METHODS FOR TESTING ROTOGRAVING CYLINDERS AND ROTOGRAVING PLATES

Description

[0001] The present invention relates to a method for detecting defects on gravure cylinders or gravure plates. In particular, the method utilizes neural networks.

[0002] In the following, we will refer to gravure cylinders. All process steps (with the exception of rotation) are analogously applicable to gravure plates or are explained separately.

[0003] In gravure printing, inks are applied to depressions in the surface of metal cylinders or plates and then transferred to the medium to be printed (paper, fabric, etc.). The depressions are formed as a multitude of individual cavities (called cells or cups).

[0004] An imaged (engraved) surface is created by indentations being created into a cylinder surface (gravure cylinder) using various processes. The main imaging processes in gravure printing are electromechanical engraving with a diamond stylus, laser light, or a chemical etching process.

[0005] Each imaging process creates typical cup shapes according to its technical procedure, with the arrangement of these cup shapes on the cylinder surface being subject to the manufacturing constraints of the process. Typically, the cups are designed as depressions with a diamond-shaped opening. The diamonds are arranged in parallel columns, with adjacent columns offset by half the distance of the diamond centers of a column, allowing the diamond columns to be arranged in an interlocking manner.

[0006] Typical designs are closed cells, in which the corners of neighboring cells in a column are separated from each other, meaning each cell exists independently and enclosed within the surface, and open cells, in which the corners of neighboring cells in a column are connected. The connection between the corners is realized either by direct contact or by means of a short channel that bridges the distance between the corners.

[0007] During the engraving process, surface damage may already be present on the gravure cylinder, or surface damage or defective cells may occur during the engraving process. Detecting and categorizing these defects is a recurring task in quality assurance for gravure cylinders.

[0008] The subject matter of DE 10 2019 117 680 A1 is a method for measuring defects in components in a quality inspection process. The method uses a captured digital image of a component to be examined and a first reference image of the component to be examined. A second reference image is created from a combination of the captured digital image and the first reference image. In addition, the method uses a trained machine learning-based classifier system that was trained with training data to create a model, whereby the model serves as the basis for a semantic segmentation of voxels of the received image into defect classes. The activated classifier system classifies voxels of the received digital image, whereby both voxels of the received digital image and voxels of the second reference image serve as input data for the classifier system. The use of an SVM is explicitly mentioned.The presence of a flawless reference image is essential for this method. The overall structure of the acquired image is compared with the overall structure of the reference image. The individual elements present in the acquired image (structures consisting of multiple pixels / voxels) are not detected or analyzed.

[0009] Known procedures involve carrying out a test print followed by checking the print result.

[0010] DE 10 2008 016 538 A1 proposes, in order to assess print results, to identify text and image areas on the printed substrate and to capture the image areas with a multispectral camera and then, after image processing, to compare the obtained values ​​with reference values.

[0011] EP 0 749 833 A2 attempts to solve the problem of inconsistent brightness when capturing images of printed materials. The proposed system is intended to ensure spatial and temporal consistency of the measurements.

[0012] The subject of EP 0 795 400 A1 is a method for aligning captured images for image comparison. For this purpose, a reference image is captured and converted into grayscale values. The image to be evaluated is captured analogously. From the comparison of the two images, a transfer function is derived that assigns the image areas to each other.

[0013] GB 2 159 622 A describes in detail the algorithms used when comparing a reference image with a comparison image. The printing speed at which the reference and comparison images were obtained is taken into account in the calculation.

[0014] JP 2012154830 is aimed at detecting defects in the printed product caused by the gravure printing plate. For this purpose, the reference is not an initial image created using the printing plate, but rather an image created previously, independently of the printing plate.

[0015] US 2007 / 0201066 A1 describes analysis methods for quality control images taken from printed material.

[0016] US 2019 / 0340740 A1 proposes attributing human assessment of print quality to individual parameters of the print image, thus creating a set of parameters for determining print quality. Images of printed substrates captured by a camera are then to be downscaled to the specified parameters and categorized.

[0017] The disadvantage of print inspection methods is the considerable effort involved. The gravure cylinders must be inserted into a printing press for testing and then removed and cleaned after the test print.

[0018] Advanced methods involve direct inspection of the surface of the gravure cylinders.

[0019] US 2017 / 0246853 A1 describes: "A printing apparatus applies ink to a surface of a printing plate in the form of a predetermined pattern and then transfers the ink to a substrate. The printing apparatus includes an image recording section that applies the ink to the surface of the printing plate; a plate surface observation unit that acquires information about the surface of the printing plate; a storage section that stores information about a reference shape serving as a reference of the surface of the printing plate; and a determination section that compares the reference shape information stored in the storage section with the information about the surface of the printing plate obtained by the plate surface observation unit and determines whether the surface of the printing plate to which the ink has been applied lies within a predetermined range of the reference shape."

[0020] In some embodiments, the removal of defects (contaminants) on the printing plate by means of a laser is also provided.

[0021] In CN 108481891A, the printing rolls are placed in an inspection frame and scanned with a camera system. The camera system and the downstream data processing unit create a three-dimensional image of each cell (cup) of the roll. The cell volume is then calculated and compared with a specified value. EP 0 636 475 B1 provides for the comparison of an image of the surface of a printing plate or printing cylinder with either a reference image or a set of rules describing the positive / negative state of sections of the printing plate / cylinder. Some of the mathematical methods used are listed.

[0022] DE 10 2005 002 351 A1 proposes a method for determining surface topology under pressure. A surface profile is obtained using light scanning. Section

[0067] briefly outlines the calculation method.

[0023] JP 2002254590 describes a frame for inspecting a gravure cylinder (printing roller). The cylinder is rotatably mounted in the frame, and a camera image of its surface is generated. Image recognition software is designed to detect defects.

