Online Submission ID: vis-1171

A Visual Approach to Efficient Analysis and Quantification ofDuctile Iron and Reinforced Sprayed Concrete

Category: Application

Fig. 1. CT-based visualization of drill cores from Steel Fibre Reinforced Sprayed Concrete (SFRSpC) enables inspection and analysisof the distribution of the steel fibres’ directions, which is crucial for the material’s mechanical properties. (a) Frequency of occurrenceof directions color-coded in a Direction Sphere Histogram (DSH) (red: highest, blue: lowest); (b) Frontal view; (c) Top-down view.

Abstract—This paper describes advanced volume visualization and quantification for applications in non-destructive testing (NDT),which results in novel and highly effective interactive workflows for NDT practitioners. We employ a visual approach to explore andquantify the features of interest, based on transfer functions in the parameter spaces of concrete application scenarios. Examplesare the orientations of fibres or the roundness of particles. The applicability and effectiveness of our approach is illustrated using twoconcrete scenarios of high practical relevance. First, we discuss the analysis of Steel Fibre Reinforced Sprayed Concrete (SFRSpC).We investigate the orientations of the enclosed steel fibres and their distribution, depending on the concrete’s application direction.This is a crucial step in assessing the material’s behavior under mechanical stress, which is still in its infancy and therefore a hot topicin the building industry. The second application scenario is the designation of the microstructure of ductile cast irons with respectto the contained graphite. This corresponds to the requirements of the ISO standard 945-1, which deals with 2D metallographicsamples. We illustrate how the necessary analysis steps can be carried out much more efficiently using our system for 3D volumes.Overall, we show that a visual approach with custom transfer functions in specific application domains offers significant benefits andhas the potential of greatly improving and optimizing the workflows of domain scientists and engineers.

Index Terms—Non-Destructive Testing, Multi-Dimensional Transfer Functions, Direction Visualization, Volume Rendering.

1 INTRODUCTION

In recent years, the introduction and rapidly increasing use of 3D com-puted tomography in non-destructive testing (NDT) applications hasopened up many powerful possibilities for analyzing and testing in-dustrial parts such as cast metal without destroying them. However,many important tasks in this area such as designating the microstruc-ture of ductile cast irons are still routinely carried out using a (destruc-tive) metallographic approach and subsequent manual visual analysis.That is, the metal under inspection is examined by sequentially re-moving layers via grinding or etching, and taking microscopic pho-tographs. Due to the complexity of the observed structures, in currentpractice the classification of important properties such as the form ofgraphite particles that are crucial for ductility is performed by humanvisual analysis and manual comparison with reference images. Thisprocess could be simplified tremendously by using computed tomog-raphy (CT) and powerful visual tools. However, currently there is aclear lack in visualization solutions that provide sufficient flexibilityfor exploration and inspection, as well as ultimately quantification, todo so. Moreover, the use of CT in non-destructive testing is increas-ingly employed for non-metallic or compound materials. For exam-ple in the analysis of different types of concrete, which probably isthe most widely used non-natural material worldwide. The inspectionof concrete by CT is becoming more widespread, but the subsequentanalyses are very time-consuming and in current practice still involvea considerable amount of manual work due to insufficient tools.

In this paper, we describe novel, visual workflows that are cus-tomized for applications in non-destructive testing which are of highrelevance for NDT practitioners. The presented approaches have beendeveloped jointly with expert users from this domain, both from amore practical engineering and testing perspective, as well as a viewrelating to materials research. The first application scenario that wedescribe is the analysis of Steel Fibre Reinforced Sprayed Concrete(SFRSpC), which is of very high importance in the building industry.We are concerned with the examination and analysis of the enclosedsteel fibres regarding their orientations and distribution with respectto the application direction in which the concrete has been sprayed.SFRSpC is a relatively new material that is used in various applica-tions in geotechnics. However, although the distribution of fibre direc-tions could strongly influence the mechanical properties of the mate-rial, they have not been properly explored yet. Instead, a continuousdistribution was assumed. A preferred orientation of the fibres, how-ever, can result from the application process of the concrete, whichcould significantly alter its mechanical behavior. Our specific exampleis concerned with the examination of this form of sprayed concrete intunnel linings. In a previous study of our domain expert collaborators,drill core samples have been taken and scanned using 3D CT. Then,each single steel fibre was measured individually, which required asignificant amount of tedious work. Additionally, different tools wererequired in order to compute and visualize statistical results. In con-

1

Fig. 2. Region growing in a pre-processing step (left) results in detected features, which can then be explored and quantified interactively (right).The specific requirements of either application scenario are accommodated by pre-computing additional global as well as per-feature statistics forsteel fibres or graphite microstructure, respectively. Default transfer functions (TFs) for the exploration phase are also generated. The user exploresfeatures and feature classes by changing TFs, interactive exploration of statistical parameters, and the 3D volume view. The selected steel fibresare visualized in the Direction Sphere Histogram (DSH), whereas the graphite microstructure is classified according to ISO standard requirements.

trast, our solution provides powerful interactive exploration and auto-matic statistical analysis, based on automatic detection of the enclosedsteel fibres, and the specification of transfer functions in the space offibre directions. Figure 1 shows CT volumes of SFRSpC drill coresand visualizations of the directions of the enclosed steel fibres.

