High Quality Offset Printing An Evolutionary Aroach Ralf Joost Institute of Alied Microelectronics and omuter Engineering University of Rostock Rostock, 18051, Germany +49 381 498 7272 ralf.joost@uni-rostock.de Ralf Salomon Institute of Alied Microelectronics and omuter Engineering University of Rostock Rostock, 18051, Germany +49 381 498 7251 ralf.salomon@uni-rostock.de ABSTRAT Print media are still very imortant for everyone's daily life. urrent efforts are concerned with the alication of the wellestablished offset-rinting technology to other media, articularly cardboards, which require some substantial adatations. To this end, this aer roosed a new secific re-rocessing stage. This re-rocessing stage can be configured by several arameters. This aer otimizes these arameter settings by using evolution strategies. It turns out that this otimization reduces the required energy and the number of wrongly generated ixels by about 15%, resectively. ategories and Subject Descritors J.7 [omuter Alications]: omuter in other Systems Industrial control, Process control. General Terms Algorithms, Performance, Design, Economics, Reliability Keywords Image rocessing, evolutionary algorithms, industrial alication 1. INTRODUTION Print media, such as newsaers, journals, books, and flyers, are still imortant information sources in everyday life. These media are the result of a well-established rinting rocess in which a rinting machine uses a rinting late to distribute liquid color onto the actual media, i.e., aer. In this rocess, the rinting content, i.e., letters, symbols, and grahics, is encoded on the rinting late's surface in form of a hysical structure. This structure consists of reverses and dots, with the latter being the means to bring the color onto the aer. This rocess is widely known as offset rinting, and sketched in Figure 1. Every age requires a age-secific rinting late, which is generally reared in three stes. First, a steel or aluminum late, covered with some light-sensitive olymeric material, is brought into another machine, called a setter, which is deicted in Figure 2. Then, the setter's exosing unit transfers the age's content onto the rinting late's surface. This content is normally stored as a digital image, which is a collection of individual ixels, on a (host) comuter. During the exosure, the olymeric material changes its hysical roerties: those areas that are exosed to light hardens, the others not. Finally, a subsequent chemical rocess removes the nothardened material such that later on, the exosed areas are able to transfer the liquid color. It might be mentioned here that some setters utilize laser light, others ultraviolet light with a wave length of about 360nm to 450nm. Due to the rinting late's size and the required resolution, the image cannot be transferred in a monolithic form. Rather, as is illustrated in Figure 3, the image is divided into stries. In stateof-the-art setters, an entire strie is rocessed by moving the exosure head from left to right or vice versa. During these movements, the image data is loaded and aroriately shifted line by line through the head's otical system (see, also, Aendix A). For regular rint media, such as aer, the offset rinting rocess is well established. Since offset rinting is of high quality and cost effective, current efforts try to also utilize it for other material, such as cardboard, wall aers, and the like. These new materials, however, are characterized by a much higher roughness, which in turn requires substantial modifications of the rinting late, called flexo rinting late: In order to ensure highquality rints, the difference between the rinting late's dots and reverses has to be much more ronounced. This in turn imoses new roblems to the exosing rocess: the main roblems are, as Section 2 discusses, that the rinting late s dots lack hysical stability and that the image runs out of focus due to the olymeric material's thickness. The exosing rocess is defined by several arameters, including the light's intensity, the exosure time, and the oint of focus. rinting late dot reverse Permission to make digital or hard coies of all or art of this work for ersonal or classroom use is granted without fee rovided that coies are not made or distributed for rofit or commercial advantage and that coies bear this notice and the full citation on the first age. To coy otherwise, or reublish, to ost on servers or to redistribute to lists, requires rior secific ermission and/or a fee. GEO 07, July 7 11, 2007, London, England, United Kingdom. oyright 2007 AM 978-1-59593-697-4/07/0007...$5.00. ink disenser rint material (aer, cardboard, etc.) Figure 1: Simlified rint rocess 2053