[0024] JP 2018039117A relates to a device for examining the surface condition of a gravure printing plate. Anomalies in the plate depth are detected through continuous monitoring. An automatic cleaning machine can be activated as needed. Here, too, a three-dimensional image of the surface is obtained by irradiating it with laser light, and the reflected light is captured by a camera. An automated evaluation then takes place.

[0025] WO 2008 / 049510 A1 describes a method and device for inspecting printing cylinders. The device comprises several measuring devices that mechanically and optically scan the cylinder surface and generate an image of it. The method calculates a proof image without actually printing.

[0026] US 2007 / 089625 A1 relates to a method in which defects in both the cells and the surface of a printing cylinder are detected using image analysis.

[0027] In general, the image processing methods used are not described in detail.

[0028] Methods and devices that incorporate printing press parameters into defect analysis are also proposed: JP 2014106069A uses images of both the print result and the printing plate to generate defect detection from the combined images. The image processing methods are described only generally: "The inspection processing is alignment processing by pattern matching, normalized correlation, etc., filtering processing such as averaging and maximization, centering on difference processing between the reference image obtained at the time of conveyance and the inspection image to be inspected."

[0029] Currently, there are first attempts to use methods of artificial intelligence (AI) and neural

[0030] Networks (NN) are to be included in the testing of gravure rollers: CN 106739485A stipulates that a number of parameters of the printing cylinder are recorded together with color signals from the rotating printing cylinder. The data is fed into an algorithm that determines the overprint error. Correction of the result using a neural network is provided.

[0031] JP 2019117105A proposes the use of a neural network to categorize print image defects. The defects are to be divided into two categories (harmless and non-harmless).

[0032] Although numerous solutions have already been proposed, the problem of inspecting gravure cylinders and plates cannot be considered satisfactorily solved. This is particularly true given the increased demands on the reliability and speed of the inspection procedures.

[0033] The task therefore remains to propose a method for testing the surface of gravure cylinders or plates that ensures rapid and reliable detection of defects.

[0034] According to the invention, the object is achieved by the method according to claim 1. Advantageous embodiments of the method are disclosed in the dependent subclaims.

[0035] The method according to the invention for optically testing the surfaces of gravure cylinders and gravure plates provides at least the following steps: a) Place the gravure cylinder or gravure plate into an inspection device, b) Scanning optically capture the surface of the gravure cylinder or gravure plate using an image capture unit in one or more measuring images and capture the position of the measuring images relative to the surface of the gravure cylinder or plate.of the gravure printing plate, c) detection of cells by scanning the first measured image, d) comparison of a series of specified properties with the specified target values ​​assigned to these properties and recording the deviation of each property from the respective target value as a value in a feature vector, e) detection of defects in the areas of the first measured image in which no cells were detected, f) determination of an anomaly defect value of a detected defect, g) classification of the feature vector of each cell and the anomaly defect value of each defect based on specified result weighting and permissible tolerance values ​​of each property of the feature vector by means of a pre-trained support vector machine (SVM) and, h) repetition of steps c) to g) for each further measured image, i) output of an evaluation of the gravure cylinder or the gravure printing plate, wherein the evaluation includes at least the information "fault-free" or "faulty".

[0036] Steps b) to i) are preferably carried out with the support of or in one or more data processing systems. Place the gravure cylinder or gravure plate into an inspection device

[0037] The inspection device provides a rotatable bearing for the gravure cylinder or a flat holder for gravure plates. An inspection unit scans the surface of the gravure cylinder. For this purpose, the inspection unit preferably has at least one image capture device (e.g., camera) and at least one illumination device. Suitable beam deflection devices (mirrors, prisms) ensure that the light from the illumination device falls perpendicularly or nearly perpendicularly onto the surface of the gravure cylinder and that the image capture device can capture a measurement image perpendicularly or nearly perpendicularly to the surface. The perpendicular illumination and image capture largely prevent shadow effects on the edges of cells or damage.

[0038] By moving the inspection unit relative to the surface of the gravure cylinder, an image (measurement image(s)) of the entire surface (or at least the surface to be inspected) of the gravure cylinder is captured (scanning optical capture). This image can be presented either as an overall image (measurement image) or as a plurality of individual images (measurement images). The individual images preferably overlap partially to support merging the individual images into an overall image. The measurement images(s) are preferably stored electronically and further processed. The recording position (position of the image on the surface of the printing cylinder) in which it was created is assigned to each individual image or to the overall image. This enables the localization of detected defects on the surface of the gravure cylinder.

[0039] The measurement resolution of the 2D scanned gravure cylinder surface is typically less than 10 µm. This value can be larger or smaller and depends on the defect size to be recorded and evaluated. The measurement resolution is determined by the image scale of the optical components (measurement lens) and the camera pixel size. It is essential that the measurement resolution is high enough that the smallest cells and defects to be identified can be recorded as flat image elements (imaged in a large number of pixels), which enable the shape of the recorded cells or defects to be determined. The laws of optical imaging dictate that the number of pixels in the camera must be at least twice the minimum resolution in the object field, i.e. the surface to be inspected. As an example, with a desired measurement resolution of 1 µm, this distance is mapped to at least 2 pixels in the camera.

[0040] The processing of the measurement images or the single measurement image takes place in an electronic data processing system.

[0041] Either each individual image is processed separately, or the individual images are combined into a single measurement image or a few larger measurement images and then processed. To combine the individual images into a single measurement image or a few larger measurement images, well-known image processing algorithms are used, which combine individual images into a single image.

[0042] When processing individual images or the combined single measurement image or a few large measurement images, each structure (element) detected in the measurement image is first checked to determine whether it is a cell. If this can be ruled out, a defect in the surface of the gravure cylinder is checked for. Well-known image processing methods, such as the threshold method, are used to detect structures.

