X-ray Testing: The State of the Art

Domingo MeryDepartment of Computer Science – Pontificia Universidad Catolica de Chile

Av. Vicuna Mackenna 4860(143) – Santiago de Chiledmery@ing.puc.cl - http://dmery.ing.puc.cl

Abstract

X-ray imaging has been developed not only for its use inmedical imaging for human beings, but also for materialsor objects, where the aim is to analyze –nondestructively–those inner parts that are undetectable to the naked eye.Thus, X-ray testing is used to determine if a test object de-viates from a given set of specifications. Typical applica-tions are analysis of food products, screening of baggage,inspection of automotive parts, and quality control of welds.In order to achieve efficient and effective X-ray testing, au-tomated and semi-automated systems are being developedto execute this task. In this paper, we present a generaloverview of computer vision methodologies that have beenused in X-ray testing. In addition, we review some tech-niques that have been applied in certain relevant applica-tions; and we introduce a public database of X-ray imagesthat can be used for testing and evaluation of image anal-ysis and computer vision algorithms. Finally, we concludethat the following: that there are some areas –like castinginspection– where automated systems are very effective, andother application areas –such as baggage screening– wherehuman inspection is still used; there are certain applicationareas –like weld and cargo inspections– where the processis semi-automatic; and there is some research in areas –including food analysis– where processes are beginning tobe characterized by the use of X-ray imaging.

1. IntroductionSince Rontgen discovered in year 1895 that X-rays can

be used to identify inner structures, X-rays have been de-veloped not only for their use in medical imaging forhuman beings, but also in nondestructive testing (NDT)for materials or objects, where the aim is to analyze –nondestructively– the inner parts that are undetectable tothe naked eye. NDT with X-rays, called X-ray testing, isused in many applications such as: analysis of food prod-ucts, screening of baggages, inspection of automotive parts,quality control of welds, among others. X-ray testing usu-

ally involves measurement of specific part features such asintegrity or geometric dimensions in order to detect, recog-nize or evaluate wanted (or unwanted) inner parts.

In order to achieve efficient and effective X-ray testing,automated and semi-automated systems are being devel-oped to execute this difficult, tedious and –sometimes– dan-gerous task. Compared to manual X-ray testing, automatedsystems offer advantages of objectivity and reproducibilityfor every test. Fundamental disadvantages are, however, thecomplexity of their configuration, the inflexibility to anychange in the evaluation process, and sometimes the in-ability to analyze intricate images, which is something thatpeople can generally do well. Research and developmentis, however, ongoing into automated adaptive processes toaccommodate modifications.

In this paper, we present the state of the art in X-ray test-ing using computer vision. We describe a general overviewof computer vision methodologies that have been used inthe last years (Section 2). In addition, we review sometechniques that have been applied in certain relevant appli-cations using X-ray testing (Section 3). Finally, we intro-duce a public database of X-ray images that can be usedfor testing and evaluation of image analysis and computervision algorithms (Section 4). The paper ends with someconcluding remarks (Section 5). A summarized version ofthis paper was recently presented in [37].

2. PrinciplesIn this Section, the principles that governs the X-ray test-

ing by computer vision are presented as a general model ac-cording to Fig. 1. Applications on X-ray testing –as willbe seen in next Section– follows this general schema. De-pending on the way the X-ray images are acquired and ana-lyzed, each block can be (or can be not) used. For instance,in baggage inspection there are methods that analyze sin-gle or multiple views, using monoenergetic or dual-energysystems. For each one of these four combinations, only thecorresponding blocks of Fig. 1 are to be used as illustratedwith the color squares.

This Section covers i) X-ray image formation, ii) single

1

Figure 1. General schema for X-ray testing using computer vision. X-ray images of a test object can generated at different positionsand different energy levels. Depending on the application, each block of this diagram can be (or can be not) used. For example, thereare applications like weld inspection that use a segmentation of a single mono energetic X-ray image (black square) and sometimes withpattern recognition approaches (red squares); applications like casting inspection that use mono energetic multiple views where the decisionis taken analyzing individual views (green squares) or corresponding multiple views (blue squares); applications like baggage screeningthat use dual-energy of single views (magenta squares) and multiple views (yellow squares); and finally, applications on cargo inspectionthat use active vision where a next best view is set according to the information of a single view (cyan squares). In each case, the blockswithout the corresponding color square are not used.

view analysis, iii) geometric model and iv) multiple viewanalysis.