The second application scenario is the designation of the graphitemicrostructure in ductile cast irons, corresponding to the ISO standard945-1. We illustrate how the analysis steps required by this estab-lished standard can be met much more efficiently by our system thanthe standard workflow of testing practitioners. This includes perform-ing classification and quantification of the graphite particles using avisualization-driven approach in the whole 3D volume. This leads tomore accurate results than the examination of 2D slices because theinclusions are not of completely regular shape in all dimensions andtherefore quantification in 2D cross sections has inherent drawbacks.Using slices for quantification also introduces a lot of uncertainty withrespect to the distribution of inclusions, since the results can stronglydepend on the positions where 2D microsections are taken. In our vi-sualization system, expert users can either search the 3D volume foran appropriate position to perform the quantification in a representa-tive 2D slice image, or conduct quantification in a specific part of the3D volume. We compute application-specific information such as his-tograms of particle diameter over density, as well as roundness overdensity, in a pre-process. At run-time, the interactive specification oftransfer functions in these domains allows for exploration and quan-tification of the sample in line with the requirements of the ISO stan-dard. Transfer function presets enable easy classification accordingto parameter ranges such as determining whether the cast iron is ver-micular, intermediate, or nodular, according to the graphite particles’roundness. Additionally, we compute properties such as the nodularityaccording to ISO standard 1611.

Overall, we illustrate the large potential of customized, visual work-flows to improve the daily routine of NDT practitioners, as well asopening up entirely new possibilities for exploration and analysis.

2 RELATED WORK

We first review related work in the specific application domains usedin this paper. The first international standard on graphite classifica-tion by visual analysis was adopted in 1975 as ISO 945:1975. It wasleft nearly unaltered until 2000, when work started on the first part ofa revised version that was published as ISO 945-1:2008. The use ofimage analysis systems in many production laboratories of foundriesthat could lead to a more impartial classification of the graphite mor-phology required appropriate standards for their application. Currentlythere is no information published when this part will become available.To evaluate the 3D shape of the graphite particles, metallographic se-rial sectioning techniques [15] can be used. The need for impartialclassification and evaluation of the graphite microstructure led to thedevelopment of several suggestions to analyze the form of graphite inmetallographic sections [10, 14, 17]. These processes are based on dif-

ferent evaluation schemes such as image analytical parameters, shapeparameters, or morphological transformations. However, fully auto-matic methods are very susceptible to the choice of parameters. Incontrast to the case of ductile cast iron, the norm-guidelines for SFR-SpC were evaluated by the European Committee for Standardization(CEN) in the last few years. This has resulted in general standardsfor sprayed concrete and steel fibres and for their control [13]. Thesestandardization efforts are under constant further development, wherealso national institutions such as concrete or construction societies areinvolved. We are not aware of any analysis of the influence of the fibredistribution in SFRSpC, except mostly manual approaches [1].

In visualization, the previous work most directly related to oursare size-based transfer functions [5], and specifying transfer functionsover volume density, feature size, and region growing parameter [7].The former computes a local size of the region surrounding any givenvoxel using scale space analysis. The latter detects connected regionsand enables transfer functions over the size of these regions. We buildon the second approach since it directly enables quantification of thefeatures of interest, which is crucial in our application context. Sec-tion 3 reviews this basic approach and describes extensions demandedby our application scenarios. In our system, the detection of featuressuch as steel fibres or graphite particles is performed using regiongrowing [18]. In region growing techniques, the specification of seedpoints is crucial. This can be tackled, for example, by evaluating mul-tiple seed points at the same time [2], iteratively specifying seed pointsfor region of interest selection [4], or allowing each voxel in the vol-ume to potentially become a seed [7].

For volume visualization, we use GPU-based ray-casting [12, 16]with 3D textures and multi-dimensional transfer functions [11]. Thisallows for real-time rendering combined with considerable flexibility,including efficient memory managment for large volumes [3]. Re-gion growing techniques have been integrated into volume visualiza-tion systems [8], which can also be used to define two-dimensionaltransfer functions that highlight the regions selected by region grow-ing. This has also been demonstrated to be very useful in the contextof industrial CT data [9]. However, the result is only an approximationand in general cannot reproduce the exact set of voxels of any givenfeature. In our context, it is crucial that all voxels are included in afeature to enable accurate quantification.

3 VOLUME EXPLORATION FOR FEATURE DETECTION ANDQUANTIFICATION

In previous work [7], a novel method for interactive exploration ofindustrial CT volumes such as cast metal parts was presented, whichhelps to bridge the gap between visualization and feature detection aswell as quantification. This visualization-driven approach allows fea-tures in the volume to be explored interactively without re-computingthe segmentation or detection of features. The basis for this is anunattended pre-computation stage that computes a feature volume andadditional data structures, which contain the result of feature detec-

2

Online Submission ID: vis-1171

tion over parameter domains instead of fixed parameters. The pre-computed feature volume tracks a feature size curve for each voxelover time. Thereby time is identified with the main region growingparameter such as the maximum intensity variance allowed within a re-gion. This pre-computation has to be performed only once for a givendata set and forms the basis of interactively exploring all containedfeatures. In contrast to detection of a single type of feature, the userexplores all feature classes and decides interactively which classes areof interest, instead of specifying this information beforehand.

The system computes and records the result of region growing forthe entire density domain, all different sizes of features, and the en-tire domain of the most important parameter. In case a specific regiongrowing algorithm is used, this parameter could be the maximum vari-ance, for example. For generality, this parameter is referred to as thetime parameter throughout the paper. Together, these three 1D pa-rameters (density, feature-size, time) comprise a 3D domain, which isexplored by the user via 3D transfer functions (TFs). In order to makeTF specification tractable, a 2.5D metaphor is employed, which stillprovides the necessary flexibility. This enables a powerful interactiveworkflow that tightly couples visualization and feature detection. Itbuilds on a multi-pass region growing approach, and allows for a fullexploration of the volume with no or almost no beforehand parameterspecification. The result of exploration is a classification of all featureclasses of interest through transfer functions, which can then immedi-ately be used to quantify only the corresponding features. This impliesthat the subsequent quantification is visualization-driven as well, i.e.,quantification is performed exactly for what the user has chosen to bevisualized. This empowers users who are experienced domain expertsto decide on their own and make informed decisions for quantification.This is a significant difference to systems currently used in practice,which rely on the result of a given set of parameters.