exosure head bearing rinting late late holder terminal Figure 2: The UV-Setter, courtesy of Punch Grahix Preress Germany, Ltd. In order to alleviate the roblems mentioned above, Section 3 rooses an image rerocessing method that rocesses images the same way the visual system of mammals does. This neuralnetwork-based aroach is defined by a set of arameters, called connections. With resect to the rinting's quality, these arameters are highly deendent on each other, and multi modal by their very nature. Therefore, Section 4 alies evolution strategies to evolve them. The results, as resented in Section 5, indicate that the roosed aroach leads a homogenous distribution of the light's energy, and therefore, a high rinting quality. Furthermore, otimally evolved arameter sets may reduce the exosing time by about 15%. Section 6 concludes this aer with a brief discussion. 2. PROBLEM DESRIPTION As has been mentioned in the introduction, flexo rinting lates are characterized by a significantly increased thickness of the light-sensitive olymeric material. This thickness means that the exosing light has to travel a longer distance through this material. Along this distance, the light diverges, since it has reviously assed a lens system (see, also, Aendix A). The roblem is that flexo lates cannot neglect this light divergence; rather, they have to deal with and comensate for it. Because the light diverges along its way through the olymeric material, a blob has varying diameters along the z-axis. This effect is illustrated in Figure 4, which shows four cross-sections along the z-axis. The increased size of the exosed area leads to two different effects. First: Due to the increased size of a blob in the bottom layers of the late, the energy er unit area decreases. In order to have enough energy to roerly harden the olymeric material, the total amount of light energy has to be increased. Due to mechanical constraints, this can be achieved only by slowing down the exosure rocess. This comensation, however, contradicts economical demands, which desire the rocess to be as fast as ossible. Second: an increased amount of global light energy also widens a ixels size at the to layer (light ixels are circles, rinting ixels are squares). This degrades the rinting quality significantly, because surious ixels may aear on the rinting material, In this aer, the term surious ixel refers to ixels that aear on the rint media, even though they are not in the data file. In summary, in order for single (isolated) ixels to roerly aear on the rinting late, the light energy has to be increased, which leads to other roblematic effects, e.g., surious ixels. In contrast to very thin rinting lates for regular aer material, flexo lates suffer from these effects and thus have to comensate for them. 3. THE PROPOSED APPROAH Section 2 has discussed that flexo rinting lates have to coe with certain exosure roblems. From a technical oint of view, the rearation of a rinting late can be seen as an image rocessing task in which a digital image is brought onto a hysical late. surface level moving directions image rocessing system exosed area unexosed area rintig late sensitive to UV-light Figure 3: Illustration of an exosing rocess base level Figure 4: Widening of a single, freestanding ixel (blue border) 2054
Neuronal Network central ixel energy values original image different energy rocessing values for each ixel (all ixels rocessed) Figure 5: the neural network aroach The aroach resented in this aer inserts an image rerocessing stage that is executed during or rior to the exosure of the late (see, also, Section 6). The design of this rerocessing stage draws some insiration from how the mammalian visual system rocesses images: for every ixel, the brain alies local filters all of them having the same structure and rocessing the ixel's neighborhood [7]. This visual system oerates highly arallel, fault tolerant and reliable, and adative by its very nature. Thus, the aroached roosed here is called neurally rerocessed rinting late or NP³ for short throughout this aer. Figure 5 illustrates that like the biological visual system, NP³ considers a virtual neuron for every ixel. This neuron senses its corresonding image data, called the central ixel, as well as the ixel's neighborhood, which might be of 3x3 or 5x5 in size. This neuron calculates the ixel's exosure I from the image data P as follows: I = w (1) P with w denoting a connection weight or strength. Since all ixels are handled in the same way, all neurons and thus all connections w are identical. The roosed aroach seems rather simle and straightforward. However, the connections w are not indeendent of each other. Because of the light's divergence, every ixel also influences its neighbors by injecting some light energy to them. This in turn affects the determination of otimal values for the connections w. Both the strength and the number of influenced ixels deend on the chosen light geometry as is illustrated in Figure 6. 