[0043] Thresholding methods are well-known in the art and are based on the realization that images of structures in an image exhibit, for example, a different brightness, different reflection properties, or a different color than structure-free surface sections. Here, a threshold is defined, above which exceedance of this threshold is considered a structure to be examined. Detection of cells by scanning the first measurement image

[0044] The structures in each individual measurement image or the combined measurement image are now identified. These structures can be closed cells, open cells, or other irregularities (damage such as scratches or craters, etc.).

[0045] In principle, there are three variants for well identification or evaluation according to the invention: a) Pattern recognition and subsequent evaluation of the identified cells using known methods of shape recognition or self-developed algorithms, b) Recognition of the cells with a first neural network and subsequent evaluation of the identified cells using known methods of shape recognition or self-developed algorithms, c) Recognition of the cells with a second neural network, whereby the evaluation of the cells is also carried out with the neural network.

[0046] For this purpose, various methods are used alternatively or in combination: a) Pattern recognition

[0047] Pattern recognition (template matching) is preferably implemented using normalized cross-correlation (NCC), non-maxima suppression (NMS), and contour determination based on the centers of detected cells. This can also be achieved using image transformation methods into the frequency domain, which has a positive effect on computation time. This alternative approach is based on a mathematical transformation of the measurement image and template image into their frequency domain (=Fourier transformation) and a subsequent computation in this frequency domain. This method offers, among other advantages, faster computation.

[0048] In template matching, images of different sizes are created in advance. These different templates are applied repeatedly to structures detected in the image, and the best correlation result is recorded. A template is a type of stencil that specifies the ideal well shape. It has proven advantageous to provide approximately 5 to 10 different template sizes, also called scales, each separate for open-cell and closed-cell templates.

[0049] Normalized cross-correlation (NCC) is particularly preferred, as it has the following advantages: The value range is standardized to [-1 ... 1], allowing a direct assessment of the results as a percentage. The implementation is simple, and the formula is easy to interpret. Standardization allows for comparability between different well templates or different sizes of the same template.

[0050] Preferably, different templates are used for open-cell and closed-cell cells. The template shape present on the gravure cylinder to be examined is known from the engraving process. Pattern recognition methods also produce different results for the two template shapes, making automatic differentiation possible.

[0051] After identifying the cells, their geometric center can be determined.

[0052] Based on the geometric center of the wells, the quality of the wells can be examined. Various approaches are suitable here as well: a1) An approach based on the geometric center extracts the well contour. Algorithms such as convexity defects can be used for this. This involves finding a curve that completely encloses the well and is convex (curved outwards, away from the enclosed object) along its entire length.

[0053] The enclosing convex hull is created based on a polygon. The maximum distance between the contour points and the line segment of the enclosing polygon can be used to determine potential collapses or other damage at the edge of the cell.

[0054] a2) Once the geometric centers of the cups have been found, information can be extracted based on the created cup contour. Obvious features are moments (moments in the sense of image processing - Ming-Kuei Hu, "Visual Pattern Recognition by Moment Invariants", doi:10.1109 / TIT.1962.1057692) up to the nth order; for example, moments up to n=7 are described as Hu moments, which can be extracted efficiently.

[0055] The moments of the ideal shapes of the cells, and thus their geometric deviation, are known. The determined moments can be compared with those of the ideal shapes, and one or more deviation values ​​can be determined.

[0056] a3) TemplateDistanceRefAngle method (in-house developed algorithm): The cup contour is evaluated based on selected edge points. The expected distance of an edge point from the center of the cup is compared with the actual distance and used as a measure of the deviation from the ideal shape.

[0057] For this form of contour comparison (TemplateDistanceRefAngle) of ideal (template) well and found well, eight points located at systematic points of the circumference of the well are preferably compared with the expected positions.

[0058] Determining the 45° angles is critical. To locate the relevant (discrete) point, direct points in the neighborhood are examined to determine which quadrant the points are located in.

[0059] An improvement in speed can be achieved by using a contour list of points and calculating the absolute value of the difference and sum of the coordinates. dx1 = |x - y| and dx2 = |x + y|. The values ​​without absolute value, dx and sx, assign the quadrants. Ideally, the points at 45°, 135°, 225°, and 315° can be found directly in the list created from these values: dx 1 = 0 & sx 2 > 0 folgt P x / y = 45 ° dx 1 = 0 & sx 2 = 0 folgt P x / y = 225 ° dx = 0 & dx 2 = 0 folgt P x / y = 135 ° dx < 0 & dx 2 = 0 folgt P x / y = 315 °

[0060] An example for the distances of dx1 would be dx1 = (4 3 2 0 1 3 6 ...) The point corresponding to 0 in the list dx2 allows for direct verification of the quadrant. The advantage lies in the fast calculation, as well as the detection of non-ideal points in the neighborhood. This is the case with an irregular bowl shape, which would lead to a sequence of values ​​without 0 (e.g., dx1 = (4 3 2 1 1 3 4 ...)). The first point with a value of 1 in the list can then be selected.

[0061] The described quality measure in conjunction with the developed approach of template matching on multiple scales in combination with a small amount of image preprocessing using morphological operators works very well for isolated wells.

[0062] A typical regular print pattern also contains cells that are connected to each other via a channel, i.e., are no longer isolated. To detect and evaluate these, a case distinction is made using a second template.

[0063] One way to detect connected cells is to use the second template. A regular print pattern leads to a first distinction between open cells at the edge and in the inner area of ​​the print area.

[0064] There are areas in which the cells are recognized with the known template for closed cells. There are also areas in which the template for closed cells fails.

[0065] The second template for open cells can then be used to identify all unidentified cells. Ideally, the open cell will have two curves corresponding to the unclosed well ridges.

[0066] If the template for open cells does not match the structure found, it is most likely a defect on the surface of the printing roller.