2.1. X-ray image formation

X-ray testing is a non-destructive testing (NDT) definedas a task that uses X-ray imaging to determine if a test objectdeviates from a given set of specifications, without chang-ing or altering that object in any way [21]. In X-ray testing,X-ray radiation is passed through the test object, and a de-tector captures an X-ray image corresponding to the radi-ation intensity attenuated by the object1. According to the

1X-rays can be absorbed or scattered by the test object. In this paperwe will cover the first interaction only. For an interesting application based

principle of photoelectric absorption [3]:

I = I0 exp(−µz), (1)

the transmitted intensity I depends on the incident radiationintensity I0, the thickness z of the test object, and the energydependent linear attenuation coefficient µ associated withthe material, as illustrated in Fig. 2a.

The most widely X-ray imaging systems used in X-raytesting are digital radiography (DR) and computed tomogra-phy (CT) imaging2. On the one hand, DR emphasizes high

on the X-ray scattering effect, the reader is referred to [61].2Computed tomography is out of scope of this paper. For NDT appli-

cations using CT, the reader is referred to [15].

Figure 2. X-ray image formation according to absorption law: a)X-ray image of a homogenous object, and b) X-ray image of anobject with two different materials.

throughput. It uses electronic sensors (instead of traditionalradiographic film) to obtain a digital X-ray projection of thetarget object, for this reason it is simple and quick. A flatamorphous silicon detector can be used as image sensorsin X-ray testing systems. In such detectors, using a semi-conductor, energy from the X-ray is converted directly intoan electrical signal that can be digitalized into an X-ray dig-ital image [49]. On the other hand, CT imaging provides across-section image of the target object so that each objectis clearly separated from each other, however, CT imagingrequires a considerable number of projections to reconstructan accurate cross-section image which is time-consuming.

It is wort noting that if X-ray radiation passes throughn different materials, with attenuation coefficients µi andthickness zi, for i = 1, . . . n, the transmitted intensity I canbe expressed as

I = I0 exp

(−∑i

µizi

). (2)

This explains the image generation of regions, that arepresent inside the test object, as shown in Fig. 2b, wherea gas bubble is clearly detectable. Nevertheless, sometimesX-ray images contains overlapped objects so that it is ex-tremely difficult to distinguish them.

Coefficient µ in (1) can be modeled as µ = α(Z,E)ρ,where ρ is the density of the material, and α(Z,E) is themass attenuation coefficient that depends on the atomicnumber of the material Z, and the energy E of the X-rayphotons. Values for α(Z,E) are already measured andavailable in several tables (see [22]). In order to identifythe material composition –typically for explosives or drugs

detection– the atomic number Z cannot be estimated usingonly one image, because a thin material with a high atomicnumber can have the same absorption as a thick materialwith a low atomic number [61]. For this purpose, a dual-energy system is used, where the object is irradiated with ahigh energy level E1 and a low level energy E2. In the firstcase, the absorbed energy depends mainly on the density ofthe material. In the second case, however, the absorbed en-ergy depends primarily on the effective atomic number andthe thickness of the material [55]. Using dual-energy, it ispossible to calculate the ratio R = ln(I2/I0)/ ln(I1/I0),where I1 and I2 are the transmitted intensities I obtainedby (1) using energies E1 and E2 respectively. Thus, fromR = α(Z,E2)/α(Z,E1), the term −ρz is canceled out,Z can be directly found using the known measurementsα(Z,E) [20]. From both images, a new image is generatedusing a fusion model, usually a look-up-table that producespseudo color information [13, 4], as shown in Fig. 3.

2.2. Single view analysis

A computer vision system for single view analysis, asshown in Fig. 1, consists typically of the following steps:An X-ray image of the test object is taken and stored in acomputer. The digital image is improved in order to en-hance the details. The X-ray image of the parts of interest isfound and isolated from the background of the scene. Sig-nificant features of the segmented parts are extracted. Se-lected features are classified or analyzed in order to deter-mine if the test object deviates from a given set of specifica-tions. Using a supervised pattern recognition methodology,the selection of the features and the training of the classi-fier are performed using representative images that are to belabeled by experts [11].

For the segmentation task, two general approaches canbe used: a traditional image segmentation or a sliding–window approach. In the first case, image processing al-gorithms are used (e.g. histograms, edge detection, mor-phological operations, filtering, etc. [16]). Nevertheless,inherent limitations of traditional segmentation algorithmsfor complex tasks and increasing computational power havefostered the emergence of an alternative approach based onthe so-called sliding–window paradigm. Sliding–windowapproaches have established themselves as state–of–art incomputer vision problems where a visually complex objectmust be separated from the background (see for examplesuccessful applications in face detection [59] and human de-tection [8]). In sliding–window approach, a detection win-dow is moved over an input image in both horizontal andvertical directions, and for each localization of the detectionwindow, a classifier decides to which class belongs the cor-responding portion of the image according to its features.Here, a set of candidate image areas are selected and all ofthem are fed to the subsequent parts of the image analy-

Figure 3. Generation of a pseudo-color image using dual-energy. In this example orange is for organic materials, and blue is for metals.The intensity of this image corresponds to the thickness of the materials.

sis algorithm. This resembles a brute force approach wherethe algorithm explores a large set of possible segmentations,and at the end the most suitable is selected by the classifi-cation steps. An example for weld inspection using sliding-windows can be found in [36].