For the two application scenarios described in this paper, we areusing this basic approach of visualization-driven quantification. How-ever, in order to accommodate the specific requirements of these appli-cations, several key aspects and corresponding extensions have to beconsidered. In the case of sprayed concrete (SFRSpC), the directionsof steel fibres must be computed, visualized, and analyzed in partic-ular. For classification and analysis of the graphite microstructure ofductile cast iron, the specific requirements of the corresponding ISOstandards must be met.

The specific application-triggered technical contributions of this pa-per are:

• The orientations of the steel fibres in SFRSpC are computed au-tomatically in the pre-computation phase, and visualized interac-tively in a Direction Sphere Histogram (DSH) that color-codesthe frequency of occurrence of directions. This histogram ischanged interactively when a different set of fibres is selectedvia direction transfer functions (below).

• We introduce direction transfer functions (DTFs) for the analysisand visualization of the direction-distribution of the steel fibresin SFRSpC according to the two angles α and γ , which are spec-ified according to the application direction of the concrete.

• In addition to feature-size, we have integrated the specific prop-erties of feature-diameter and feature-roundness required by theISO standards for classification of the graphite microstructureof ductile cast iron. These properties can be used in the ex-tended TF domains of (density, feature-diameter, time) and (den-sity, feature-roundness, time), which enables the advantages ofthe basic approach outlined above to be combined with the re-quirements of the respective standards.

• Quantification of the microstructure of ductile cast irons with re-spect to the contained graphite. This is done according to therequirements of the ISO 945-1 and ISO 16112 standards for sin-gle 2D slice images as well as for the entire or part of the 3Dvolume.

4 STEEL FIBRE REINFORCED SPRAYED CONCRETE

Steel fibre reinforced sprayed concrete (SFRSpC) is a composite mate-rial used for various applications in geotechnics. It combines the highcompressive strength of sprayed concrete with the strength after crack-ing through the addition of the steel fibres. The material behaves brittleuntil a crack starts and becomes ductile through the steel fibres after-wards. Ideally the steel fibres represent a continuous reinforcement, ifthe fibres are randomly and homogenously distributed. For example,SFRSpC is used when a quick application of a shell is needed. Thisis for example the case, when walls or rocks have to be secured as nosteel mashes have to be applied before. SFRSpC can endure higherdisplacements of the rock due to the quasi-ductility of the material.Furthermore, the characteristics of the SFRSpC increase the safety forthe workers, as cracks can be seen long before material failure.

4.1 Application Scenario

Although the mechanical properties of SFRSpC are generally a wellstudied problem, the steel fibres’ orientation and distribution were dis-regarded in the past. Especially the orientation of the fibres has a greatinfluence on the bearing capacity after the formation of cracks. Theaim of this special study, realized at anonymous is a statistical eval-uation of the steel fibres’ alignment in SFRSpC applied on a tunnelwall. The goal is to find out whether there is a prevailing characteristicorientation of the fibres that is due to the application direction of thesprayed concrete. Drill core samples are taken from the tunnel liningand are analyzed by means of industrial CT. In the previous, com-pletely manual approach, the CT volume was loaded into a softwaretool where for each single steel fibre the coordinates of both ends hadto be selected manually. Thereby it was important that each steel fibrewas processed sequentially, to keep the storage order. Additionally,several formatting steps were required to finally get suitable coordi-nates for other tools to accomplish the statistical interpretation. Thiswas tedious, time-consuming, and required the use of several differentsoftware tools. In contrast, our visualization-driven solution simpli-fies the steel fibre evaluation significantly, because it automaticallycomputes and visualizes the directions of the fibres, and also providesdifferent possibilities for statistical evaluation.

4.2 Processing Pipeline

The important stages of the processing pipeline are shown in Figure 2.The additional calculations required for this application are describedin Section 4.2.2. They are performed in the automatic pre-processingphase, which is the basis of the interactive exploration and quantifica-tion process as described in Section 4.2.3.

4.2.1 Data Acquisition

The samples used in this section are core samples of a sprayed concreteshell of a tunnel. When drilling the cores, the axis of the cylindrical

Fig. 3. Drill core SFRSpC 2 taken from the tunnel lining. (a) The longestaxis of the OBB (black rectangle) enclosing the fibre (yellow) representsthe feature’s main direction. The angles α and γ are required for the di-rection quantification of the steel fibres. (b) Direction Sphere Histogram(DSH) of all fibres. (c) The color coding from the DSH can be used inthe 3D volume view for inspection of fibres according to orientation.

3

Fig. 4. The DTFs for α (a) and γ (d), where the x axes show the density, and the y axes are the angle distribution of the steel fibres. For thisapplication, it was sufficient to select one representative 2D time step of the 3D TF [7]. The 3D volume view immediately shows the fibres selectedby the widgets in the α-DTF (b), or the γ-DTF (e), as well as their distributions in the corresponding Direction Sphere Histogram (DSH), (c) and (f).

sample was aligned with the surface normal of the inner tunnel lin-ing. This direction can be assumed to be the direction along which thesprayed concrete has been applied, which is mapped to the z-directionin our application. The data is acquired from core samples using amicrofocus x-ray CT-system scanned on a Phoenix|x-ray v|tome|x cof GI Sensing & Inspection Technologies. For these samples, the res-olution has to be high enough to correctly identify the steel fibres,which have a high x-ray absorption compared to the concrete matrix.A single fibre should be detected as one feature, and the detection offibre fragments should be avoided because these can influence the de-tected orientations. We have investigated two drill cores in particular,denoted as SFRSpC 1 and SFRSpC 2, respectively. Figures 1 and 4depict SFRSpC 1, and Figure 3 depicts SFRSpC 2. See also Table 4.

