阅读说明：本技术 借助神经网络进行印刷品质分析 (Print quality analysis by means of neural networks ) 是由 S·内布 N·R·诺瑞克 N·马丁 于 2019-07-03 设计创作，主要内容包括：本发明涉及一种用于借助计算机(6)补偿印刷机(7)中印刷过程的局部对版错误和/或重影错误的方法,其中,至少一个图像传感器(5)检测、数字化特定的印刷图案(8)并将这些数字图像数据(9)传递至所述计算机(6),所述计算机(6)检查所述数字图像数据(9)的可能的局部对版错误和/或重影错误并引入可能有必要的措施用于补偿这些错误,该方法的特征在于,所述计算机(6)将所述数字图像数据(9)供应给神经网络(12),该神经网络(12)借助训练数据(10,10a,10b)这样地被学入,使得它从被供应的数字图像数据(9)求得相应的对版值和/或重影值。(The invention relates to a method for compensating local register errors and/or double image errors of a printing process in a printing press (7) by means of a computer (6), wherein at least one image sensor (5) detects, digitizes and transmits specific printed patterns (8) to the computer (6), the computer (6) checks the digital image data (9) for possible local registration errors and/or ghost errors and introduces measures which may be necessary for compensating these errors, the method is characterized in that the computer (6) supplies the digital image data (9) to a neural network (12), the neural network (12) is learned by means of training data (10,10a,10b) in such a way that it determines corresponding register values and/or ghost values from the supplied digital image data (9).)

1. A method for compensating local register errors and/or double image errors of a printing process in a printing press (7) by means of a computer (6),

wherein at least one image sensor (5) detects, digitizes and transmits digital image data (9) to a computer (6) which checks the digital image data (9) with respect to possible local register errors and/or ghost errors and introduces possibly necessary measures for compensating the errors,

it is characterized in that the preparation method is characterized in that,

the computer (6) supplies the digital image data (9) to a neural network (12) which has been learned by means of training data (10,10a,10b) in such a way that the neural network determines corresponding register values and/or ghost values on the basis of the supplied digital image data (9).

2. The method of claim 1, wherein the first and second light sources are selected from the group consisting of,

it is characterized in that the preparation method is characterized in that,

-applying a suitable register mark (8) for said determined printing pattern (8).

3. The method according to any one of the preceding claims,

it is characterized in that the preparation method is characterized in that,

the digital image data (9) is cleared by a computer (6) of disturbing effects, such as static lens errors and objective lens distortions of the at least one image sensor (5).

4. The method according to any one of the preceding claims,

it is characterized in that the preparation method is characterized in that,

the training data (10,10a,10b) are generated manually by the computer (6) for the learning of the neural network (12) by generating a test data set (10) in the form of a digitized printed pattern (9,10,10a,10b) with known register and/or ghost errors and using the test data set as input variables for the neural network (12).

5. The method of claim 4, wherein the first and second light sources are selected from the group consisting of,

it is characterized in that the preparation method is characterized in that,

for the purpose of learning the neural network (12), a digitized printed pattern (9,10,10a,10b) with known registration and/or ghost errors is first transmitted by a computer (6) to the neural network (12) with a high image resolution and then the image resolution is continuously reduced by means of downsampling until the image resolution of the at least one image sensor (5) is reached.

6. The method according to any one of the preceding claims,

it is characterized in that the preparation method is characterized in that,

as at least one image sensor (5), a camera system (5) of an image detection system (2) is used, which is installed in an inline manner in the printing press (4) downstream of the printing unit.

7. The method according to any one of the preceding claims,

it is characterized in that the preparation method is characterized in that,

the method is carried out in an offset printing press (4) or an inkjet printing press (4).

Technical Field

The invention relates to a method for compensating for local register and ghost errors of a printing process by using a neural network (neuronalen Netz).

The technical field to which the invention belongs is print quality analysis.

Background

Different print control elements are used in offset printing presses. Some of these print control elements are printed only on demand, others are printed on each sheet. These so-called register elements or SID elements are used when the machine performs proofing (Abdrucken). They are used to analyze and eliminate the register errors and ghosting phenomena of the printing press. Here, the term "plate" (Passer) means: the individual color separations of the process colors of the printing press are superimposed in a targeted and accurate manner. The error classes which can be identified by these elements are, for example: shorter longer prints, narrower wider prints, cadence errors (e.g., ghosting), etc.