4. METHODS AND SIMULATION This section describes the evolutionary aroach, defines the fitness function, and sketches the simulation environment used for the exeriments. The evolutionary aroach: Because the arameters w are real-valued, exhibit significant eistatic interaction, and are multi-modal with resect to the fitness function 1, this aer uses evolution strategies [4] for the otimization of the connections w. 1 These roerties have been observed by reliminary exeriments that have involved classical steeest-descent methods. neighbour ixel energy values satial energy distribution Figure 6: Exosure settings Evolution strategies in general are a member of the class of heuristic oulation-based search rocedures known as evolutionary algorithms that incororate random variation and selection. Evolutionary algorithms rovide a framework that mainly consists of genetic algorithms [1], evolutionary rogramming [2], [3], and evolution strategies [5]. A genetic algorithm maintains a oulation of µ individuals, also called arents. In each generation, it generates λ offsring by coying randomly selected arents and alying variation oerators, such as mutation and recombination. It then assigns a fitness value (defined by a fitness or objective function) to each offsring. Deending on their fitness, each offsring is given a secific survival robability. Unless otherwise stated, this aer has used a (1+5)-evolution strategy, since the ertinent literature [4] indicates that this configuration of the number of arents and offsring yields the highest sequential fitness. The chosen notation also indicates that the arent of the next generation is selected from union of the current offsring and the revious arents. In addition, the chosen evolution strategies have also emloyed a ste size adatation. Furthermore, this aer considers rather small oulation sizes, since the ractical fitness evaluation is a very time-consuming rocess. It should be mentioned here that the evolutionary algorithm initializes its oulation members with w (i!=4) = 0 and w 4 = 8.5, which reresents the conditions without the rerocessing stage. The fitness function: The goal of the evolutionary rocess is to evolve a weight matrix, such that all ixels aear as given in the digital data file: that is, for black ixels the total light energy has to be above a threshold T, such that it withstands the chemical rocess; and obviously for white ixels, the total light energy has to be below this threshold. Mathematically, the fitness F for a single ixel is calculated as follows: 2 ( T E ) E > T; = 1 3 5( E T ) + 100 E < T; = 1 = 3 5( E T ) + 100 E > T; = 0 2 ( E) E < T; = 0 F (2) 2055
with T denoting the threshold value, E denoting the energy value of Pixel and denoting the current state of ixel in that way that 1 constitutes a black ixel and 0 constitutes a white ixel, resectively. The offset 100 has been introduced in order to ronounce correctly exosed ixels. For illustration uroses, Figure 7 deicts the two cases, i.e., black and white ixels, of F. Then, the fitness function F sums the individual contributions F overall ixels P as follows: F = P F = 1. (3) Simulation environment: Because the evaluation of every fitness value by means of rearing a hysical rinting late is way too costly and time consuming, this aer has used a hysically lausible simulation software. This simulation software allows for the rocessing of arbitrarily configured test images, rovides the user with an aroriate grahical user interface, and suorts the setting off all relevant arameters, such as oulation sizes, lighting conditions, late arameters, etc. For the exeriments reorted in this aer, the simulation software has used the test image shown in Figure 8. In order to yield a recise evaluation, the simulation subdivides every ixel into 3x3 subixels. This way, the simulation software accounts for the fact that a light ixel is of circular geometry, whereas the image data contains square-shaed ixels. An examle is illustrated in Figure 6. The Figure also shows that the simulator treats the light beams as fine-grained normally distributed values. It can be seen that the Gaussian-shaed light beams account for the effect of overlaing ixels in a natural way. 5. RESULTS In order to asses the quality of the NP³ aroach roosed in this aer, this section also considers rinting lates that are exosed in the usual way, that is without any further rerocessing. 