2.3. Geometric model

The X-ray image of a test object corresponds to a pro-jection in perspective, where a 3D point of the test object isviewed as a pixel in the digital X-ray image, as illustratedin Fig. 4. A geometric model that describes this projectioncan be very useful for 3D reconstruction and for data asso-ciation between different views of the same object. Thus,3D features or multiple view 2D features can be used to im-prove the diagnosis performed by using a single view.

For the geometric model, four coordinate systems areused (see Fig. 4):• OCS (X,Y, Z): Object Coordinate System, where a 3Dpoint is defined using coordinates attached to the test object.• WCS (X, Y , Z): World Coordinate System, where the

origin corresponds to the optical center (X-ray source) andthe Z axis is perpendicular to the projection plane of thedetector.• PCS (x, y): Projection Coordinate System, where the 3Dpoint is projected into projection plane Z = f , and the ori-gin is the intersection of this plane with Z axis.• ICS (u, v): Image Coordinate System, where a projectedpoint is viewed in the image. In this case, (x, y)–axes areset to be parallel to (u, v)–axes.

The geometric model OCS → ICS, i.e. transformationP : (X,Y, Z) → (u, v), can be expressed in homogeneouscoordinates as [32]:

λ

uv1

= P

XYZ1

, (3)

where λ is a scale factor and P is a 3 × 4 matrix modeledas three transformations:i) OCS → WCS, i.e. transformation T1 : (X,Y, Z) →

Figure 4. Geometric model: a 3D point (X,Y, Z) is projected as a 2D point (u, v).

(X, Y , Z), using a 3D rotation matrix R, and 3D transla-tion vector t;ii) WCS → (PCS), i.e. transformation T2 : (X, Y , Z) →(x, y), using a perspective projection matrix that dependson focal distance f ; andiii) PCS → ICS, i.e. transformation T3 : (x, y) → (u, v),using scales factor αx and αy , and 2D translation vector(u0, v0).The three transformations OCS→WCS→ PCS→ ICS areexpressed as:

P =

αx 0 u00 αx v00 0 1

︸ ︷︷ ︸

T3

f 0 0 00 f 0 00 0 1 0

︸ ︷︷ ︸

T2

[R t0T 1

]︸ ︷︷ ︸

T1

(4)The parameters included in matrix P can be estimated usinga calibration approach as illustrated in Fig. 5 [19].

In order to obtain multiple views of the object, m differ-ent projections of the test object can be achieved by rotatingand translating it (for this task a manipulator can be used asshown in Fig. 6). For the k-th projection, for k = 1 . . .m,the geometric model Pk used in (3) is computed from (4)including 3D rotation matrix Rk and 3D translation tk. Ma-trices Pk can be estimated using a calibration object pro-jected in them different positions [32] or using a bundle ad-justment algorithm where the geometric model is obtainedfrom the m X-ray images of the test object [35].

2.4. Multiple view analysis

It is well known that an image says more than thou-sand words, however, this is not always true if we havean intricate image. In certain X-ray applications, e.g. bag-gage inspection, there are usually intricate X-ray imagesdue to overlapping parts inside the test object, where each

Figure 6. Generally a manipulator is used to locate the test objectin a desired position.

pixel corresponds to the attenuation of multiple parts, as ex-pressed in (2).

In some cases, active vision can be used in order to ad-equate the viewpoint of the test object to obtain more suit-able X-ray images to analyze. Therefore, an algorithm isdesigned for guiding the manipulator of the X-ray imagingsystem to poses where the detection performance should behigher [48] (see Fig. 1).

In other cases, multiple view analysis can be a pow-erful option for examining complex objects where uncer-tainty can lead to misinterpretation. Multiple view analy-sis offers advantages not only in 3D interpretation. Two ormore views of the same object taken from different view-points can be used to confirm and improve the diagnosticdone by analyzing only one image. In the computer visioncommunity, there are many important contributions in mul-tiple view analysis (e.g. object class detection [56], motionsegmentation [64], simultaneous localization and mapping(SLAM) [24], 3D reconstruction [2], people tracking [12],breast cancer detection [58] and quality control [6]). Inthese fields, the use of multiple view information yields asignificant improvement in performance.

Figure 5. Geometric calibration of a computer vision X-ray imaging system: the modeled projection of a 3D model coincides with realgeometry of the calibration object.