4.2.2 Pre-computationThe computations necessary for the direction estimation of the steelfibres are performed during the region growing step in the pre-processing phase. After region growing, for each segmented featurean object oriented bounding box (OBB) is calculated, using principalcomponent analysis (PCA). An OBB is a rectangular bounding boxthat approximates the arbitrary orientation of an object in 3D space [6].PCA yields the three eigenvectors for the main axes, where the largesteigenvalue corresponds to the principal axis. The principal axis of thefeature’s OBB is used as direction vector in additional orientation cal-culations, as shown in Figure 3(c). To check the samples for preferredorientations of the contained fibres according to the requirements ofthe real-world analysis to be performed, we compute two main angles:

• Alpha (α) is the angle in the xy-plane normal to the concrete’sapplication direction. It is in the range [0◦,180◦], since a half-circle covers all orientations. See Figure 3.

• Gamma (γ) is the angle between the z-axis (the application di-rection of the sprayed concrete) and the feature direction in 3D.The range for γ is [0◦,180◦] as well. See Figure 3.

For each angle, a 3D histogram is created, where the distributions areshown along the vertical axis. See Figure 4(a) for one time step inthe α-histogram and Figure 4(d) for one time step in the γ-histogram.The angles replace the feature-size described in Section 3. The hor-izontal axis shows the density distribution, and the third dimensioncorresponds to the change of the main region growing parameter overtime, as also described in Section 3. The color coding indicates thenumber of features, where red stands for a high and blue for a low oc-currence of features at the current position. These histograms are thedomains of transfer functions for interactive exploration and quantifi-cation described below.

4.2.3 Interactive Exploration and QuantificationFor this application, the basic framework described in Section 3 hasbeen extended for the visualization, exploration, and quantification ofthe steel fibres in SFRSpC. The exploration of the direction distribu-tion of the steel fibres is performed according to their two main an-gles, via the respective direction transfer function (DTF). Figures 4(a)

and (d) depict the distributions of the fibres according to the angles α

and γ in [0◦,180◦]. By means of these DTFs, the user can explore thesteel fibres according to their angles, using quantification widgets tointeractively control the selection. By dragging the widgets with themouse, their size and position can be adapted to colorize the desiredfeatures in the 3D volume view (Figures 4(b) and (e)). Additionalvalues like the degree range, number of features, and the percentageof the volume covered by features selected by the current widget areconstantly updated and shown, which provides extra information andaids navigation (Table 1). The 3D volume view allows one to inspectthe current selection at any time, and determine if it is satisfactory.Afterwards, the mean and standard deviation of the selected features’angles are calculated on demand. The frequency distribution of theselected fibres’ directions are color-coded on a direction sphere his-togram (DSH). This histogram stores the direction frequency of theselected fibres, which is Gaussian-distributed, at each sphere-vertex.The implementation details are described below. Figure 1(a) showsthe DSH for all features in the sample, where red color indicates ahigh frequency and blue color a low one. Because the two angles α

and γ are in the range [0◦,180◦], it is sufficient to plot the directions ona hemi-sphere. For other applications, this can be switched to an entiresphere. The sphere’s alignment regarding the two angles α and γ , andthe application direction of the sprayed concrete, is demonstrated inFigure 1(a) and Figure 3(a). This alignment is the same for all DSHimages shown here. Additionally, the color coding of the DSH can alsobe used in the 3D volume view, where each fibre is assigned the coloraccording to the frequency of fibres in the same direction (Figure 1(b)and (c)).

4.2.4 Direction Sphere Histogram (DSH) ComputationFor the Direction Sphere Histograms (DSHs) displayed by our systemand shown in Figure 1(a), Figure 3(b), and Figure 4(c,f) we use a tes-sellated sphere as the domain, where each vertex is a histogram binwith a corresponding floating point count that can easily be mapped toa color using a simple 1D color table. The spheres are rendered usingOpenGL, where the colors are mapped to vertex colors and interpola-tion is used within each triangle. For each fibre selected by means ofthe DTFs, its direction is entered into the DSH by splatting a Gaus-sian kernel centered at the fibre’s direction onto the sphere. That is,the histogram count of all sphere vertices covered by the (truncated)Gaussian is incremented by a value corresponding to its location withrespect to the center of the Gaussian. The integral of the Gaussiankernel is 1, and thus we make sure that we obtain a good trade-off be-

Angle degree feature count % of part vol.Alpha 15◦- 25◦ 53 0.14

Fig. 4 (a,b,c) 81◦- 106◦ 37 0.08Gamma 80◦- 110◦ 234 0.60

Fig. 4 (d,e,f) 136◦- 153◦ 9 0.02

Table 1. Quantification of the steel fibres in SFRSpC 1 selected by theclassification widgets shown in Figure 4.

4

Online Submission ID: vis-1171

Fig. 5. The domain of a Direction Sphere Histogram (DSH) is a sphere (or hemi-sphere) that is tessellated using a simple latitude/longitudetessellation. The directions of steel fibres are entered into the histogram by splatting a Gaussian kernel centered at the direction vector onto thesphere vertices (d). The computation is simplified by using a 1D look-up table for one half of a Gaussian (a), which is then accessed to obtainvalues from a 2D Gaussian distribution (b,c).

tween accurate counts at any direction, and a distribution over a smallneighborhood of directions that results in smooth histograms.

