric, then the arc angle θ length is initialised to shape the surface of hperbola.

11 GPR Objects Hperbola Region Feature Etraction 799 As an eample, if straight points d straight, d straight are identified, the minimum number of points ie., i=1 is computed in first iteration, if the right and left point is identified. Then this process is repeated for successive number of iteration to fit the hperbola in d right, d right and d, directive perpendicular and straight left d left lines to converge them at their points. Figure 4 and 5 illustrate the proposed hperbola fitting process. Figure a) Input GPR Data Set Figure b) Feature Class Segment of figure a Figure c) After preprocessing : feature constant i=, of figure b. Figure d) After preprocessing : feature constant i=4, of figure b. Figure e) Identified hperbolic signatures, of figure b. Figure f) Hperbolashaped clusters with straight curves, of figure b. Figure g) Hperbolashaped clusters with proposed Thresholding, of figure b. Figure h) Output image of proposed algorithm., of figure b. Figure i) Output image with fitted Hperbolae ( in circle shape), of figure b.

12 8 K. Rajiv, G. Ramesh Chandra & B. Basaveswara Rao Figure j) 3D surface plot of figure a Figure k) 3D surface plot of figure b Figure 4: GPR Data Analsis : Proposed Hperbola fitting process with 3D surface feature plot. Figure a) Input GPR Data Set Figure b) Feature Class Segment of figure a Figure c) After preprocessing : feature constant i=, of figure b. Figure d) After pre-processing : feature constant i=4, of figure b. Figure e) Identified hperbolic signatures, of figure b. Figure f) Hperbolashaped clusters with straight curves, of figure b. Figure g) Hperbolashaped clusters with proposed Thresholding, of figure b. Figure h) Output image of proposed algorithm, of figure b. Figure i) Output image with fitted Hperbolae ( in circle shape), of figure b.

13 GPR Objects Hperbola Region Feature Etraction 81 Figure j) 3D surface plot of figure a Figure k) 3D surface plot of figure b Figure 5 : GPR Data Analsis : Proposed Hperbola fitting process with 3D surface feature plot. Figure 4 and 5 shows the different scenarios of the proposed hperbola recognition in input GPR images. In figure 4, figure a, shows the GPR Data Set taken for analsis. Figure b shows the feature class segment of figure a, taken for analsis of proposed hperbola fitting function. Figures c and d contains the pre-processing results of figure a for feature constant i= and 4 respectivel. Figure e displas the identified hperbolic signature of proposed method. Figures f and g displas the hperbola shaped clusters with straight shaped curves resulting straight, right and left hperbola curves and the varing thresholding at.7 is shown. Figures h and i displas the output hperbola feature etractions of the proposed method. Figure 5 shows the same scenarios, with an another GPR data set input. Figure a) Input GPR Data Set Figure b) After pre-processing :the peak and local curvature corners. Figure c) After pre-processing : Hperbola shape point for =34 and =41.

14 8 K. Rajiv, G. Ramesh Chandra & B. Basaveswara Rao Figure d) After pre-processing : Hperbola shape point for =34 and =39.5 Figure e) After pre-processing : Hperbola shape point for =34 and =38 Figure f) After pre-processing : Hperbola corner detection Figure g) After pre-processing : Proposed Hperbola Fitting result of figure b. Figure 6: GPR Data Analsis : Proposed Hperbola Fitting Method. Figure 6 shows the results of proposed hperbola fitting method, figure 6 c, 6 d and 6 e shows the Hperbola shape point feature detection for d right, d right and d left, d left locations for respective and coordinates. Figure f shows the Hperbola shape feature etraction and figure g shows the proposed Hperbola Fitting of the data set taken in figure 6 b. III.COMPARATIVE RESULTS AND DISCUSSIONS In this section, we have taken GPR snthetic data and real data to compare with C.Maas at al. [3], P.L.Rosin [] and Qingu Dou [1] works. We applied our data set on snthetic data and etracted the different hperbolic fitting features and in second the same methods were used to etract the hperbola in GPR data set taken for analsis. The eperiments results were tabulated in figures 1,, 3, 4 and 5. The performance comparison of our proposed method with [1], [] and [3] are tabulated in table 6. The computation time of the proposed method comparison is shown in table 7.

15 GPR Objects Hperbola Region Feature Etraction 83 Table 6: Feature Detection rate Comparisons Method Recall Precision F-measure Fitting rates of [3] Fitting rates of [] Fitting rate of [1] Fitting rate of our method Table 7: Computation Time Comparison Method Fitting time per hperbola(s) Hough transform based fitting method.97 Proposed in [1].73 Our proposed method.7 From table 6, we can sa, proposed method has improved comparativel in feature etraction and from table 7, the time taken for hperbola fitting is slightl reduced according to the performance of the method proposed. IV.CONCLUSIONS In this paper, an approach for the detection of hperbola feature is proposed. The distance parameters for hperbola fitting method is robust and efficient for fitting right-straight and left-straight hperbole. The method of feature etraction from GPR data samples ehibits ver good performance compared with [1], [] and [3], in terms of accurac and fast enough for real time applications. These features of proposed method can be used in application of GPR data analsis for real time pipes detection. REFERENCES [1]. Qingu Dou, Lijun Wei, Derek R. Magee, and Anthon G. Cohn, 'Real Time Hperbolae Recognition and Fitting in GPR Data', pages []. P. L. Rosin, Unimodal thresholding, Pattern Recognition, pp , 1. [3]. C. Maas and J. Schmalzl, Using pattern recognition to automaticall localize reflection hperbolas in data from ground penetrating radar, vol. 58, pp , 13. [4]. 3D Model Based Approach for Data Visualization [5]. Geometrical Properties of 3-D Images and Its Uses in 3-D Image Processing, "K. Rajiv, G. Ramesh Chandra, B. Basaveswara Rao", International Journal of Advanced Research in Computer and Communication Engineering, ISO 397:7 Certified, Vol. 5, Issue 7, Jul 16.

16 84 K. Rajiv, G. Ramesh Chandra & B. Basaveswara Rao [6]. G. Ciochetto, S. Delbo, P. Gamba, and D. Roccato, Fuzz shell clusering and pipe detection in ground penetrating radar data, IGARSS 99, vol. 5, pp , [7]. R. Janning, A. Busche, T. Horvath, and L. Schmidt-Thieme, Buried pipe localization using an iterative geometric clustering on GPR data, Artifical Inteligence Review, vol. 4, pp , 13. [8]. Rafael C Gonzalez; Digital Image Processing;Pearson Education India, 9.