Multiple view analysis in X-ray testing can be used toachieve two main goals: i) analysis of 2D correspondingfeatures across the multiple views, and ii) analysis of 3Dfeatures obtained from a 3D reconstruction approach. Inboth cases, the attempt is made to gain relevant informationabout the test object. For instance, in order to validate asingle view detection –filtering out false alarms– 2D corre-sponding features can be analyzed [38]. On the other hand,if the geometric dimension of a inner part must be measureda 3D reconstruction must be performed [44].

As illustrated in Fig. 1, the input of the multiple viewanalysis is the associated data, i.e. corresponding points (orpatches) across the multiple views. To this end, associated2D cues are found using geometric constraints (e.g. epipolargeometry and multifocal tensors [19, 33]), and local scale-invariant descriptors across multiple views (e.g. like SIFT[29]).

Finally, 2D or 3D features of the associated data can beextracted and selected, and a classifier can be trained us-ing the same pattern recognition methodology explained inSection 2.2.

Depending on the application, the output could be a mea-surement (e.g. the volume of the inspected inner part is3.4cm3), a class (e.g. the test object is defective) or an inter-pretation (e.g. the baggage should be inspected by a humanoperator because the uncertainty is high).

3. Applications

In this Section, the state of the art of certain relevant ap-plications on X-ray testing will be described. This sectioncovers X-ray testing in i) casting, ii) weld, iii) baggage, iv)food, and v) cargo. In Table 1, approaches in each appli-cation are summarized showing how they use the differentsteps of general diagram according to Fig. 1.

3.1. Casting

Light-alloy castings produced for the automotive indus-try, such as wheel rims, steering knuckles and steeringgear boxes are considered important components for overallroadworthiness. Non-homogeneous regions can be formedwithin the work piece in the production process. These aremanifested, for example, by bubble-shaped voids, fractures,inclusions or slag formation. To ensure the safety of con-struction, it is necessary to check every part thoroughly us-ing X-ray testing. In casting inspection, automated X-raysystems have not only raised quality, through repeated ob-jective inspections and improved processes, but have alsoincreased productivity and consistency by reducing laborcosts. An example is illustrated in Fig. 7. A survey can befound in [34]. Selected approaches are summarized in Ta-ble 1. In this area, we conclude that automated systems arevery effective, because inspection task is fast and achieves

Table 1. Applications on X-ray testing

Appli- First Refe- Energy Geometric Single Views(∗∗) Active Multiple Views(∗∗)

cation author Year rence Mono Dual Model(∗) 1 2 3 Vision 1 2 3Casting

Carrasco 2011 [5] × N × × × × × ×Li 2006 [26] × – × ×Mery 2002 [38] × C × × × × × ×Pieringer 2010 [46] × C × × × × × ×Tang 2009 [57] × – ×Mery 2011 [35] × N × × × × × ×

WeldLiao 2008 [27] × – × × ×Liao 2009 [28] × – × × ×Mery 2011 [36] × – × × ×Shi 2007 [51] × – ×

BaggageAbusaeeda 2011 [1] × × C ×Bastan 2012 [4] × × – × × ×Chen 2005 [7] × × – ×Ding 2006 [9] × × – ×Franzel 2012 [13] × × C × × × × × ×Heitz 2010 [20] × × – × × ×Mansoor 2012 [31] × × – × × ×Mery 2011 [35] × N × × × × × ×Mery 2013 [40] × N × × × × × ×Mery 2013 [41] × N × × × × × ×Qiang 2006 [30] × × – × ×Rahman 2010 [47] × × – ×Riffo 2011 [48] × N × × ×Schmidt 2012 [50] × × – × × ×Singh 2005 [54] × × – ×

FoodHaff 2004 [17] × – × × ×Jiang 2008 [23] × – × × ×Ogawa 2003 [45] × – × × ×Kwon 2008 [25] × – × × ×Mery 2011 [39] × – × × ×

CargoDuan 2008 [10] × – × × ×Frosio 2011 [14] × – × × × ×Zhu 2008 [63] × – × × ×

(*) C: Calibrated, N: Not calibrated, –: not used. (**) See 1 , 2 , 3 in Fig. 1

high performance.

3.2. Weld

In welding process, a mandatory inspection using X-raytesting is required in order to detect defects like porosity,inclusion, lack of fusion, lack of penetration and cracks. In-dustrial X-ray images of welds is widely used for the detect-ing those defects in the petroleum, chemical, nuclear, naval,aeronautics and civil construction industries, among others.An example is illustrated in Fig. 8. A survey can be foundin [52, 53]. Selected approaches are summarized in Table 1.

As we can see there is much research on weld inspection.Achieved performance of the developed algorithms is stillnot high enough, hence it is not suitable for fully automatedinspection.