For efficient run-time computation of DSHs, we pre-compute asampling of the weights of a Gaussian distribution with a user-definedstandard deviation σ that controls the size of the splat. For the full sizeof the truncated Gaussian kernel, we use a radius of 4σ . For simpleimplementation, the function values according to the distance from thecenter of the Gaussian kernel are stored in a 1D gauss array, as shownin Figure 5(a), for use during splatting. During the tessellation of thesphere, the angular distances and number of steps between two succes-sive longitudes or latitudes of the tessellation, respectively, are storedin a sphere distance array. This enables a fast computation of the areaon the sphere that is covered by a Gaussian of a given size.

To calculate the values for the histogram bins, the contributions ofthe directions of all selected fibre are summed up. According to thepreviously calculated OBBs, the fibres’ main direction is calculatedand the vertices covered by the truncated Gaussian centered at thatlocation are computed. These vertices are then assigned a value ac-cording to their distance to the center of the Gaussian (Figure 5(b)).Figure 5(c) shows this procedure for one fibre, whose Gaussian splatcovers a certain area in longitude and latitude on the sphere. At eachvertex/bin, the contribution of each fibre is summed up. In the laststep, the values of the bins are mapped to colors. Figure 5(d) showsthe DSH Gaussian distribution of one selected fibre. This techniqueensures a fast recalculation of the sphere colors when the selectionof fibres has changed, because only the Gaussian values for the ver-tices have to be recalculated. This is done by simple look-up oper-ations in the gauss array according to the distance looked up in thesphere distance array. The whole sphere initialization process mustonly be recalculated when the σ -value of the Gaussian distributionchanges.

4.3 Discussion and Results

Normally, the quantification is needed for all steel fibres in the vol-ume as shown in Figure 1. In this case, the DTFs are mainly usedto get a quick overview of the fibres’ occurrences at selected angularranges of interest to the user by using the classification widgets. Forthe direction quantification as shown in Figures 1 and 3, a single clas-sification widget was used to mark all steel fibres in the volume, forwhich it does not make a difference which of the two DTFs is taken.Nevertheless, the user has the possibility to explore and quantify onlya selected part of the inclusions interactively, which assists in a fastcomprehension of the data. An example of the selective classificationis shown in Figure 4. Just a part of the steel fibres are selected by theuser employing the quantification widgets, as given in Table 1.

For practical evaluation of our system we have made a comparisonwith the results generated by the completely manual approach per-formed at anonymous. The result of our system was exactly the sameas the manual approach. However, it could be achieved in a fractionof time and provided several additional possibilities for real-time ex-ploration of the data set. The distributions in the γ-DTF (Figure 4(d))as well as in the DSH (Figure 1(a)), clearly show that there occurs ahigh fibre concentration at about γ = 90◦, which is close to a right an-

gle with respect to the concrete’s application direction. Also for theα-direction, the DTF (Figure 4(a)) and the DSH (Figure 1(a)) showa preferred orientation between α = 10◦ and α = 90◦. Our expertcollaborators assume that the application direction in reality was notexactly orthogonal to the tunnel lining, which has lead to the observedorientation according to α . As was often tested by using flexural ten-sion tests, the SFRSpC keeps a remaining force after the building ofcracks because the steel fibres aligned orthogonal to the crack interactagainst the bending direction. Due to the observed preferred orienta-tion of the fibres, its effectiveness increases if they lie parallel to thesurface, and decreases if they lie parallel to the spraying direction, i.e.,normal to the surface.

A direction visualization of the second drill core sample (SFR-SpC 2) taken from the same tunnel lining, including more steel fibres,is shown in Figure 3. This analysis emphasizes the observations de-scribed above. The guidelines (EFNARC European Specification forSprayed Concrete 1996) and standards (EN 14488-7) for sprayed con-crete, currently only provide the examination of the fibre content infresh or already hardened probes. In fresh concrete, the fibres must bewashed out. In already hardened concrete, the fibres have to be me-chanically quarried out of the concrete and cleaned afterwards. Thenthe number of fibres in the sample can be estimated. With our system,it is now possible to estimate not only the number of fibres, but alsotheir orientation in the sample at hand. Because the fibres’ orientationhas a significant impact on the mechanical behavior of SFRSpC, ourexpert collaborators from the application domain think that in generalthe quantification of SFRSpC and the analyses of the related propertiescan be improved tremendously in the future by using such a system,as well as enabling new scientific discoveries in this engineering dis-cipline.

5 DUCTILE CAST IRON

In ductile iron, the majority of the carbon is present as graphite spheresin the microstructure, which exhibits higher strength and ductility thanfor lamellar shaped graphite particles. The main share of ductile castiron is used in the production of pressure pipes, followed by automo-tive parts and different parts for mechanical engineering. To achievethe desired properties of the material, an appropriate microstructurehas to be ensured during production. One major influence is the for-mation of the graphite. Ideally the microstructure consists of small,finely dispersed spheres of graphite all over the casting which giveshighest strength and ductility as well as high fatigue strength at thesame time. Due to different production constraints this can not beachieved in every casting or even in every part of the casting. Forthis reason the examination of the graphite formation is often requiredin quality specifications. To harmonise these examinations the stan-dard ISO 945-1 (ICS: 77.080.10) was established. The classificationof the graphite’s microstructure in cast irons is based on a visual com-parison. In quality specifications for castings, minimum requirementsconcerning graphite form, particle density and maximum particle sizeare common in the casting industry.

5

Fig. 6. The quantification of ductile cast iron according to the ISO 945-1and the 16112 standards is currently accomplished on the basis of 2Dmicroscopic slice images (a), where a representative section is selected(b) to execute the quantification process. (a) and (b) are edited imagesof microscopic photographs. Our system quantifies a 3D part of thevolume to make the process more efficient. (c) shows a cut in the xz-plane, of the part in the CT volume we used for the quantification.

