The result is a file that describes the page to be printed with all the images, graphics and text. In order to process ..... [14] Android, âAndroid development area.
Detection of Commercial Offset Printing in Counterfeited Banknotes Anton Pfeifera , Eugen Gillicha , Volker Lohwega and Johannes Schaedeb a inIT
Institute Industrial IT, Langenbruch 6, D-32657 Lemgo, Germany; SA, Avenue du Grey 55, CH-1018 Lausanne, Switzerland
b KBA-NotaSys
ABSTRACT Today, a major part of counterfeits is produced by commercial offset printing machines. The counterfeiters use mainly raster printing (rosette printing). Our approach aims at analysing the banknotes for the presence of commercial printing procedures by means of intrinsic features. For this purpose, the images are analysed in view to periodic printing patterns by means of the Discrete Fourier Transformation (DFT). The typical raster frequencies are detected. The rosette pattern is separated by suppression of the remaining image frequencies and the subsequent back-transformation. Keywords: Commercial Printing, Printing Features, Feature Extraction, Pattern Recognition
1. INTRODUCTION Banknotes are important ambassadors of a country. They contribute to an eye-catching image of states, represent identification symbols for their inhabitants and leave first impressions on visitors. The confidence expressed by the users attributes to their value. Therefore, counterfeit protection is and will always remain an important objective of the central banks.1 In previous approaches2, 3 we had presented a reliable authentication approach for security prints. The so-called Soundof-Intaglio method is based on the intrinsic characteristics of Intaglio print and is applied as a possible authentication for many banknotes worldwide. However, it is worthwhile to extend the above mentioned approach to other printing methods like Offset. Many large volume forgeries are printed in Offset. This counterfeit type is often produced with traditional commercial offset printing methods, but also mimics partly Intaglio print. To further enhance the detection of forgeries we apply our commercial offset recognition method in combination with Sound-of-Intaglio. For this purpose, the images under test are analysed in respect to periodic printing patterns or intrinsic periodic motifs. In addition to the periodic pattern detection the raster of each printed colour is detected and analysed in order to avoid any confusion with a possible banknote structure. The paper is organized as follows: In section two necessary information about commercial printing techniques is summed up. In the third section the approach is presented, comprising a mathematical model for the description of the characteristic offset pattern features and image-processing methods to analyse them. Following, results are presented in section four before the paper is concluded in section five.
2. PREREQUISITES Nowadays, a large number of counterfeit banknotes are produced by commercial offset printing machines.4 The counterfeiters use mainly raster print. (offset-print). Also commonly used Laser and Ink Jet printing systems are suitable for producing high-quality counterfeit banknotes. Here, the production flow of counterfeits respectively printed products is in general divided into three steps (cf. Fig. 1): prepress, press, and postpress.5 Prepress describes the preparatory steps of the artwork for the subsequent press. This can be divided in printing technologies with permanent printing master and Non-Impact Printing (NIP). Printing technologies with a printing master are e.g. Lithography (Offset Printing), Gravure Printing, Screen Printing, and Letterpress Printing. Among the most widely used NIP methods are: Electrophotography and Ink Jet.5 In the subsequent postpress optical effects (gloss increase) or a protection against mechanical pollution are applied. The different press techniques together with the prepress approach are the theoretical foundations on which the algorithms for the detection of commercial offset printing are developed. In the following specific details regarding prepress are explained. Further author information: (Send correspondence to Anton Pfeifer) E-mail: {anton.pfeifer, eugen.gillich, volker.lohweg}@hs-owl.de
Optical Document Security, San Francisco, 10-12 February 2016
© Reconnaissance International 2016
Original
Prepress
Press
Printed product
Postpress
Figure 1: Production flow for printed products.5