3.3. Baggage

Since 11/9, automated (or semi-automated) 3D recog-nition using X-ray images has become a very importantissue in baggage screening [62, 43]. The inspection pro-cess is very complex because threat items are very difficultto detect when placed in close packed bags, superimposed

Figure 9. Detection of a guns based on the trigger identification in multiple views [40].

Figure 7. Detection of very smalls flaws in an aluminum wheelusing multiple views [35].

by other objects, and/or rotated showing an unrecognizableview. During the last decade, however, relevant researchhave been done in multiple view analysis and dual-energyimaging. The use of multiple view information yields a sig-nificant improvement in performance because certain itemsare difficult to be recognized using only one viewpoint [60].On the other hand, unsing dual-energy it is possible to

Figure 8. Weld inspection using a sliding-window: a) X-ray image,b) detected windows, c) activation map, d) detection [36].

identify the material composition –typically for explosives,drugs and organic materials– [55]. An example is illustratedin Fig. 9. A survey on explosives detection can be found in[55, 61]. Selected approaches are summarized in Table 1.We conclude, that in baggage screening, where human se-curity plays an important role and inspection complexity isvery high, human inspectors are still used. For intricate con-ditions, multiple view X-ray inspection using dual-energy isrequired.

3.4. Food

In order to ensure food safety inspection, several applica-tions have been developed in food industry. The difficultiesinherent in the detection of defects and contaminants in foodproducts have limited the use of X-ray into packaged foodssector. However, the need for NDT has motivated a consid-erable research effort in this field spanning many decades[18]. The most important advances are: detection of for-eign objects in packaged foods [25]; detection of fishbonesin fishes [39]; identification of insect infestation in citrus[23]; detection of codling moth larvae in apples [18]; fruitquality inspection like split-pits, water content distributionand internal structure [45]; and detection of larval stages ofthe granary weevil in wheat kernels [17]. In these applica-tion, only single view analysis is required. An example isillustrated in Fig. 10. A survey can be found in [18]. InTable 1, the mentioned applications are summarized.

3.5. Cargo

With the ongoing development of international trade,cargo inspection becomes more and more important. X-raytesting has been used for the evaluation of the contents ofcargo, trucks, containers, and passenger vehicles to detectthe possible presence of many types of contraband. Someapproaches are presented in Table 1. There still is not muchresearch on cargo inspection, and complexity of this inspec-tion task is very high. For this reason, we conclude thatthese X-ray systems still are semi-automatic only, and theyrequire human supervision.

4. Data Bases

Public databases of X-ray images can be found for med-ical imaging3, however, to the best knowledge of the au-thor of this article, up until now there have not been publicdatabases of digital X-ray images for X-ray testing.

As a service to the X-ray testing community, we col-lected more than 3.000 X-ray images for the development,testing and evaluation of image analysis and computer vi-sion algorithms. The images are organized in a public

3See for example a good colection in http://www.via.cornell.edu/databases/

Figure 11. Some X-ray images of our public database GDXray:wood, razor blade, aluminum wheel, bag with a gun, fishbones,weld, pen case shrunken, calibration pattern and apples.

database called GDXray: The GRIMA X-ray database4.The database includes three groups of X-ray images: metalobjects (castings, welds, razor blades, ninja stars (shuriken),guns, knives and sink strainers), baggages (bags and pencases); and natural objects (fruits, fishbones and wood).Some examples are illustrated in Fig. 7–11.

5. ConclusionsIn this paper we presented a general overview of com-

puter vision methodologies that have been used in X-raytesting, as illustrated in diagram of Fig. 1. The presentedapplications on X-ray testing follows this general schema,where depending on the way the X-ray images are acquiredand analyzed, each step can be (or can be not) used.

In the presented applications, we observe that thereare some areas, like casting inspection, where automatedsystems are very effective; and other application areas,like baggage screening, where human inspection is stillused. Additionally, there are certain application areas, likeweld and cargo inspection, where the inspection is semi-automatic. Finally, there is some research in food analysiswhere they are beginning to be characterized using X-rayimaging.

It is clear that many research directions have been ex-ploited, some very different principles have been adoptedand a wide variety of algorithms have been developed forvery different applications. Nevertheless, automated X-ray testing remains an open question because it still suf-fers from: i) loss of generality because approaches devel-oped for one application may not be used in other one; ii)deficient detection accuracy because commonly there is afundamental trade–off between false alarms and miss de-tections; iii) limited robustness because prerequisites for theuse of a method are often fulfilled in simple cases only; and

4GRIMA is the name of our Machine Intelligence Group at the De-partment of Computer Science of the Pontificia Universidad Catolica deChile http://grima.ing.puc.cl. The X-ray images included inGDXray can be used free of charge, for research and educational purposesonly. Redistribution and commercial use is prohibited. Any researcher re-porting results which use this database should acknowledge the GDXraydatabase by citing [42].