The elements are arranged distributed over the sheets and are designed in such a way that a register sub-element is printed in each printing unit. The mentioned error classes can be quantified based on the relative positions of these sub-elements with respect to each other. The reading in of these elements is done manually, which requires moving a special high-resolution camera from element to element. These digital image data are subsequently further processed by means of image recognition methods by means of suitable software.

Furthermore, the printing press section has an inline camera system within the scope of an image detection system which performs image monitoring for quality analysis, but this type of image monitoring is currently only suitable for determining the registration offset. For register-incorrect characteristic values, similar arrangements are also conceivable in digital (or inkjet) printing and have also been used in part for experimental purposes.

Disclosure of Invention

The object of the present invention is therefore to provide a method which is more effective and better for detecting local register errors and double image errors in printing processes than the methods known hitherto from the prior art.

This object is achieved by a method for compensating local register errors and/or ghost errors of a printing process in a printing press by means of a computer, wherein at least one image sensor detects, digitizes a defined printing pattern and transmits the digital image data to the computer, which checks the digital image data for possible local register errors and/or ghost errors and introduces possible measures for compensating the errors, characterized in that the computer supplies the digital image data to a neural network, which is learned (einglernt) by means of training data in such a way that the neural network determines the corresponding register values and/or ghost values on the basis of the supplied digital image data. The use of a neural network is necessary because the resolution of the image sensor is significantly lower than the printing resolution of a high resolution hand-held camera. This means that: the printed sub-elements are present after digitization (or detection) by the image sensor with a significantly lower resolution than when measured with a hand-held camera. When a simple analysis is carried out with the aid of image detection systems and conventional image processing algorithms, information about the printed sub-elements, which is actually necessary for evaluating the plate, is lost on account of this lower resolution. In order to be able to carry out the evaluation even in this case by the monitoring system and thus to replace the manual register check, in the method according to the invention the image data digitized by the image detection system are instead supplied to a neural network in the form of a self-learning algorithm, which then evaluates these image data with respect to the characteristic values of the printing nozzles to be determined. In order for such a self-learning algorithm to be able to achieve the goal, in addition to the knowledge of the underlying digital image processing tools, the self-learning algorithm must be learned beforehand by means of training data in digital form. It makes no sense to use an unlearned neural network. Therefore, the neural network must be learned based on digital training data (e.g., proofreading sub-elements) having specific, specifically known characteristics. This can be achieved by supplying the neural network with sub-elements in digital form which contain completely defined values (in most cases defined error images in respect of erroneous register). The neural network then learns the training data for a long time until the neural network can determine the correct characteristic values. Only then is the neural network used for evaluating the actually detected and digitized sub-elements and for subsequent registration.

Advantageous and therefore preferred developments of the method result from the dependent claims and the description with the figures.

In this case, a preferred development of the method according to the invention is to use suitable register marks for these defined printing patterns. Of course, it makes the most sense to use register sub-elements in the form of register marks as the printed pattern.

In this case, a further preferred development of the method according to the invention provides that the digital image data are cleared from disturbing effects (such as static lens errors and objective distortions of the at least one image sensor) by a computer. In order for the neural network to be able to correctly evaluate the current digital image data as well, the above-mentioned interference effects must be removed from the digital image data. Otherwise, it may happen that the neural network finds false errors or fails to correctly find the actual errors that are overlapped by these interference effects. However, it is more difficult to accurately obtain the register characteristic value. Of course, it is also possible to train the neural network in such a way that it learns to ignore these interference effects. However, this makes the actual learning process longer and more complicated, so that it is preferable to remove the disturbing effect from the digital image data by a computer.

In this case, a further preferred refinement of the method according to the invention provides that those training data for the learning of the neural network are generated manually by a computer in that test data sets in the form of digitized printed patterns with known register and/or ghost errors are generated and used as input variables for the neural network. This presents a particular advantage of using neural networks. Instead of programming a computer and an image monitoring system controlled by the computer with the aid of various digital image processing operations, as in the prior art, in order to detect errors in the subelements and to evaluate them accordingly with regard to the registration, the method according to the invention trains an artificially generated, digitally existing neural network with manually introduced error registration subelements for registration evaluation. Since this entire process is performed automatically by a computer, it is possible to: the neural network is supplied with a large amount of artificial training data generated by a computer in a relatively short time, and then the neural network realizes learning in a short time. For example, the training data can be generated manually by a computer in such a way that the training data contains sub-elements with specific image elements which represent specific registration errors in the printed image. The computer can be programmed in such a way that the errors contained in the artificially generated training data can be changed in almost any combination. In this way, a large number of different training data sets can be generated in the shortest time and used for learning the neural network.