2.1 Digital Prepress Digital prepress is a sub-process in the reproduction technology and sums up all the steps prior to the actual printing process. These include e.g. scan works, data processing, image editing, word processing, and the design of the page layout. The result is a file that describes the page to be printed with all the images, graphics and text. In order to process the data, three methods are available: Computer to Film (CTF), Computer to Plate (CTP), and Computer to Press (Direct Imaging).5 All these methods have in common that the associated equipment or machine technology is driven by a customized raster image processor (RIP). It has the task of adjusting the provided data to the print job.5 An important part of the adjustment is screening. Screening. Many of today’s printing processes are binary processes: colour is applied to the printed medium or not. However, in most cases the image data colour depth is 8 − 32 Bit and cannot be reproduced with binary printing processes. Therefore, in monochromatic images reproduction of tonal values is executed through a specific arrangement of binary dots in screen cells. By using this halftone screening continuous colour transitions are simulated.5 There are essentially two screening methods to simulate the colour transitions: The frequency-modulated screening (FM) and the amplitudemodulated screening (AM).5 In case of frequency-modulated screening all dots have the same diameter and are arranged in different distances to one another within the screen cell. In amplitude-modulated screening, the dots are aligned centred in the screen cell and arranged in the same distances to one another. Here, the dot size is varied depending on the desired hue. If the surface is bright, the dots are small. If the surface is dark, the dots are large. The closer the dots are next to each other, the finer colour transitions and finer detail can be reproduced in print. The spacing of the halftone dots to each other is defined by w. The reciprocal of the dot spacing w is a measurement for the resolution of the screened images and is referred as screen frequency L. This indicates how many dots are printed per unit length and is expressed in lines per centimetre (l/cm):5 1 . (1) L= w Multicoloured pictures can be reproduced by overprinting of the individual colour screens. This is only possible without interference if the individual screens are aligned in specific angles. If this is not the case moir´e patterns5 will arise, which impair the impression of the printed media. In empirical studies it was found that the rotation of the screen at 30◦ in relation to each other, the moir´e patterns are almost completely eliminated on the printed image. However, a rotation by 30◦ allows only three different screen angles due to the symmetrical screen. For this reason the least distinctive colour, e. g. yellow, is aligned in an angle of 15◦ in relation to magenta and cyan. Thus, the standard screen angles of 0◦ , 15◦ , 45◦ and 75◦ arise.5 The individual screen lines and screen angles are sketched in Fig. 2 for the two-, three-, and four-colour printing with exemplary colour assignments. Second Colour 75° 30 °
Black 45°
Second Colour 75° 30
Black 45°
°
Magenta 75° 30
Black 45°
°
°
° 45
30
°
30
Third Colour 15°
15°
Cyan 15° Yellow 0°
(a)
(b)
(c)
Figure 2: Different colour screen angles of two- (a), three- (b) and four-colour press (c).5
Optical Document Security, San Francisco, 10-12 February 2016
© Reconnaissance International 2016
Despite the optimal screen angle to reduce interferences the superimposition of the screens may form a moir´e pattern called offset rosette. In Fig. 3 the structures are shown. On closer examination there are two types of structures. The first one is the dot-centred rosette, which has a point in the centre of the rosette (cf. Fig. 3b). The arrangement of points is referred as clear-centred rosette (cf. Fig. 3a). In practice, accurate representations of the two forms are hard to find, since every smallest deviation has influence regarding the shape of the rosette.5
(a)
(b)
Figure 3: Comparison between clear-centred rosette (a) and dot-centred rosette (b).
3. APPROACH Today, various printing methods are available. In the consumer area the most used printing techniques are based on Ink Jet or Laser which are getting better in terms of print quality. For high copy runs they are unsuitable because of the high printing costs per sheet. For commercial use various printing methods are applied depending on the application, such as Lithographic printing (offset printing), screen printing, rotogravure printing or digital printing.5 Our approach is based on the detection of commercial offset printing. The development of the model is divided into two essential steps — i) the description of two relevant characteristics of printed media; ii) the development of image processing algorithms.
3.1 Screen Model The description for the screen model is based on the reproduction of continuous colour transitions in a digital printing process. In this process only a limited number of inks is available. To reproduce a variety of colours from this limited number of pure inks, the multicoloured images are broken down into dots of different sizes. This process is called halftone screening and is described in Sec. 2. From the halftone screening two characteristics features are derived and described mathematically: the colour screen rotation and the offset rosette. Colour Screen Rotation. In order to reproduce multicoloured images the original is broken down into screens of the basic colour of the printing process. To avoid interferences (moir´e patterns), the screens are overprinted at different angles. Commonly used angles for the screens colours cyan, magenta, yellow, and key (black) are 15◦ , 75◦ , 0◦ and 45◦ .