Figure 10. Detection of fish bones using sliding-windows [39].

iv) low adaptiveness because it may be very difficult to ac-commodate an automated system to design modifications ordifferent objects.

Compared to manual X-ray testing, automated systemsoffer advantages of objectivity and reproducibility for ev-ery test. Fundamental disadvantages are, however, the com-plexity of their configuration, the inflexibility to any changein the evaluation process, and sometimes the inability to an-alyze intricate images, which is something that people cangenerally do well. Research and development is, however,ongoing into automated adaptive processes to accommodatemodifications.

Acknowledgments

This work was supported by grant Fondecyt No.1130934 from CONICYT–Chile. The author wants to thankM. Bastan from Image Understanding and Pattern Recogni-tion Group, at Technical University of Kaiserslautern, Ger-many, for sharing pseudocolor image of Fig. 3.

References[1] O. Abusaeeda, J. Evans, D. D., and J. Chan. View synthesis

of KDEX imagery for 3D security X-ray imaging. In Proc.4th International Conference on Imaging for Crime Detec-tion and Prevention (ICDP-2011), 2011.

[2] S. Agarwal, N. Snavely, I. Simon, S. Seitz, and R. Szeliski.Building Rome in a day. In IEEE 12th International Confer-ence on Computer Vision (ICCV2009), pages 72–79, 2009.

[3] J. Als-Neielsen and D. McMorrow. Elements of modern X-ray physics. Willey, second edition, 2011.

[4] M. Bastan, M. R. Yousefi, and T. M. Breuel. Visual wordson baggage X-ray images. In Computer Analysis of Imagesand Patterns, pages 360–368. Springer, 2011.

[5] M. Carrasco and D. Mery. Automatic multiple view in-spection using geometrical tracking and feature analysis

in aluminum wheels. Machine Vision and Applications,22(1):157–170, 2011.

[6] M. Carrasco, L. Pizarro, and D. Mery. Visual inspec-tion of glass bottlenecks by multiple-view analysis. In-ternational Journal of Computer Integrated Manufacturing,23(10):925–941, 2010.

[7] Z. Chen, Y. Zheng, B. R. Abidi, D. L. Page, and M. A.Abidi. A combinational approach to the fusion, de-noisingand enhancement of dual-energy x-ray luggage images. InIEEE Conference on Computer Vision and Pattern Recogni-tion (CVPR-2005), Workshop CDROM, 2005.

[8] N. Dalal and B. Triggs. Histograms of oriented gradientsfor human detection. In European Conference on ComputerVision (ECCV), 2005.

[9] J. Ding, Y. Li, X. Xu, and L. Wang. X-ray image segmen-tation by attribute relational graph matching. In 8th IEEEInternational Conference on Signal Processing, volume 2,2006.

[10] X. Duan, J. Cheng, L. Zhang, Y. Xing, Z. Chen, andZ. Zhao. X-ray cargo container inspection system with few-view projection imaging. Nuclear Instruments and Methodsin Physics Research A, 598:439–444, 2009.

[11] R. Duda, P. Hart, and D. Stork. Pattern Classification. JohnWiley & Sons, Inc., New York, 2 edition, 2001.

[12] R. Eshel and Y. Moses. Tracking in a dense crowd usingmultiple cameras. International Journal of Computer Vision,88:129–43, 2010.

[13] T. Franzel, U. Schmidt, and S. Roth. Object detection inmulti-view X-ray images. Pattern Recognition, pages 144–154, 2012.

[14] I. Frosio, N. Borghese, F. Lissandrello, G. Venturino, andG. Rotondo. Optimized acquisition geometry for X-ray in-spection. In Instrumentation and Measurement TechnologyConference (I2MTC), 2011 IEEE, pages 1–6, 2011.

[15] J. Goebbels. Handbook of Technical Diagnostics, chapterComputed Tomography, pages 249–258. Springer, 2013.

[16] R. Gonzalez and R. Woods. Digital Image Processing, 2nded. Addison-Wesley, third edition, 2007.

[17] R. Haff and D. Slaughter. Real-time X-ray inspection ofwheat for infestation by the granary weevil, sitophilus gra-narius (l.). Transactions of the American Society of Agricul-tural Engineers, 47:531–537, 2004.

[18] R. P. Haff and N. Toyofuku. X-ray detection of defects andcontaminants in the food industry. Sensing and Instrumenta-tion for Food Quality and Safety, 2(4):262–273, 2008.