[0067] One way to use existing methods for evaluating the detected cell or other irregularity on the surface of the printing plate (or in its measured image) is to insert a frame into the examining neural network or other software used for measured image processing and evaluation. This (preferably rectangular) frame encloses the detected cell or other irregularity. This frame, in turn, can be used to determine the center of the cell. Proven contour tracking and evaluation algorithms, similar to those used for closed cells, can then also be applied to open cells. Algorithms based on the "TemplateDistanceRefAngle" are also preferred. Alternatively, suitable contour tracking and evaluation algorithms can also be used without an additional frame. These are then oriented, for example, to the outer contour of the cell or other irregularity.the other irregularity.

[0068] If, in the case of a broken bridge, more than two closed contour surfaces result between two actually connected cups, this can be concluded to be a defect.

[0069] The currently used algorithm for defect detection consists of the following steps: 1. Image preprocessing: contrast enhancement, binarization, morphological opening 2. Determination of well centers: template matching using normalized cross-correlation (NCC) and multiple scales (template sizes) Maximum search in the result space of the scaled templates Non-maxima suppression 3. Evaluation of the wells: Contour evaluation using convexity defects from OpenCV Contour evaluation by template comparison at eight core points Contour evaluation by template comparison of the ideal ridge shape for open-cell wells 4. Differentiation of detected defects: Case a: Well contour without recognizable deformation: Well Case b: Well contour with recognizable deformation: Defect on well Case c: No well contour - but recognizable: Defect

[0070] The detection of defects in the well based on selected features is carried out using the parameters θ , which are derived from the evaluations already mentioned. For example, for nFeatures of the vector θ = [ θ 1 θ 2 ... θn ]

[0071] Given the rating used, as follows: θ = θ 1 θ 2 θ 3 θ 4 θ 5 θ 6 = d 45 ° d 135 ° d 225 ° d 315 ° d conv θ NCC

[0072] Herein refers dx ° the distance of the contour to the ideal (value of the template at this angle) at x degrees. The value of the convexity defects is found in dconv and the quality of the normalized cross-correlation in θ NCC .

[0073] Optionally, the process of determining (and evaluating) the fractal dimension of the detected cell can also be included. The box counting method, for example, is suitable for this. In this method, the cell is completely covered one after the other with squares of a given side length. The side length of the squares is reduced several times. If the number of required squares is plotted against the inverse of their side length using a double logarithmic scale, a straight line is created whose slope can be interpreted as the fractal dimension. Deviations of this slope from a predetermined value can be used to assess the cell's accuracy. The advantage of this method is that it works regardless of the cell's size.

[0074] A number of other similar definitions for fractal dimensions are known and suitable (https: / / de.wikipedia.su / wiki / Fraktale_Dimension#Boxcounting-Dimension).

[0075] In summary, the series of predefined properties used to evaluate the cells can include the following: area size deviation (deviation of characteristic sizes such as length, height, etc. from the template or deviating ratios of the sizes to each other), convexity errors (see a1)), convexity hull of the cells, shape matches using HU moments (see a2)), contour distances of the cells (TemplateDistanceRefAngle), fractal dimension and a method for evaluating the contour deviation (see a3)).

[0076] The results of comparing the specified properties with the target values ​​are recorded as components in a feature vector. This can be expressed as a percentage or absolute deviation from the target values. Other quantifications (e.g., logarithmic deviations, etc.) are also possible. The selection of suitable methods for generating the components of the feature vector depends on the magnitude of the influence of fluctuations in the result on the accuracy of the evaluation.

[0077] Optionally, the method for optically inspecting the surfaces of gravure cylinders and gravure plates can be supplemented by checking, based on the known images engraved on the gravure cylinder or gravure plate, whether there is a deviation of a cell position from the image data, whereby a value is determined that is included in the feature vector.

[0078] Furthermore, optionally, a value can be determined from the deviation of the position of individual cells from the regular pattern resulting from the large number of cells (deviation in the cell pattern), which is included in the feature vector. b) First neural network

[0079] Another approach to cell identification involves an initial pre-trained neural network. This network detects and marks the cells. Subsequent determination of the geometric center of gravity and contour of the detected cell enables cell examination similar to the further processing of cells detected by template matching. This again includes the determination of convexity defects or moments, as described under a1) or a2).

[0080] For cup identification, either one-stage or two-stage neural network architectures are preferably used. Two-stage methods often achieve the highest accuracy, but are significantly slower than single-stage methods (approximately a factor of 20). The most popular example is Faster-RCNN (Shaoqing Ren et al., "Faster R-CNN: Towards Real-Time Object Detection with RegionProposal Networks." In: CoRR abs / 1506.01497 (2016). arXiv: 1506.01497. url: http: / / arxiv.org / abs / 1506.01497). However, single-stage methods are also suitable for the given problem; useful architectures are Retina-Net (Tsung-Yi Lin et al. "Focal Loss for Dense Object Detection". In: CoRR abs / 1708.02002 (2017). arXiv: 1708.02002. url: http: / / arxiv.org / abs / 1708.02002) and YOLOv4 (Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. "YOLOv4: Optimal Speed ​​and Accuracy of Object Detection". In: CoRR abs / 2004.10934 (2020). arXiv: 2004.10934.URL: https: / / arxiv.org / abs / 2004.10934). YOLOv4 is particularly preferred. c) Second neural network

[0081] In one embodiment of the method according to the invention, a second pre-trained neural network detects the cells and evaluates them immediately. The result is a confidence value that represents a measure of the accuracy of the detected cell. This value is transferred as a component to the feature vector. In this embodiment, the second neural network is essentially the first neural network, extended by the evaluation function.

[0082] Preferably, both the first and second neural networks are YOLOv4 networks. Detecting defects in the unimaged areas of the gravure cylinder

[0083] In principle, the print template determines which areas of a gravure cylinder are to be imaged and which are to be unimaged. A comparison with the reference image (the expected image, which is known from pre-printing) can narrow down the areas of the measured image to be examined.