Optical Document Security, San Francisco, 10-12 February 2016
© Reconnaissance International 2016
The assignment of the individual colours on the angle varies depending on the application. However, the rotation in relation to each other will remain equal. In case of independent colour screen angles, it is deduced that an images is printed in commercial offset. For the subsequent image processing algorithms it is assumed that the screen dots are arranged on a line (cf. Fig. 4b): ρ = x · cos (θ) + y · sin (θ) . (2) The orthogonal distance from the origin is ρ and θ is the angle to the x-axis. With this parametrization, the differences between the individual screen angles are computed, resulting in the identification of the printing process, if the resulting rotation to each other will remain equal to 15◦ , 30◦ or 45◦ . Offset Rosette. The offset model is based on continuous colour transitions in the digital press flow. In this process a variety of different colour screens are overprinted in defined angles. As a result of the superimposed screens of cyan, magenta, and black a moir´e pattern is formed, called offset rosette (cf. Sec. 2). In order to analyse this pattern a rosette is defined as follows: (x − xp )2 + (y − yp )2 − rp2 = 0,
(3)
with the point (xp , yp ) for the centre of the circle and the point rp for the radius of the rosette.
y
Magenta 75°
y
30
(a)
k
k
Cyan 15° Yellow 0°
θ ρ
15°
θc
c
x
°
y
ρ
°
θy
ρ
xp
45
p
30
y
Black 45°
°
rp
θm
x
ρ
m
(b)
Figure 4: Description of an offset rosette (a) and the characteristic colour screen rotation (b).
3.2 Image Processing The approach for analysing commercial offset is divided into 5 image processing steps: image acquisition, preprocessing, segmentation, feature extraction, and decision making. After an image is acquired a Region-of-Interest (ROI) is selected. In the preprocessing step colour separation is generated. This includes the individual spectral colour components generation.6 Based on the above mentioned step the corresponding colour screens are separated. In the next step a segmentation of each raster is performed. Search strategies are applied to identify distinctive regions and discontinuities. Thus leads to separation of the colour rasters and rosette structures of irrelevant background information. The final step describes the evaluation of the segmented colour screens. Based on the model description of commercial offset the characteristic features are determined. In detail: Preprocessing. After an image is acquired, the raw image data are projected in the RGB space (Red-Green-Blue). In order to separate the individual colour screens, the first preprocessing step consists of a colour transformation6, 7 (cf. Fig. 5). For this purpose, raw image data are converted into the CMYK colour space (Cyan-Magenta-Yellow-Key). In the second step the individual colour channels are transformed by means of the two-dimensional discrete Fourier transform (2D DFT). This step allows to identify the periodical screens within the separated channels. For the transformation of a two-dimensional image signal g(u, v) with the width M (u, columns) and the height N (v, rows) the 2D DFT is
Optical Document Security, San Francisco, 10-12 February 2016
© Reconnaissance International 2016
(a)
(b)
(c)
Figure 5: Separation of the colour channels yellow (b) and cyan (c). defined as follows:8 G(m, n) = √
M −1 N −1 X X mu nv 1 g(u, v) · e−i2π M · e−i2π N , M N u=0 v=0
(4)
with the spectral components n = 0 . . . N − 1 and m = 0 . . . M − 1. For the following analysis of the spectrum, the centred amplitude spectrum |G(m, n)| is computed8 (cf. Fig. 6): q |G(m, n)| = G2Re (m, n) + G2Im (m, n) . (5)
(a)
(b)
Figure 6: Fourier transform of an original (a) and a reproduced five Euro banknote (b). In Fig. 6 the Discrete Fourier spectrum of an original and a reproduced five Euro banknote is shown. The image pattern of the banknote is represented in the centre (low-frequency coefficients). An examination of Fig. 6b shows that
Optical Document Security, San Francisco, 10-12 February 2016
© Reconnaissance International 2016