[19] R. I. Hartley and A. Zisserman. Multiple view geometry incomputer vision. Cambridge University Press, second edi-tion, 2003.

[20] G. Heitz and G. Chechik. Object separation in X-ray imagesets. In IEEE Conference on Computer Vision and PatternRecognition (CVPR-2010), pages 2093–2100, 2010.

[21] C. Hellier. Handbook of Nondestructive Evaluation. Mc-Graw Hill, second edition, 2013.

[22] J. Hubbell and S. Seltzer. Tables of X-Ray mass attenuationcoefficients and mass energy-absorption coefficients from 1keV to 20 MeV for elements Z = 1 to 92 and 48 additionalsubstances of dosimetric interest. http://www.nist.gov/pml/data/xraycoef/index.cfm, 1996.

[23] J.-A. Jiang, H.-Y. Chang, K.-H. Wu, C.-S. Ouyang, M.-M.Yang, E.-C. Yang, T.-W. Chen, and T.-T. L. a. An adaptiveimage segmentation algorithm for x-ray quarantine inspec-tion of selected fruits. Computers and Electronics in Agri-culture, 60:190–200, 2008.

[24] K. Konolige and M. Agrawal. FrameSLAM: from bundleadjustment to realtime visual mapping. IEEE Transactionson Robotics, 24(5):1066–1077, 2008.

[25] J.-S. Kwon, J.-M. Lee, and W.-Y. Kim. Real-time detectionof foreign objects using x-ray imaging for dry food manu-facturing line. In Proceedings of IEEE International Sym-posium on Consumer Electronics (ISCE 2008), pages 1–4,April 2008.

[26] X. Li, S. K. Tso, , X.-P. Guan, and Q. Huang. Improv-ing automatic detection of defects in castings by applyingwavelet technique. IEEE Transactions on Industrial Elec-tronics, 53(6):1927–1934, 2006.

[27] T. Liao. Classification of weld flaws with imbalanced classdata. Expert Systems with Applications, 35(3):1041–1052,2008.

[28] T. W. Liao. Improving the accuracy of computer-aided ra-diographic weld inspection by feature selection. NDT&E In-ternational, 42:229239, 2009.

[29] D. Lowe. Distinctive image features from scale-invariantkeypoints. International Journal of Computer Vision,60(2):91–110, 2004.

[30] Q. Lu and R. Conners. Using image processing methods toimprove the explosive detection accuracy. IEEE Transac-tions on Applications and Reviews , Part C: Systems, Man,and Cybernetics, 36(6):750–760, 2006.

[31] M. Mansoor and R. Rajashankari. Detection of concealedweapons in X-ray images using fuzzy K-NN. InternationalJournal of Computer Science, Engineering and InformationTechnology, 2(2), 2012.

[32] D. Mery. Explicit geometric model of a radioscopic imagingsystem. NDT & E International, 36(8):587–599, 2003.

[33] D. Mery. Exploiting multiple view geometry in X-ray test-ing: Part I, theory. Materials Evaluation, 61(11):1226–1233,November 2003.

[34] D. Mery. Automated radioscopic testing of aluminum diecastings. Materials Evaluation, 64(2):135–143, 2006.

[35] D. Mery. Automated detection in complex objects using atracking algorithm in multiple X-ray views. In Proceedingsof the 8th IEEE Workshop on Object Tracking and Classifi-cation Beyond the Visible Spectrum (OTCBVS 2011), in Con-junction with CVPR 2011, Colorado Springs, pages 41–48,2011.

[36] D. Mery. Automated detection of welding discontinuitieswithout segmentation. Materials Evaluation, (June):657–663, 2011.

[37] D. Mery. X-ray testing by computer vision. In Proceedingsof the 9th IEEE CVPR workshop on Perception Beyond theVisible Spectrum, Portland, 2013.

[38] D. Mery and D. Filbert. Automated flaw detection in alu-minum castings based on the tracking of potential defects ina radioscopic image sequence. IEEE Trans. Robotics andAutomation, 18(6):890–901, December 2002.

[39] D. Mery, I. Lillo, V. Riffo, A. Soto, A. Cipriano, and J. Aguil-era. Automated fish bone detection using X-ray testing. Jour-nal of Food Engineering, 2011(105):485–492, 2011.

[40] D. Mery, G. Mondragon, V. Riffo, , and I. Zuccar. Detectionof regular objects in baggage using multiple x-ray views. In-sight, 55(1):16–20, 2013.

[41] D. Mery, V. Riffo, I. Zuccar, and C. Pieringer. Automated X-ray object recognition using an efficient search algorithm inmultiple views. In Proceedings of the 9th IEEE CVPR work-shop on Perception Beyond the Visible Spectrum, Portland,2013.