[0084] Furthermore, it is possible to determine the unimaged area by subtracting the areas in which cells were detected from the measurement image(s), so that only the areas without cells remain as areas to be examined.

[0085] The two approaches above can be used to speed up processing.

[0086] In general, however, a complete measurement image is examined for error characteristics.

[0087] The inspection of a completely unimaged gravure cylinder corresponds in terms of manufacturing technology to that of a copper surface or an imaged gravure cylinder with a chrome-plated surface in areas where no cup shapes are present.

[0088] A copper surface appears dull and reflective with a surface structure and roughness created by grinding stones or sanding belts.

[0089] Typical causes of defects include metallic copper defects caused by problems in the galvanic copper plating process, as well as streaks, scratches, holes, dents, etc.

[0090] The detection of defects in the unimaged areas of the gravure cylinder is preferably carried out by means of i. thresholding for object detection, and / or ii. a pre-trained third neural network

[0091] Thresholding methods are well-known in the art and are based on the knowledge that images of defects exhibit, for example, different brightness, different reflection properties, or a different color than undamaged surface sections. Here, a threshold is set based on empirical values, optionally depending on the surface coating, above which exceeding a threshold is considered a defect.

[0092] Alternatively or additionally, the unimaged area can also be examined with a third neural network. A convolutional neural network using a Patch Distribution Modeling Framework (PaDiM) is preferred. PaDiM uses a pre-trained convolutional neural network (CNN) for patch embedding and multivariate Gaussian distributions to obtain a probabilistic representation of the normal class. It also leverages correlations between the different semantic layers of CNN to better localize anomalies. PaDiM outperforms current approaches for anomaly detection and localization. PaDiM's state-of-the-art performance and low complexity make it a good candidate for many industrial applications.

[0093] Defects detected by thresholding or the third neural network are then evaluated using an anomaly defect value. The anomaly defect value is a characteristic value for the severity of a defect and thus provides a measure of the defect. The anomaly defect value is evaluated, as explained below, by incorporating the anomaly defect value into a feature vector and using a support vector machine (SVM) to check whether the feature vector (and thus also the anomaly defect value) lies within a permissible volume. The anomaly defect value is determined by taking into account some or all of the following properties of the defect: the area, the extensions in at least two different directions, and the angle of incidence relative to the unwinding direction of the gravure cylinder.

[0094] For example, an "ellipse fitting" is applied to the defect contour detected by PaDiM using OpenCV, resulting in the defect area being identified with a rectangle and a direction vector. Area, maximum distance between two points, or direction vector are all suitable quality parameters here.

[0095] An example calculation method for the anomaly defect value is given in equation (1).

[0096] Anomaly defect value for scratches: Adw = L B x A Amax

[0097] Where Adw. is the anomaly defect value, L is the length, B is the width, A is the area, F1, F2 are the vertices of the enveloping ellipse, Amax is the empirical value of a critical defect area, α is the angle to the cylinder axis. If necessary, a function (Func(α)-) on the angle α is added to equation (1), which describes the effect of the orientation of the angle to the cylinder axis. This function is a function of properties of the printing ink, such as viscosity. It is determined empirically, so no generally valid function can be specified.

[0098] In other words, the magnitude of the anomaly defect value is calculated from the ellipse parameters. The ellipse belongs to the conic sections and forms a closed curve that has the shape of a compressed circle and runs around two fixed points, the foci (where the distance from one focal point and the distance from the other focal point add up to the same sum everywhere). If the values ​​of the major vertices L and B, the ellipse area A, and optionally the ellipse major axis angle α obtained from the ellipse fitting are entered into the formula, the magnitude is calculated. The value of Amax is an empirical value similar to a threshold and indicates when a defect or scratch has a negative impact on product quality. This depends on the product requirements.

[0099] The results of the individual methods for evaluating cell shapes and surface shapes are incorporated into the feature vector with different, predetermined proportions. Preferably, the methods for evaluating cell shapes provide a common value that is 100% included in the feature vector. The methods for evaluating deviations in cell positions and patterns are also each included 100% in the feature vector. The same applies to the anomaly defect value for structures from the unimaged area. Thus, the results for each structure from the imaged area (for each cell) and also from the unimaged area (for each defect, such as scratches, holes, etc.) are combined into a feature vector.

[0100] Optionally, all features can be combined into a single feature vector, along with a note of their origin. From this, a suitable subset can then be passed to one or more Support Vector Machines (SVMs) for classification.

[0101] Once all features have been calculated and recorded as components in the feature vector, they are classified. The classification results provide a final quality assessment of the tested product. A Support Vector Machine (SVM) is used as the classifier, which is characterized by high processing speed and good results.

[0102] Support Vector Machines are procedures for supervised Machine Learning,which are primarily used for binary classification. The training data is plotted in n-dimensional space, and the algorithm attempts to draw a boundary with the greatest possible distance to the nearest sample.

[0103] The RBF (radial basis function) kernel used provides better classification results than a linear kernel. The two other SVM parameters are the C and the gamma parameter, where C describes the size of the penalty when a sample is misclassified and gamma encompasses the influence of each individual sample. The results of the study yielded values ​​for C of 1.0 and gamma of 0.8.

[0104] The properties of the SVM allow the characteristics in the feature vector to be transferred to higher dimensions (kernel trick). A hyperplane is fitted in this space. The SVM was trained with test data so that values ​​within the enclosed volume are considered good and values ​​outside are considered errors. It has been shown that this trained SVM model, using data from real measurements, can achieve very good evaluations.

[0105] The evaluation is carried out for all structures in all measurement images (or the one large measurement image).

[0106] If all structures in all measurement images lie within the enclosed volume, the gravure roller is considered defect-free. Otherwise, the defect location is determined from the coordinates of the defective structure.