the reproduced note is additionally superimposed with peaks around the centre. To ensure that the peaks represent the periodic colour screens, the characteristic dots are localized in the frequency domain and the associated patterns are extracted. Segmentation. In order to extract a pattern from the image the first step is to localize characteristic frequencies in the Fourier domain and to suppress unimportant frequencies. In the second step the remaining image frequencies are transformed back with the inverse DFT. Finally, a model-based segmentation is applied to determine the halftoning screens and offset rosette pattern. The circularly arranged peaks in the amplitude spectrum |G(m, n)| represent a periodic structure in the separated colour channel. To segment the associated pattern, the peaks in the high frequency domain are located. In case of a greyscale amplitude spectrum, the peaks have a higher brightness (greyscale) than their direct neighbours. This brightness difference is sufficient to segment the corresponding pattern by means of a radial histogram. In a radial histogram all coefficients of |G(m, n)| with the same distance rh to the centre are added. The corresponding radial histogram h for |G(m, n)| contains exactly K entries: s 2 2 M −1 N −1 + , K∈Z. (6) K= 2 2 Each radial histogram entry is defined as: h(rh ) = the sum of the intensity values with the distance rh to the centre, s 2 2 N −1 M −1 rh = + n− , for all 0 ≤ rh ≤ K , rh ∈ Z . m− 2 2
(7)
In the histogram shown in Fig. 7 the local maxima represent the characteristic coefficients in the amplitude spectrum (e.g. r1 ). The maximum is defined in a local area which starts at the origin or after a previous local area and ends when the amount of h(rh ) is less than 35% (empirical value) of the current local maximum. In order to extract the characteristic
3 .104
2.5 .104
h(rh )
2 .104
1.5 .104
1 .104
0.5 .104
0
0
50
100
r1
150
r
200
250
K
h
Figure 7: Radial histogram of the spectrum in Fig. 6b. coefficients and associated pattern all non-essential coefficients in the Fourier domain G(m, n) are suppressed by means of a threshold operation:9 G(m, n), for r1 − B < rh < r1 + B GF (m, n) = (8) 0, otherwise , with B = 2.5 as empirically determined value that specifies the range of the threshold operation.
Optical Document Security, San Francisco, 10-12 February 2016
© Reconnaissance International 2016
Subsequently, the coefficients are transformed back with the inverse 2D DFT8 (cf. Fig. 8): M −1 N −1 X X um vn 1 gF (u, v) = √ GF (m, n) · ei2π M · ei2π N , u ∈ [0 . . . M − 1] , v ∈ [0 . . . N − 1]. M N m=0 n=0
(9)
The extracted pattern represents the characteristic frequency at the point r1 , of the cyan colour channel. In an enlarged detail the individual dots are seen, which are aligned in a defined screen angle. The other colour screens are extracted similarly.
(N-1)/2
(M-1)/2
m
rh
n
(a)
(b)
(c)
(d)
Figure 8: Filtered spectrum (a) and the corresponding structure in original size (b) and in an enlarged form (c). Superimposed colour screens are shown in (d). Next, the rosette pattern is examined. Thhe segmented colour screens have to be superimposed (cf. Fig. 8d). The method of extracting rosette pattern features and colour screens will described at follows: Feature Extraction. The rotation determination is based on a model-based feature extraction. This extraction consists of two main parts: the screen model (cf. Sec. 3.1), and a comparison of the image data with the model. The specified geometric property of a colour screen is a straight line. With the straight alignment of the colour screen the rotation angle θ is computed. To determine this unknown parameter from the segmented structure the Hough transform10 (HT) is used. This transform provides the possibility to locate any shape that is defined with parameters within a distribution of points.11, 12 Most of geometrical shapes such as lines, circles or ellipses are described using simple equations with only a few parameters.8 Since the geometrical shape of aligned screen dots is a line we consider ρ as a function of θ in order to determine parameter combinations for a given screen point (xm , ym ): ρ(θ) = xm · cos(θ) + ym · sin(θ).
(10)
Each screen point projects a curve in the parameter space H(ρ, θ). This procedure describes all possible parameter combinations of the straight line through a point (xm , ym ) and is referred as Hough Space.8 The transformation in this space is the Hough domain. A further screen point (xn , yn ) on the straight line generates another curve in the Hough space (cf. Fig. 9). y
ρ(θ) (xn , yn )
(x , y ) m
m
ρ
1
θ1
ρ
x
1
θ1
θ
Figure 9: Hough transform of two screen points. The intersection of the two curves shows the possible parameter combinations for a straight line passing through two points (xm , ym ) and (xn , yn ). By considering other points more curves will intersect at the same position. If all screen
Optical Document Security, San Francisco, 10-12 February 2016
© Reconnaissance International 2016
dots are mapped in the Hough space (cf. Fig. 10) the point with the most intersections describes the required line with the parameters ρ and θ.