5.1 Application ScenarioDuring the designation of microstructure in cast irons according to therequirements of the standard ISO 945-1, the metallographic samplesof ductile cast iron materials are examined under a microscope, wherethe graphite inclusions are classified by its (1) form, (2) distributionand (3) size. The graphite distribution designation is only specifiedfor a form, which is not allowed to occur in ductile cast iron and istherefore ignored in our process. The standard ISO 945-1 providesreference images of schematic microstructures of cast iron, as a clas-sification basis. Form and size of the graphite inclusions observed atthe sample surface, are determined by comparison with the referenceimages that resemble them most closely. Depending on the locationwhere the metallographic samples were taken, the results can show asignificant variation. Additionally, the preparation of the sample itselfis crucial. Careful grinding and polishing of the sample is important.The graphite particles should appear in their original form and size, sothat an unacceptable alteration of the microstructure is avoided. Thewhole conventional quantification process is limited to 2D, where thecomparison of the actual microstructures with the schematic imagesdepends on the subjective impression of the metallographer. Our 3Dapproach via CT-scanning and interactive visualization-driven explo-ration of the whole data set, can overcome the drawbacks of the con-ventional, manual approach. A higher number of individual particlesare taken into account, and scattered graphite particle deviations canbe found more easily. Additionally, our 3D approach should help toresolve the statistical aberrations occuring when calculating the formfactor of an inherently 3D graphite inclusion by means of a single 2Dslice image. By automizing the quantification step, the results are nolonger constrained to the metallographers’ subjectivity.

5.2 Processing PipelineAn overview of the individual stages of the processing pipeline for thequantification of ductile cast iron is shown in Figure 2.

5.2.1 Data AcquisitionThe sample presented in this section was taken from a thin walledcasting from serial production, which had to be checked for its castingquality. The data is acquired from material samples using a microfo-cus x-ray CT system. The resolution must be chosen high enough toensure that the graphite particles can be represented by several vox-els in the reconstructed CT volume. The accuracy of the analysis isdependent on the average number of voxels per graphite particle.

5.2.2 Pre-computationThe parameters required for graphite microstructure quantification arecomputed in the pre-computation phase described in Section 3 duringregion growing. The relevant indications for quantifying the graphitespheres according to the ISO 945-1 standard are their size, form, and

the nodule count. These are measured by means of the smallest sphereenclosing a particle In the ISO standard, the form of each graphiteparticle is denoted by the roundness of the particle. The nodularityis calculated as a parameter of the dominant form in the microstruc-ture, according to the ISO standard 16112, from the percentages of thedifferent graphite forms in the sample. For each particle, we estimatethe smallest enclosing bounding sphere. In this context, the size of aparticle is defined as the diameter of its enclosing sphere. Figure 7(d)describes the sizes according to the classes defined in ISO 945-1. Toestimate the form factor for a given graphite particle, we have to cal-culate its roundness, which is computed in 3D as follows:

Roundness =3Vparticle

4r3π, (1)

where Vparticle is the volume of the particle, and r is the radius of thesmallest enclosing bounding sphere. The roundness therefore is theratio between the volume of the particle and the volume of the en-closing sphere. According to this property, the particles are classifiedas belonging to one of three relevant (of six overall) different forms,representing the typical types of graphite particles for the investigateddata:

• vermicular (form factor III): 0.0≤ Roundness < 0.525.

• intermediate (form factor V): 0.525≤ Roundness < 0.625.

• nodular (form factor VI): 0.625≤ Roundness < 1.0.

The nodularity, regarding the three different morphology groups asdefined in accordance to ISO 16112, is calculated as follows:

Nod[%] = 100∑Vnodular +0.5∑Vintermediate

∑Vnodular +∑Vintermediate +∑Vvermicular, (2)

where ∑Vnodular, ∑Vintermediate, and ∑Vvermicular are the number ofvoxels of all particles with roundness factor nodular, intermediate,and vermicular, respectively. To estimate the nodule count, only theparticles of form factor V and VI (see Figure 7(c)) per mm3 are of rele-vance. We extend the planimetric method to the 3D domain, such thatwe compute the nodule count n as n = N/Vre f , where N is the num-ber of nodules (i.e., particles classified as forms V and VI) countedin a reference subset of the whole volume chosen for measuring thenodule count, and Vre f is the volume of this subset.

To provide an interactive exploration and quantification process, ac-cording to the parameters relevant for ISO 945-1 and ISO 16112 de-scribed above, we provide two new 3D TFs. Features, i.e., particles arenow classified by means of the 3D domain of density, feature-diameteror feature-roundness, and time. Whereas the density- and time param-eter stay the same as described in Section 3, the feature-size parameteris exchanged by either the feature’s diameter or roundness factor. The2D TFs of one selected time step are shown in Figure 7(a) and (d).The number of features is color-coded, where red corresponds to ahigh and blue to a low occurrence of features at the current position.The backgrounds of the histograms are color-coded according to thegraphite’s size or roundness classes, defined by the ISO standards.

5.2.3 Interactive Exploration and QuantificationFor the user, the exploration stage is the most important part of ourpipeline. During exploration, the current classification is shown inreal-time in a 3D volume view and three orthogonal slice views. Inorder to explore and classify feature classes, the corresponding fea-tures are mapped to color and opacity, so that the user can observe andcontrol the selections made. The quantification of the ductile cast ironaccording to the form parameter of the ISO standard 945-1, is done viathe (density, feature-roundness, time) TF shown in Figure 7(a). Theform-distribution of the graphite’s microstructure is explored via thequantification widgets, where the selection of the graphite particles isconstantly updated in the 3D volume view. As already explained forthe SFRSpC, additional parameters like the roundness- and densityrange, or the nodularity of the features, selected by the quantification

6

Online Submission ID: vis-1171

Fig. 7. To quantify the graphite microstructure according to the form and size parameters of the ISO 945-1 standard, one appropriate time step ofthe respective TFs is used: (a) and (d). The color coding according to the three form factors (b) and the eight size-classes (e) is also applied in the3D volume view. For the nodule count, only the graphite microstructure of forms V I and V are taken into account (c).