[42] D. Mery, U. Zscherpel, V. Riffo, G. Mondragon, I. Lillo,I. Zuccar, and H. Lobel. GDXray – the GRIMA database ofX-ray images. http://dmery.ing.puc.cl/index.php/material/gdxray/. Department of ComputerScience, Universidad Catolica de Chile.

[43] S. Michel, S. Koller, J. de Ruiter, R. Moerland, M. Hoger-vorst, and A. Schwaninger. Computer-based training in-creases efficiency in X-Ray image interpretation by aviationsecurity screeners. pages 201–206, Oct. 2007.

[44] A. Noble, R. Gupta, J. Mundy, A. Schmitz, and R. Hart-ley. High precision X-ray stereo for automated 3D CAD-based inspection. IEEE Trans. Robotics and Automation,14(2):292–302, 1998.

[45] Y. Ogawa, N. Kondo, and S. Shibusawa. Inside quality eval-uation of fruit by x-ray image. volume 2, pages 1360–1365vol.2, July 2003.

[46] C. Pieringer and D. Mery. Flaw detection in aluminium diecastings using simultaneous combination of multiple views.Insight, 52(10):548–552, 2010.

[47] S. M. Rahman, M. O. Ahmad, and M. Swamy. Contrast-based fusion of noisy images using discrete wavelet trans-form. Image Processing, IET, 4(5):374–384, 2010.

[48] V. Riffo and D. Mery. Active X-ray testing of complex ob-jects. Insight, 54(1):28–35, 2012.

[49] J. Rowlands. The physics of computed radiography. Physicsin medicine and biology, 47(23):R123, 2002.

[50] L. Schmidt-Hackenberg, M. R. Yousefi, and T. M. Breuel.Visual cortex inspired features for object detection in X-rayimages. In Pattern Recognition (ICPR), 2012 21st Interna-tional Conference on, pages 2573–2576. IEEE, 2012.

[51] D.-H. Shi, T. Gang, S.-Y. Yang, and Y. Yuan. Researchon segmentation and distribution features of small defectsin precision weldments with complex structure. NDT & EInternational, 40:397–404, 2007.

[52] R. Silva and D. Mery. State-of-the-art of weld seam inspec-tion using X-ray testing: Part i image processing. MaterialsEvaluation, 65(6):643–647, 2007.

[53] R. Silva and D. Mery. State-of-the-art of weld seam inspec-tion using X-ray testing: Part ii pattern recognition. Materi-als Evaluation, 65(9):833–838, 2007.

[54] M. Singh and S. Singh. Optimizing image enhancement forscreening luggage at airports. In Computational Intelligencefor Homeland Security and Personal Safety, 2005. CIHSPS2005. Proceedings of the 2005 IEEE International Confer-ence on, pages 131 –136, 31 2005-april 1 2005.

[55] S. Singh and M. Singh. Explosives detection systems (eds)for aviation security. Signal Processing, 83(1):31–55, 2003.

[56] H. Su, M. Sun, L. Fei-Fei, and S. Savarese. Learning a densemulti-view representation for detection, viewpoint classifi-cation and synthesis of object categories. In InternationalConference on Computer Vision (ICCV2009), 2009.

[57] Y. Tang, X. Zhang, X. Li, and X. Guan. Application of a newimage segmentation method to detection of defects in cast-ings. The International Journal of Advanced ManufacturingTechnology, 43(5-6):431–439, 2009.

[58] J. Teubl and H. Bischof. Comparison of Multiple ViewStrategies to Reduce False Positives in Breast Imaging. Dig-ital Mammography, pages 537–544, 2010.

[59] P. Viola and M. Jones. Robust real-time object detection.International Journal of Computer Vision, 57(2):137–154,2004.

[60] C. von Bastian, A. Schwaninger, and S. Michel. Do Multi-view X-ray Systems Improve X-ray Image Interpretation inAirport Security Screening?, volume 52. GRIN Verlag, 2010.

[61] K. Wells and D. Bradley. A review of X-ray explosives de-tection techniques for checked baggage. Applied Radiationand Isotopes, 2012.

[62] G. Zentai. X-ray imaging for homeland security. ImagingSystems and Techniques, 2008. IST 2008. IEEE InternationalWorkshop on, pages 1–6, Sept. 2008.

[63] Z. Zhu, L. Zhao, and J. Lei. 3d measurements in cargoinspection with a gamma-ray linear pushbroom stereo sys-tem. In Proceedings of the 2005 IEEE Computer SocietyConference on Computer Vision and Pattern Recognition(CVPR05), 2005.

[64] V. Zografos, K. Nordberg, and L. Ellis. Sparse motionsegmentation using multiple six-point consistencies. InProceedings of the Asian Conference on Computer Vision(ACCV2010), 2010.