-45°
45°
°
°
°
°
°
°
°
°
°
Figure 10: Hough transform of a screen (black). Maxima in the Hough space are highlighted in red. In the Hough space (cf. Fig. 10) the highlighted local maxima represent the parameter combinations for ρ and θ for a black screen. In order to locate the parameters several solutions are possible. As only the angle θ is required for the rotation detection, the simplest solution would be a horizontal image projection Ph .8 If the Hough space is assumed to be a greyscale image this projection represents a one-dimensional sum of each pixel value in the Hough space H(ρ, θ). It has the length Nh corresponding to the width of H(ρ, θ) and the length Mh corresponds to the Hough space height: Ph =
M h −1 X
H(ρ, θ0 ), 0 ≤ θ0 ≤ Nh − 1 and H(ρ, θ0 ) ∈ [0, 1] .
(11)
ρ=0
The maximum of Ph represents the main rotation θK of the examined screen. In this example (cf. Fig. 10) the angle θ is equal to −45◦ and 45◦ due to the symmetrical screen. For subsequent analysis it is irrelevant which angle is used since the relation to other angles remains equal. Screen angles of the other colours cyan θC , magenta θM , and yellow θY are calculated accordingly and are assigned to the vector a: θC θM a= (12) θY . θK The extraction of the offset rosettes is executed accordingly. To represent a circle respective a rosette in 2D three parameters (x, y, r) (cf. Eq. (3)) are required. Therefore, the parameter space of the HT requires three dimensions HC (x, y, r) to locate positions and radius of rosettes.8 These positions are detected at local maxima in the Hough space and assigned to the matric C: x0 x1 x2 · · · xj C = y0 y1 y2 · · · yj . r0 r1 r2 · · · rj The parameters—used as features—in a and C are applied for subsequent decision making.
(13)
Decision Making. The detected screen angles (a) and rosettes (C) are characteristics that are applied to determine rosette offset printing (cf. Sec. 3.1). For this purpose the difference between five screen angles aD is computed: θC θM θCM θC θY θCY θCK θ θ aD = − = (14) C K θM θY θMY θM θK θMK
Optical Document Security, San Francisco, 10-12 February 2016
© Reconnaissance International 2016
If one of the corresponding angles is equal to 15◦ , 30◦ , or 45◦ it is deduced that the investigated region of interest is related to commercial offset printing. Besides the screen angles, the offset rosette is computed. Therefore, the nearest neighbours for a point position are computed (x1 , y1 ) using the parameters in C. These are determined by the smallest distance dx between two points (x1 , y1 ) and (x2 , y2 ). The distance is computed by: p (15) dx = (xm − xn )2 + (ym − yn )2 . The positions (xm , ym ) and (xn , yn ) represent two random circle centres. This process is continued until the neighbours of all points are determined. Subsequently, it is determined whether the rosettes are homogeneously distributed in the ROI. In case that a significantly high number of equal distances is present, it is determined being a rosette pattern. Results of the feature extraction and the classification are discussed in the following section.
4. RESULTS On the basis of our approach two demonstrators have developed and tested. The first demonstrator has been realized on the basis of a Data Collection Unit (DCU)13 which consists of a standard PC, a Dalsa-Teledyne Line chip camera with a resolution of approx. 1,000 ppi (pixel per inch) and a drive controller. The second demonstrator is based on an Android14 platform and has been developed as an App for a cell phone. Furthermore, the OpenCV library was incorporated.15 As testing devices we applied a Sony Xperia Z (13 MP)16 and a Samsung Galaxy S5 (16MP).17 Due to the low image quality of the mobile image sensor (cf. Fig. 11) it is currently not possible to implement the algorithms on the Android App.