Fig. 8. The same visualization and quantification principle as shown in Figure 7, but for one selected 2D slice, taken from the middle of thevolume.Here, the exploration and quantification is based on the same values, but calculated for the 2D case.

widget, are constantly displayed for better navigation. Our systemprovides automatic color coding according to the form factors definedin the ISO standard. Figure 7(a) shows one selected time step t ofthe (density, feature-roundness, time) TF. The feature’s roundness isshown on the vertical axis, where the roundness increases from bot-tom to top in the range [0.0;1.0]. Generally, the user can select dif-ferent feature classes, which are colored corresponding to the widgetcolors and opacity. If desired, the form factor’s color coding can betransferred to the 3D volume view to gain a better all-over impressionof the data, as shown in Figure 7(b) and (c). In contrast to the SFR-SpC, the quantification of individual form-groups may of interest. Inthis case different widgets are placed to mark the desired areas, andare quantified separately. For example, to estimate the nodule count,only features of form factor V and V I are of interest and can thereforebe selected individually (see Figure 7(c)). The same procedure usingthe (density, feature-diameter, time) TF, is performed to estimate thesize-distribution of the graphite particles, as shown in Figure 7(d) and(e). With this TF the user can explore and quantify the different sizeclasses defined in the ISO 945-1 standard. The size on the verticalaxis ranges from 0 (class 8) to > 100 (class 1). The color coding ofthe size-distribution can also be applied to the 3D volume view.

5.3 Discussion and Results

The established standard workflow is currently based on determiningthe classification according to the ISO standards in microscopic 2Dslice images such as the one shown in Figure 6(a,b). Due to the sig-nificant amount of manual work this entails, all properties are only de-termined in chosen 2D reference regions, for example the highlightedrectangle in Figure 6(a). This has huge implications for the accuracyof the obtained results. First, the properties under investigation canchange considerably from one region in the volume to another. Sec-ond, everything is only determined in 2D cross-sections, i.e., com-pletely independent from the actual 3D shape or volume of the con-tained particles. Therefore, efforts are underway to move the wholeprocess to 3D and leverage the possibilities provided by 3D CT ac-quisition. The advantages of extending the quantification to the 3Ddomain were discussed in Section 5.1. We have compared the resultsobtained by using our system with manually obtained results in the

2D slice image shown in Figure 6(b). Table 2 gives the quantificationresults for the evaluated parts shown in Figure 7 and Figure 8. Theresults for 2D data are comparable considering that differences due todifferent reference regions used for inspection are to be expected. Itwas not possible, however, to compare the properties obtained usingour system in 3D, since there are no existing reference data. How-ever, as explained above, the results of quantification in 2D and 3Dare not directly comparable in general. A manual quantification of the3D structures in the data to accurately evaluate our system is underprogress. Our results show that using the coefficients for form classi-fication defined in ISO 16112 for the 2D case have to be exhaustivelyreviewed when applying this procedure to a potentially much more ac-curate 3D application scenario. However, such a review can only bedone using a bigger database and is therefore an ongoing process.

Furthermore, graphite inclusions with a diameter less than threevoxels must not be included in the analysis to avoid the segmenta-tion of noise. Also shrinkage porosity, which is in most cases muchbigger than the graphite particles are not relevant. Therefore, a mini-mum and maximum feature size can be specified for the region grow-ing process to exclude a priori undesired features from quantification.But this is only an option when the user has a-priori knowledge of thedata. However, in our system these undesired features can also easilybe removed interactively using transfer functions. Once the feature in-formation has been calculated in the pre-processing step, the user canquantify the sample under investigation at any desired slice positionand for the whole 3D volume in real time. This overcomes the prob-lems mentioned in Section 5.1 which occur during the conventionalsample preparation. The feature volume is cached and can be quicklyreloaded for further quantification.

6 GENERAL DISCUSSION

In the case of SFRSpC, our approach is much faster and reduces the re-quired labor time considerably, since the previous method was almostcompletely manual. The manual quantification process takes hours forobtaining the same results which we have achieved, even when theshort pre-processing times shown in Table 4 are considered. For theductile cast iron, our quantification process generally takes more timethan the manual approach if only one 2D slice is prepared, because in

7

application form % feature count size feature count % nodularity nodule countDuctile Cast Iron VI: 8.67 1349 3: 504automatic, 3D V: 7.44 1268 4: 2446 12.41 1653 mm−3

Fig. 7 (b,c and e) III: 83.87 9518 5: 7079

6: 2064Ductile Cast Iron VI: 45.30 1485 3: 16automatic, 2D V: 15.23 286 4: 254

Fig. 8 (b,c and e) III: 39.47 799 5: 1017 52.91 102 mm−2

6: 974

7+8: 344Ductile Cast Iron VI: > 90 n/a n/a

manual, 2D V: < 10 n/a 6 n/a 74,80 173 mm−2

Fig. 6 (a and b) III: traces n/a n/a

Table 2. Comparison of the quantification results from the manual 2D approach using the images (a) and (b) in Figure 6, and the 2D and 3D resultsof our system, shown in Figure 7 and Figure 8.

that case the CT scan takes more time than manual sample prepara-tion. But if more 2D samples are required, it takes an additional 20minutes per sample. Table 3 compares times of the manual and ourautomatic approach. Additionally, our system provides many power-ful possibilities for real-time exploration that enable domain expertsto gain more insight about the material sample at hand. Currently, oursystem for automatic quantification according to the ISO 945-1 andISO 16112 standards in 3D is only used for research purposes. It isa first step towards specific quantification tasks in the 3D domain, butfurther research and refinement has to be performed in the future untilthe system can be incorporated in the daily quantification process ofmetallographers. Table 4 gives typical numbers for pre-computationtimes, memory consumption and volume rendering frame rates for thedata sets used in this paper. Performance has been measured on anIntel Core i7 2.95GHz and a GeForce 285 GTX.