In this case, a further preferred development of the method according to the invention is that, for the learning of the neural network, those digitized printed patterns with known registration errors and/or ghost errors are first transmitted by the computer to the neural network with a high image resolution, and then the image resolution is continuously reduced by means of down-sampling (downscaling) until the image resolution of the at least one image sensor is reached. In order to enable a neural network to correctly learn digitized printed image data with known errors, the invention proposes that known errors be introduced into digitized printed image data with a higher image resolution, and that the neural network be first trained with the higher image resolution and then reduce the image resolution step by step until the actual image resolution of the image sensor is reached. This makes it easier for the neural network to learn in and find known errors introduced.

In this case, a further preferred development of the method according to the invention provides that, as at least one image sensor, a camera system of an image detection system is used, which is installed in an inline manner in the printing press downstream of the printing unit. The invention proposes that most of the image detection systems which are present in inkjet printers, are used for image monitoring and are usually installed in the printer downstream of the printing unit in an inline manner, and which are intended to check the quality of the print obtained, also be used for register monitoring. Here, a camera of an image detection system is used as an image sensor for detecting and digitizing the printed test pattern. Of course, external image sensors or those internal image sensors which are not part of the image detection system can also be used for the method according to the invention. The method is necessary in particular in printing presses which do not have an inline image detection system. However, in principle, for efficiency reasons, using existing image detection systems with their cameras is the most significant and efficient way.

In this case, a further preferred development of the method according to the invention provides that the method is carried out in an offset printing press or an inkjet printing press. Although the method according to the invention is developed primarily for register monitoring in offset printing presses, the method according to the invention can also be used in inkjet printing presses. In any case, correct register is also a fundamental prerequisite for high-quality multicolor printing for inkjet printing.

Drawings

The invention and its structurally and/or functionally advantageous refinements are further described below on the basis of at least one preferred embodiment with reference to the drawings. In the drawings, elements corresponding to each other are denoted by the same reference numerals, respectively.

FIG. 1: examples of image monitoring systems in sheet offset printing presses;

FIG. 2: the detected digitized test pattern, existing at the actual print resolution and the detected camera resolution, respectively; and

FIG. 3: schematic flow of learning into neural networks for image analysis processing.

Detailed Description

Fig. 1 shows an example of an image detection system 2 employing a method according to the invention. The image detection system 2 comprises at least one image sensor 5 (typically a camera 5 integrated into the sheet-fed printing press 4). The at least one camera 5 captures the print image generated by the printing press 4 and transmits the data to the computers 3,6 for evaluation. The computers 3,6 can be separate computers 6 which are inherent to themselves (for example one or more specialized image processing computers 6) or can also correspond to the control computer 3 of the printing press 4. At least the control computer 3 of the printing press 4 has a display 7, on which display 7 the result of the image monitoring is displayed to the user 1.

The aim of the method according to the invention is to train the neural network 12 in such a way that it can determine the register values with sufficient accuracy and rapidity on the basis of the relatively low-resolution images 9 of the inline camera system 5. The input variables of the method proposed by the invention are a digital camera image 9 of a suitable register sub-element 8 (preferably a register mark 9 in digital form), which digital camera image 9 can be corrected for static lens errors, objective distortions, etc., if necessary. In principle, other sub-elements may also be used. What is important is only that: this sub-element has sufficient information about the aspect of the register. No further preparation/processing steps in terms of signal technology take place here. The output parameter is the corresponding register value. The association between input and output is not inherently "learned" by the interlaced image processing algorithms (versachtelebeldverarbeitungsalgorithmen) as in the prior art, but by the neural network 12 in dependence on the so-called training data 10,10a,10 b. The emphasis is that the training data 10,10a,10b of the neural network 12 can be generated manually. Since these register elements 8 can now perform inline detection, online print quality analysis can be performed. Based on these register data, information can be continuously obtained which can ultimately also be used for continuously observing and classifying the printing quality of the printing press 4 with regard to the aspects of condition monitoring and predictive maintenance. To date, these systems have only been able to employ indirectly correlated data (such as x-rotary encoder data). The present invention, however, provides a direct correlation between the printed image 11 on the sheet and the characteristic value.