(a)
(b)
Figure 11: Comparison of the image quality of the DCU (a) and the mobile device (b). Fig. 11 shows a comparison of the result of a DCU image and an image of a low cost image sensor (mobile system). While on the left side individual printing screens are detected by the image processing algorithms, it is not possible to separate and analyse the colour screens on the right side. Therefore, only the algorithms for the rosette detection are implemented on the Android platforms. In the framework of our research we examined offset prints from sample print brochures supplied by commercial printing companies.18, 19 The sample prints were produced with a screen frequency from 40 l/cm up to 80 l/cm. Examples with various refinement options on printed media between 80 g/m2 and 400 g/m2 were used. Additionally, we randomly selected 100 sections from brochures and journals. Results of the examination are presented on two sample prints (cf. Fig. 12). Findings from these two results can be transferred to all other offset sample prints. The above mentioned prints were selected for several reasons: the first print sample has a low screen frequency of 40 l/cm; whereas the second print sample has a screen frequency of 80 l/cm. Furthermore, in both samples little and heavily printed areas are found. We examined three image segments of each sample. The selected segments have a size of 400 × 400 pixels. Colour Screen Rotation. The first examined characteristic is the colour screen orientation. Its property is the existence of independent colour screen angles in relation to each other. In Fig. 13 the results for the first print sample from Fig. 12a are shown. In the first segment, only the screens of the colours magenta and black are segmented correctly with corresponding angles of 15◦ and 45◦ . For the colour Cyan no screen angle could be determined. One reason for this may be found in improper preprocessing: Here individual screen points of one colour are not clearly separated from other screen points. Furthermore, it is shown that the colour screen from the second and the third segment has been successfully segmented. The highlighted red lines indicate the computed angles. These correspond to 15◦ , 0◦ , and 45◦ for the colours Magenta, Yellow, and Black in both cases. The test result for the second print sample (cf. Fig. 12b) is shown in Fig. 14. A
Optical Document Security, San Francisco, 10-12 February 2016
© Reconnaissance International 2016
(a)
(b)
Figure 12: Comparison of examined print samples.
(a)
(b)
(c)
(d)
(e)
Figure 13: Analysis of screen angle for the first print sample (a), split for the colours cyan (b), magenta (c), yellow (d), and black (e). Note: an empty field means, that the colour does not exist in the sample. closer examination shows that the Cyan screen has been segmented correctly with the corresponding screen angle of 75◦ . Beside that artefacts are shown also. The colour screens of the second and third segment are not clearly extracted. One reason for this is the too low image resolution of the used camera chip. The individual screen dots cannot clearly be separated in the preprocessing step. As a result wrong patterns are segmented which cannot be used for further examination. Offset Rosette Detection. The second examined characteristic is the detection of offset rosettes. Since the rosettes occur especially in homogeneous printed surfaces it is expected that in these areas the rosettes are detected. In Fig. 15 the results for the two samples are shown. The red highlighted circles represent rosettes with a valid distance to each other. As expected, only few rosettes are localized in the homogeneous regions in the first segment and in the third one. In contrast, significantly more rosettes are detected in the second segment. The examination of the second offset print sample shows a similar result. In the light-printed areas only few rosettes are detected while inside the heavy-printed areas many rosettes are located. As mentioned before, the rosette detection has been implemented also on a cell phone. Despite the poor image quality, the results remain the same as shown in Fig. 15. In summary, the investigation shows that the rosettes are recognized in all heavily printed areas up to a resolution of 80 l/cm. In Addition the detection of screen angles is heavily dependent on the resolution.
Optical Document Security, San Francisco, 10-12 February 2016
© Reconnaissance International 2016
nicht vorhanden
(a)
(b)
(c)
(d)
(e)
Figure 14: Failed analysis of screen angle for the second print sample (a), split for the colours cyan (b), magenta (c), yellow (d), and black (e). Note: an empty field means, that the colour does not exist in the sample.
(a)
(b)
(c)
Figure 15: Determined rosettes in the first sample (upper half) and the second sample (lower half). Split into segment one (a), segment two (b) and segment three (c).
5. CONCLUSION AND OUTLOOK 5.1 Conclusion We have presented an approach for detection of commercial offset printing in counterfeited banknotes. The approach is a complement to the intrinsic analysis of the detection of Intaglio printing. The so-called Sound-of-Intaglio method is based upon intrinsic features of Intaglio printing and is applied as an authentication method of almost all banknotes. Our approach is applied for the detection of offset counterfeited banknotes. For this purpose, the images are separated in the individual spectral colour components. On this basis the corresponding colour screens are separated. After the preprocessing the segmentation of the individual screens and the rosette structures takes place. In order to extract a pattern from the image
Optical Document Security, San Francisco, 10-12 February 2016
© Reconnaissance International 2016
the first step is to localize characteristic points in the frequency domain of the image. In the second step unimportant frequencies are suppressed and the remaining image frequencies are transformed back. Finally, a model-based feature extraction determines the screen angles and the individual offset rosettes. Based on this approach two demonstrators have been developed and tested. The first demonstrator has been designed on the basis of a Data Collection Unit (DCU). The second demonstrator is based on an Android platform and has been developed as an App for a smartphone. The evaluation of the algorithms show that the results are heavily dependent on the resolution of the used camera chip. A reliable detection of screen angles is only possible up to a screen frequency of 40 l/cm. The rosette detection is far less depending on the camera resolution. The rosettes are discovered in all print samples of up to 80 l/cm. Furthermore, the rosettes are also recognized in images with poor quality. However, a disadvantage is that in light printed areas no rosettes are detected.