[0107] The SVM provides a corresponding output that includes at least "error-free" or "errored". Training of neural networks and support vector machines

[0108] The training of the neural networks used preferably takes place in three stages: 1. Training using synthetic images (images of roller surfaces) with ideal shapes of cells (not applicable for the third neural network), 2. Training using real images of surfaces of gravure cylinders and corresponding identification of the objects in the images, 3. Training using real images that were categorized by a Support Vector Machine (SVM).

[0109] In a first training step, the first and second neural networks are trained using ideal shapes of the respective cup shapes, which are generated as 2D reference shapes. Subsequently, in a second training step, they are trained using known data (real images) obtained from human inspections of the surfaces of gravure cylinders and gravure plates.

[0110] In addition to the real images of the surfaces of gravure cylinders or gravure plates, the data for training the first and second neural networks in the second step also contained information about whether image objects in the images were cells.

[0111] In the second step, the data for training the second neural network contains, in addition to the images of the surfaces of gravure cylinders or gravure plates and the information on whether image objects in the images are cells, an evaluation parameter for the severity of a defect on a respective cell, i.e. whether these cells are to be assessed as damaged or usable.

[0112] For the third neural network, which is designed to scan the unimaged areas for defects, the first step is omitted. The subsequent steps are analogous to the other two neural networks. Here, too, the data for training the third neural network in the second step contains not only images of the surfaces of gravure cylinders or gravure plates, but also information about whether image objects in the training images are defects.

[0113] The third neural network is preferably a PaDiM network – A Patch Distribution Modeling and Localization – for anomaly detection. The third neural network was trained with good and defective images and demonstrated very good error detection.

[0114] The training steps can also be integrated into one another by starting with the ideal shapes of the respective cup shapes, which can be mathematically formulated and generated as a 2D reference shape (blueprint). In subsequent optimization runs, these are increasingly expanded with data (cup values) from the real process.

[0115] The images of the two image sequences Fig. 5 and Fig. 6 have the following meaning, from left to right. Input: Image = Original image Input: GroundTruth = Masked error area Output, Result: Predicted heat map, Anomaly in false colors Output, Result: Predicted mask, Mask of the error area Output, Result: Segmentation result, Marking of the anomaly

[0116] The Support Vector Machine is defined in its parameters for classification by correcting the parameters through human intervention until the evaluation of existing measurement images by humans and SVM agree to a specified extent.

[0117] From the evaluation of the feature vector of each well and the anomaly defect value of each defect and the one or more measurement images, the SVM provides a database for training data to improve the detection of wells by the first neural network and / or to improve the detection and evaluation of wells by the second neural network and / or to improve the detection of defects by the third neural network. The neural networks are trained with this training data in a second training step. In particular, the training data is prepared for the training process of the neural networks as part of a human pass / fail assessment of the wells based on the actual manufacturing process. About the neural networks used

[0118] To determine the image quality measure, an algorithm trained using a machine learning method can be used. A neural network, particularly a convolutional neural network, can be used as the algorithm. Machine learning algorithms, especially convolutional neural networks, are particularly suitable for recognizing and considering a wide variety of image properties at different resolution levels. Compared to known image quality measures, a machine learning algorithm can learn a wide variety of suitable image properties and combine them in the best possible way.

[0119] The algorithm can process known, ideally generated, synthetic image data as input data. However, it is also possible to enrich this input data with real-world image data. In its simplest form, the algorithm can output a single scalar as an output value, which is the measure of image quality. This enables, as will be explained in more detail below, simple training of the algorithm and simple further processing of the output data. However, feature vectors or similar can also be advantageously output, for example, to analyze multiple dimensions of image quality separately.

[0120] In neural networks, especially convolutional neural networks, image features can be extracted from the input data in several successive layers by convolution, application of activation functions, and subsampling. A so-called "fully connected layer" can be used as the final layer to combine the image features into a scalar, which serves as a measure of image quality. The function of neural networks, and especially convolutional neural networks, is well known in the state of the art from other application areas and will therefore not be described in detail.

[0121] The algorithm can be trained, in particular, through supervised training, in which individual training data sets each contain input data for the algorithm and a target result. The training data sets can each contain two-dimensional image data or reference template data at multiple scales.

[0122] The two-dimensional image data or template data can be constructed with a known error value. This allows a measure describing this error to be included as the target result in the respective training data set. For this purpose, error feedback can be used, for example. This can be done manually, with appropriate effort, or through a programmed software function, e.g., to simulate the wear of the engraving stylus. This function produces faulty two-dimensional template data, which, as such, are entered into the network training with the attribute "defective template."

[0123] Various possibilities for training machine learning algorithms are known in the state of the art and will not be explained in detail.

[0124] The algorithm can be trained using training sets, each of which contains image data (e.g., for defective cup shapes) and a target value for the image quality measure to be determined for this image data. The image data are constructed from predefined reference images in such a way that, during construction, a modified target value is determined according to a modification specification as a function of the modification specification. As explained above, this makes it possible to construct the two-dimensional image data with a known error, and the training data set includes a measure describing this error.

[0125] In this case, reference images can first be determined whose imaging geometry is known with high accuracy. Such reference images can be determined, for example, by imaging a known object, in this case gravure cells, at high resolution. The creation of these reference images and the associated quality values ​​is carried out by a calibrated device. The highly precise and essentially error-free acquisition geometry of these cell images can then be modified. A variety of different modifications can be made to a set of cell images. For example, the same or different 2D contour deformations can be applied to different cell images. An optimized training dataset can then be created using one or more reconstructed image datasets.

[0126] Since the extent to which the cup geometry has been modified is known, this training dataset can be used as an extended template dataset in the new training run. This results in new limits for the quality criteria in the feature vector.