The following explains the sequence of the preferred embodiment variant in which only the subelements 9 are present. In the production case, possible register deviations are determined and corrected on the basis of the camera image 9 thus present, by means of the digital image data 9 being guided through a trained neural network 12. The training of the neural network 12 is further illustrated in fig. 2 and 3 and proceeds as follows:

presence of high resolution register marks 8.

If it is not present at a sufficiently high resolution, the resolution of the register marking 8 is artificially increased by tens or even hundreds of times → so-called Upsampling (Upsampling).

○ in this high resolution, register shifts are predefined in the circumferential direction, the transverse direction and introduced into the register marks 8;

○ if an inkjet printer 4 is involved, inaccuracies caused by the printing method (such as distortion, fraying, etc.) may additionally occur → these then also being introduced into the register marks 8;

the resolution of the high-resolution image 8 is then artificially reduced to the resolution of the camera 5 by means of so-called down-sampling.

Fig. 2 shows an example of a test pattern on the left side, on the one hand, in which case this should present the registration marks in digital form at a high starting image resolution of 8(2540dpi), and on the right side the same test pattern at a lower camera resolution of 9(200dpi) on the one hand.

In this way a manual test data set 10 with image data and position information (including known errors) is generated as input data. Many such test data sets 10 are required for training the neural network 12. Fig. 3 schematically shows the flow of this training. It is well seen here how the neural network 12 is trained in multiple stages. The data record 10 is divided into a training data record 10a and a test data record 10b, for example in the scale 60/40. The network 12 is "trained" on the basis of such artificially generated training data sets 10a, while the test data sets 10b examine the trained network 12. That is, the network 12 is first trained by means of training data 10 a. If a sufficient level is then reached, it is verified by means of the test data 10 b. This validated network 12 is then used for the actual image data 11, which actual image data 11 is present as tested image data 11a with position information after the examination by the validated neural network 12. Since these training data 10,10a,10b are generated manually in the computers 3,6, the use of many different test data sets 10 is not a problem. The more test data sets 10, the better the neural network 12 learns. Furthermore, a great advantage of generating these training data 10,10a,10b on the computer 3,6 is that the errors introduced are made known. The result is the register value of the printing mechanism under consideration. As the computers 3,6, an image processing computer 6 of the image detection system 2 is preferably used. However, in some cases, the control computer 3 of the printing press 4 may alternatively be used.

In another embodiment, a modified imposition sub-element may also be used to derive the same output data.

The method according to the invention using a neural network 12 has many advantages over the prior art with strict image processing algorithms. Thereby eliminating the need to take and manipulate multiple parameters. The success of the approach of the method and its robustness do not depend on partly randomly selected parameters but are derived on the basis of the quantity and quality of training data 10,10a,10b that can be generated manually and thus are almost limitlessly available. Therefore, almost any number of training data 10a and test data 10b can be generated. Furthermore, since it is generated manually, the errors and the authenticity are known with arbitrary accuracy. Since these register elements 8 can now perform inline detection, online print quality analysis can be performed. Information is continuously available on the basis of these register data, which information can ultimately also be used for continuous observation and classification of the printing quality of the printing press 4 with regard to the aspects of condition monitoring and predictive maintenance.

The computer power required in training the neural network 12 is in most cases significantly lower than in most cases the case of complex image processing algorithms, as it is currently used. By manually reading in these printing elements 8, the printing can be limited to thirty sheets in the case of a print, which limitation can likewise be dispensed with. The printing monitoring can then be carried out on the basis of, for example, five hundred sheets. This greatly increases the persuasiveness and accuracy of the data and saves cost and time when finally installing the printing press 4.

List of reference numerals

1 user

2 image detection system

3 control computer

4 printing machine

5 image sensor

6 image processing computer

7 display

8 register element with high original image resolution

9 lower camera resolution register elements

10 artificially generated image data with position information

10a artificially generated training data

10b artificially generated test data

11 actual image data

11a actual image data with position information

12 neural network