1800 1600 1400 Fitness function Figure 8: Test image For the chosen test image, Figure 9 shows how the fitness evolves over time. It can be seen that in comarison to the standard exosure rocess, the evolution strategies reduce the fitness values (actually, error values) by about 60%. Figure 11 shows how the imroved fitness affects the rinting late. It can be seen that in comarison to the standard exosure rocedure, shown in Figure 12, the evolved rerocessing stage has significantly reduced number of surious ixels. For resentation uroses, Figure 10 shows two corresonding blow us. Already in this very small subfigure, the evolutionary algorithm has reduced the number of wrongly exosed ixels (red) from 32 to three. In addition, the rerocessing stage significantly reduces the total amount of the light energy by about 12%. Furthermore, the energy is much more homogeneously distributed, which has additional advantages for the hoto/chemical rocess, which is, however, beyond the scoe of this aer. It might be argued that other configurations than the chosen (1+5)-evolution strategy might yield better or worse erformance in terms of the exerimentation time. However, this question is not in the focus of this aer; rather, the goal is to achieve suerior solutions by alying evolutionary algorithms, because standard aroaches yielded solutions with limited utility. Furthermore, it might be mentioned again that with resect to hysical validations, the oulation sizes have to be small. 100000 80000 fitness value over 150 generations Fitness value 1200 1000 800 600 =0 =1 fitness value 60000 40000 400 20000 200 0 0 2 4 6 8 10 12 14 Energy value 0 0 25 50 75 100 125 150 generation Figure 7: Fitness function Figure 9: Fitness rogress 2056
Figure 10: Exosed lates (detail view) 6. DISUSSION This aer has resented an alication from the area of offset rinting. The core of this alication is a machine that transfers a digital image onto a rinting late, which is being finished in a subsequent chemical rocess. Due to economical reasons, current develoments try to also utilize offset rinting to other media, such as cardboards. These media, however, are characterized by other hysical roerties, and thus require a modified exosure rocess. Since the direct usage of the image data leads to the generation of ixels that are actually not in the image data (surious ixels), this aer has roosed a rerocessing ste, called neurally rerocessed rinting late, that yields significant imrovements over state-of-the-art, rather standard exosing rocedures. These achievements are significant from an economical as well as quality-of-rinting oint of view. urrent develoments are integrating these otimizations into available market roducts. In the current aroach, the otimal values for the (virtual) connection weights w deend on the actual image data. This will require an additional, very short adatation hase for every new image. Even though this adatation can be executed offline, i.e., on a searate comuter, it can be considered a flaw of the current aroach. Therefore, further research will be dedicated to the develoment of an enhanced rerocessing stage in which the otimal arameter settings do not deend on the actual image content. Figure 11: NP³ Figure 12: lassical exosed rinting late Furthermore, future research will be develoing a custom-made hardware design, which will be based on state-of-the-art fieldrogrammable gate arrays (FPGAs) [6]. The exected advantage will be that both the adatation and the actual rerocessing can be done on the fly during the rearation of the rinting late. 7. AKNOWLEDGMENTS The authors gratefully thank Dr. Horst Steat, managing director at Punch Grahix Preress Germany Ltd. and John Hedde, director of research and develoment at Punch Grahix Preress Germany Ltd., for roviding the existing technology and the excellent cooeration. This research was suorted in art by the German federal ministry of education and research, grant number 03i4919B. 8. REFERENES [1] D.B. Fogel. Evolutionary omutation: Toward a New Philosohy of Machine Learning Intelligence. IEEE Press, NJ, 1995. [2] L.J. Fogel, Autonomous Automata. Industrial Research, 4:14-19, 1962. [3] D.E. Goldberg. Genetic Algorithms in Search, Otimization and Machine Learning. Addison-Wesley, Reading, MA, 1989. [4] I. Rechenberg, Evolutionsstrategie. Frommann-Holzboog, Stuttgart, 1994. [5] H.-P. Schwefel. Evolution and Otimum Seeking. John Wiley and Sons, NY. 1995. [6] R. Joost, R. Salomon. Hardware-Software o-design in Practice: A ase Study in Image Processing, In Proceedings of the 32 nd Annual onference of the IEEE Industrial Electronics Society (IEON), Paris, France, Nov. 2006. [7] D. Hubel. Eye, Brain, and Vision (Scientific American Library, No 22), W. H. Freeman, 1995. 2057
9. APPENDIX A Figure 13 shows the lens system of the UV setter used in this aer. A key feature of this articular lens system is the digital mirror device (DMD 2 ), which emloys 1024*768 individually controllable, tiny mirrors. An attached hardware controller switches every single mirror on and off according to the resent image data. The DMD is thus able to individually control the exosure of every single ixel. Figure 13: Otical system of the UV-Setter, courtesy of Punch Grahix Preress Germany, Ltd. 2 The DMD is manufactured by Texas Instruments Incororated 2058