[0127] In summary, the invention thus also relates to a method for training an algorithm using a machine learning method. In particular, supervised learning can be carried out, and the training data sets can be specified as described above. The invention also relates to an algorithm trained using such a training method for determining a measure of a cup quality of image data reconstructed from reference images, or, as a method result, to parameters of an algorithm that are determined within the scope of such a machine learning method. Furthermore, the invention relates to a computer-readable medium that stores the trained algorithm in the form of readable and executable program sections, or that stores parameters determined within the scope of the machine learning method for parameterizing such an algorithm.

[0128] In the process for determining deviations from target specifications in cell images, the measurement results of the individual cells are assigned to a feature vector. This can serve as an input data set for a classifier to obtain a quality measure of the overall set of measurement results from the sum of the individual values. Positive experiences have been made with the use of a support vector machine (SVM). The support vector machine (SVM) is a mathematical method used in the field of machine learning. It allows the classification of objects and has a wide range of applications.

[0129] The class boundaries determined by the support vector machine are so-called large margin classifiers and leave as wide a region as possible free of objects around the class boundaries. The SVM supports both linear and nonlinear classification. For nonlinear classification, the so-called kernel trick is used, which extends the object space by additional dimensions (hyperplanes) to represent nonlinear separation surfaces.

[0130] As a result of the pyramid shown (see figure Fig. 12 ) the hyperplane is displayed in 3D space on the result data set of the feature vector.

[0131] Furthermore, the invention relates to a computer program which can be loaded directly into a memory device of a processing device, with program sections in order to carry out all steps of the method according to the invention when the computer program is executed in the processing device. About the YOLO Object Detection and the training process

[0132] YOLO is a widely used object detection method published in 2015 by Redmon et al. (Joseph Redmon et al., You only look once: Unified, responsive object detection. Techn. Ber. 2016, pp. 779-788. DOI: 10.1109 / CVPR.2016.91. arXiv: 1506.02640v5. URL: http: / / pjreddie.com / yolo / ) and has since been further developed in several steps. Object detection is viewed as a regression problem in which a neural network is used to identify objects from the pixels of the input image. A detected object is described by a bounding box that uniquely defines its position and size. Furthermore, for each detected object, as many class probabilities are determined as there are possible classes in the respective application. For each class, a class probability value is determined - the sum of these values ​​is one.A detected object is assigned to the class with the highest class probability. A central goal of YOLO is real-time object detection. Due to its properties as a one-stage object detector, only one neural network is used for the entire object detection process, which is run through each image – the image is viewed only once (you only look once). This allows YOLO to be optimized end-to-end for its object detection performance. In contrast, the group of two-stage object detectors can be mentioned, which includes the methods Region Based Convolutional Neural Network (RCNN) (Ross Girshick et al. Rich feature hierarchies for accurate object detection and semantic segmentation Tech report (v5). Techn. Ber. arXiv: 1311.2524v5), Fast RCNN (Ross Girshick. Fast R-CNN. Techn. Ber. arXiv: 1504.08083v2. URL: https: / / github.com / rbgirshick / ) and Faster RCNN (Shaoqing Ren et al.Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. Techn. Ber. 6. 2017, pp. 1137-1149. DOI: 10.1109 / TPAMI.2016.2577031. arXiv: 1506.01497). In simple terms, two-stage methods create so-called region proposals in a first step, i.e. suggestions for areas in which objects could be located. In a second step, these regions are used and the content is viewed as a classification problem, i.e., the class of the imaged object is determined (Ross Girshick et al. Rich feature hierarchies for accurate object detection and semantic segmentation Tech report (v5). Techn. Ber. arXiv: 1311.2524v5).

[0133] Since convolutional neural networks are state of the art and are known from other application areas, a more detailed description is unnecessary for the person skilled in the art.

[0134] The training dataset includes approximately 2,000 images, each with closed cup templates and open cup templates of various sizes. The selected template image sizes were based on the dimensions of the actual gravure cup mold.

[0135] The YOLOV4 configuration parameters were selected according to the training dataset of template images and the task to generate a high-quality trained network. An image size of 416x416 pixels was chosen, with Scale_x_y = 1.2, jitter = 0.3, momentum = 0.949, and exposure = 1.5.

[0136] To optimize training time, the YOLOV4 training was conducted with the support of GPU processors. A GPU is a graphics processing unit (GPU), a computer chip that performs mathematical calculations particularly quickly, offloading the CPU (central processing unit) or training neural networks for artificial intelligence (AI) applications.

[0137] It was possible to reduce training time by a factor of five to eight with GPU support.

[0138] The suitability of a CNN network for the described application was confirmed by the positive optimization curve in the loss function graph. The output of the last CNN layer is a tensor in the form (S,S,(Bx5+C)), whose content represents the network's prediction for an input image and achieved a value of more than 95%. Ultimately, the optimization graph of a loss function is obtained during the training run, which can take several days depending on PC power. Corresponding figures can be found at https: / / www.researchgate.net / figure / Improved-YOLOv4-training-loss-function-curve_ fig1 _350756829.

[0139] Preferred developments of the invention result from the combinations of the claims or individual features thereof.