5.2 Outlook In our future research we aim to combine different approaches in one demonstrator and applying the tests on real banknotes and counterfeits. As mentioned is Sec. 4 problems during the preprocessing are present, resulting in not clearly extracted patterns and wrongly computed screen angles. In order to prevent these flaws we aim to replacing the camera of the DCU. In case of a camera with a higher resolution (e.g. 2000 ppi) we assume to get better results especially for the screen angle computation.
ACKNOWLEDGMENTS This work was partly funded by KBA-NotaSys SA, Lausanne, Switzerland.
REFERENCES [1] Orell, “Orell F¨ussli Security Printing Ldt:.” www.ofs.ch/en/products/banknotes/ (as-of 20 Mai 2015). [2] Gillich, E. and Lohweg, V., “Banknote Authentication,” in [1. Jahreskolloquium “Bildverarbeitung in der Automation”, ISBN 978-3-9814062-0-7], Centrum Industrial IT, Lemgo (2010). [3] Lohweg, V., D¨orksen, H., Hoffmann, J. L., Hildebrand, R., Gillich, E., Schaede, J., and Hofmann, J., “Banknote authentication with mobile devices,” Media Watermarking, Security, and Forensics , 03–07 (2013). [4] National Research Council of the National Academies, “Is That Real? Identification and Assessment of the Counterfeiting Threat for U.S. Banknotes,” (2006). [5] Kipphan, H., [Handbook of Print Media: Technologies and Production Methods, ISBN: 978-3540673262], Springer Science & Business Media (2001). [6] Kang, H. R., [Computational color technology, ISBN: 978-0819461193], Spie Press Bellingham (2006). [7] Fairchild, M. D., [Color appearance models, ISBN: 978-0470012161], John Wiley & Sons (2005). [8] Burger, W. and Burge, M. J., [Principles of Digital Image Processing: Core Algorithms, ISBN: 978-1848001947], Springer (2009). [9] Burger, W. and Burge, M. J., [Principles of Digital Image Processing: Fundamental Techniques, ISBN: 9781848001909], Springer (2009). [10] Hough, P. V., “Method and means for recognizing complex patterns.” US Patent 3069654 A (1962). [11] Illingworth, J. and Kittler, J., “A survey of the Hough transform,” Computer vision, graphics, and image processing 44(1), 87–116 (1988). [12] Duda, R. O., Hart, P. E., Stork, D. G., et al., “Pattern classification,” International Journal of Computational Intelligence and Applications 1, 335–339 (2001). [13] Gillich, E. and D¨orksen, H. and Hofmann, J. and Schaede, J. and Lohweg, V., “Data Collection Unit - A Platform for Printing Process Authentication,” in [Optical Document Security - The Conference on Optical Security and Counterfeit Detection V], (2015). [14] Android, “Android development area.” www.developer.android.com/ (as-of 20 June 2015). [15] OpenCV, “Opencv - open source computer vision library.” www.opencv.org/ (as-of 20 June 2015). [16] Sony Mobile Communications Inc., “ Sony Xperia Z .” www.sonymobile.com/de/products/phones/xperia-z/ (as-of 20 June 2015). [17] Samsung, “ Samsung Galaxy S5 .” www.samsung.com/de/consumer/mobile-devices/smartphones/galaxy-s/SMG900FZWADBT (as-of 20 June 2015). [18] Saxoprint, “SAXOPRINT GmbH.” www.saxoprint.de/producktmuster/druckmuster (as-of 20 Mai 2015). [19] H¨auser KG, “Buch- und Offsetdruckerei H¨auser KG.” www.druckdiscount24.de/druckraster (as-of 20 Mai 2015).
Optical Document Security, San Francisco, 10-12 February 2016
© Reconnaissance International 2016