7 CONCLUSIONS AND FUTURE WORK

We have presented interactive exploration and quantification of fea-tures for two real-world applications of high practical relevance inthe building industry. We have demonstrated that our system helpsto bridge the gap between visualization, feature detection and quan-tification, and has the potential of significantly improving the currentworkflow of NDT practitioners. In the future we would like to inte-grate more possibilities for interactive measurement, as well as per-forming further evaluations of correctness of the resulting quantifica-tions. This can be accomplished by further comparisons of manualquantifications, as intended for the 3D case of the graphite microstruc-ture in ductile cast iron. The integration of sub-voxel precision is alsoan important future goal.

REFERENCES

[1] anonymous.[2] R. Adams and L. Bischof. Seeded region growing. IEEE Trans. Pattern

Anal. Mach. Intell., 16(6):641–647, 1994.

∼ time for... manual 2D automatic 3Dpre-processing 60 min (sample preparation) 90 min (CT scan)further samples 20 min -

analysis 10 min 10 min70 min - ... 100 min

Table 3. Time table comparing the conventional manual approach withour automatic method for quantification of ductile cast iron.

data set resolution pre-computation renderingSFRSpC 1 440x440x705 1.7 min 20-33 fpsSFRSpC 2 400x400x570 1.2 min 25-33 fps

Ductile Cast Iron 932x932x128 21.8 min 37-69 fps

Table 4. The data sets from this paper, with typical pre-computationtimes and typical volume rendering frame rates (viewport 512x512).

[3] J. Beyer, M. Hadwiger, T. Moller, and L. Fritz. Smooth mixed-resolutiongpu-based raycasting. In IEEE/EG Symposium on Volume and Point-Based Graphics 2008, pages 163–170, 2008.

[4] M. F. Cohen, J. Painter, M. Mehta, and K.-L. Ma. Volume seedlings. InProc. ACM Symp. on Interactive 3D Graphics, pages 139–145, 1992.

[5] C. Correa and K.-L. Ma. Size-based transfer functions: A new volumeexploration technique. IEEE Transactions on Visualization and ComputerGraphics, 14(6):1380–1387, 2008.

[6] S. Gottschalk, M. C. Lin, and D. Manocha. Obbtree: A hierarchicalstructure for rapid interference detection. Computer Graphics, 30(AnnualConference Series):171–180, 1996.

[7] M. Hadwiger, L. Fritz, C. Rezk-Salama, T. Hollt, G. Geier, and T. Pabel.Interactive volume exploration for feature detection and quantificationin industrial ct data. IEEE Transactions on Visualization and ComputerGraphics, 14(6):1507–1514, 2008.

[8] R. Huang and K.-L. Ma. RGVis: Region growing based visualizationtechniques for volume visualization. In Proc. Pacific Graphics, 2003.

[9] R. Huang, K.-L. Ma, P. McCormick, and W. Ward. Visualizing industrialct volume data for nondestructive testing applications. In ProceedingsIEEE Visualization 2003, pages 547–554, 2003.

[10] H. Kerber, G. Schindelbacher, G. Ruess, A. Kneissl, and K. Kutschej.Microstructure characterization of spheroidal graphite cast iron by usingvarious image analytical systems. Practical Metallography 40 (7), pages335–342, 2003.

[11] J. Kniss, G. Kindlmann, and C. Hansen. Interactive volume rendering us-ing multi-dimensional transfer functions and direct manipulation widgets.In Proceedings IEEE Visualization 2001, pages 255–262, 2001.

[12] J. Kruger and R. Westermann. Acceleration techniques for GPU-basedvolume rendering. In Proceedings IEEE Visualization 2003, pages 287–292, 2003.

[13] H. Schorn. Classification of teel fibre reinforced shotcretes according torecent european standards. In Bernard, E.S. (Hg.): Shotcrete: Engineer-ing Developments, pages 225–229. Swets & Zeitlinger, 2001.

[14] A. Scozzafava, L. Tomesani, and A. Zucchelli. Image analysis automa-tion of spheroidal cast iron. Journal of Materials Processing Technology,153–154:853–859, 2004.

[15] H. Singh, A. Gokhale, Y. Mao, and A. Tewari. Reconstruction, visualiza-tion, and quantitative characterization of multi-phase three-dimensionalmicrostructures of cast aluminum alloys. In TMS Shape Casting Sympo-sium, pages 15–19, 2009.

[16] S. Stegmaier, M. Strengert, T. Klein, and T. Ertl. A simple and flexiblevolume rendering framework for graphics-hardware-based raycasting. InVolume Graphics, pages 187–195, 2005.

[17] I. W. Steller, W. Stets, J. Ohser, and D. Hartmann. Computer-aidedgraphite classification: An approach for international standardization.Transactions of the American Foundry Society Vol. 113, pages 587–594,2005.

[18] S. Zucker. Region growing: Childhood and adolescence. ComputerGraphics and Image Processing, 5:382–399, 1976.

8