[0140] The invention will be explained in more detail below using an exemplary embodiment and the accompanying figures. The exemplary embodiment is intended to illustrate the invention without limiting it. Figures

[0141] Fig. 1 This diagram shows the structure of the optical inspection unit, which scans the measurement image(s) of the surface of the gravure cylinder. The arrows indicate the beam path. Fig. 2 explains the scanning process of measuring images in more detail. Fig. 3 shows the photograph of an area with closed cells. Fig. 4 shows a photograph of an area with open cells. Fig. 5 shows example data used to train the third neural network, here the surface defect "hole and small scratch" as well as the output data of the third neural network. Fig. 6shows example data used to train the third neural network, here the surface defect "prominent scratch" as well as the output data of the third neural network. Fig. 7 schematically explains the process of template matching with predefined patterns (templates) of different sizes. Fig. 8 shows an example of the result of the normalized cross-correlation for closed cells and the geometric centroids determined from it. A template of 183 x 128 pixels was used. Fig. 9 shows an example of the result of the normalized cross-correlation for open cells Fig. 10 shows examples of the results for the assessment of the cell contour using "convexity defects". Fig. 11 shows an example of the use of the "convex hull" method, in which the convexity of the enveloping curve is considered. Fig. 12shows a schematic representation of the evaluation space of a support vector machine. Structures whose values ​​lie within the pyramid are considered error-free, while structures whose values ​​lie outside the pyramid are considered faulty. Fig. 13 shows the flow of the method according to the invention. The "improve learning" -> "yes" branch is only processed during the training phase of the neural networks or SVMs. Fig. 14 Explains the determination of the anomaly defect value and names the parameters of a fitted ellipse around a defect, in this case a scratch, which are used in the formula for calculating the anomaly defect value. Including: Adv. anomaly defect value, L - length, B - width, A - area, F1, F2 - vertices of the enveloping ellipse, Amax - empirical value, α - angle to the cylinder axis. Example

[0142] The process described here comprises the optical inspection of gravure cylinders (printing cylinders) in the micrometer range, a multi-stage, fully automated process for image acquisition, image processing, and quality assessment. Intelligent and self-learning software engineering components, including AI (artificial intelligence) and machine learning, are used.

[0143] The image is captured by an optical scan of the cylinder surface. Typically, the workpiece to be inspected, the gravure cylinder (gravure roller), is rotated and the optical image capture unit is moved parallel to the axis across the entire width of the gravure cylinder. Alternatively, the test specimen could be fixed and a movable image capture unit on an xy carriage could scan a 2D surface, which would be used in the inspection of rectangular test plates, e.g. a Ballard skin. The cylinder is typically held in a rotation unit by clamping units at the respective ends of the axis. Alternatively, mounting on rotating rollers is possible. In this case, the printing cylinder would be driven and guided on its surface in the rollers. Other design variants are also conceivable, the arrangement of which would be selected according to the application.This also includes inline inspection in existing production machines, e.g., the copper surface treatment of the gravure cylinder after the finishing process, as a final quality control step. An inline incoming inspection in one of the subsequent process steps, including the imaging process, represents another inline option.

[0144] To determine the position of the recorded measurement image on the cylinder surface, the respective angular position is read via a rotary encoder on the drive unit or an alternative mounting. Combined with the second location information from the axis-parallel linear unit, each recorded camera image can be located in 2D space (unfolded cylinder surface). Subsequently, several individual images are combined into a single measurement image using "stitching" (combining the individual images).

[0145] The entire cylinder installation and removal, as well as the cylinder handling, can be carried out independently within an existing inspection unit or automated process line.

[0146] In this context, after completion of the setup process, the mechanical reference points of the printing cylinder can be automatically recorded, including the cylinder width, its diameter as well as the automatic reference point positioning of the image recording unit using autofocus.

[0147] An optical inspection unit is used for image capture: Camera, lens and coaxial illumination component Image acquisition and evaluation: Using a camera system and image acquisition software, the entire cylinder surface is scanned and then fed into a specially developed software post-processing.

[0148] The components of the optical image capture unit include an area scan camera with a lens, a light source with a condenser lens, and a deflecting mirror for coaxial illumination. The surface conditions of the gravure cylinders to be inspected are, on the one hand, the unimaged copper surface, which is subsequently imaged and chrome-plated. In the unimaged state, the copper surface appears metallic with a matte finish; in the imaged state, it has a reflective, high-gloss finish due to the chrome plating. Both surfaces contain a desired degree of surface roughness. According to the laws of physics, the sharpness of the image increases with shorter-wavelength light, which is why a pulsed LED in the blue frequency range of 450 nm is used as the light source.

[0149] To make qualitative statements regarding critical surface defects in the micrometer range, a correspondingly high resolution of the measurement field is necessary. To achieve this, the measurement field size, lens, and camera must be coordinated. The required magnification, i.e., the ratio of object size to image size, can be achieved via a camera setting (binning, etc.) or alternatively by zooming the lens.

[0150] This setup allows the object to be inspected to be analyzed fully automatically at two or more optical resolutions. An optimal measurement resolution adapted to the defect size has a positive effect on the large amount of data and significantly reduces it.

[0151] When capturing the surface of a gravure cylinder (30 cm cylinder diameter, 100 cm barrel length) with a surface to be inspected of approx. 10000 cm 2<, approx. 150000 images are created with a resolution of 5120 x 5120 pixels at an object resolution of 20µm.

[0152] The images are saved and represent a data volume of approximately 190 MB per image.

[0153] After the individual images are combined using software, a 3.41 terabyte measurement image is created. The measurement image is combined, where possible, in parallel with the image acquisition to optimize the measurement time. The measurement image is saved for further processing.

[0154] Next, the detection of cells in the measurement image takes place. For this purpose, the measurement image is examined for structures (elements) using well-known image processing algorithms. This is where the threshold algorithm comes into play.

[0155] When a structure is found on the surface, it is checked whether it is a cell.

[0156] The threshold algorithm is used for this purpose. A threshold value of 0.8 was set as the threshold by which a structure must differ from the untreated or undamaged substrate (the background in the measurement image).

[0157] Once a structure has been detected, the template matching procedure is used to check whether it is an open well.

[0158] The test for open wells is performed with 10 templates of different sizes for closed wells using normalized cross-correlation. If all templates remain below the specified correlation value of 0.8, the detected structure is examined with templates for open wells. Here, the correlation threshold is also 0.8. The correlation value indicates that wells that exceed the correlation value sufficiently match the respective template.

[0159] If the structure detected is neither a closed nor an open cell, it is clearly another surface defect. Therefore, this structure is treated using the procedure for non-imaged areas.

[0160] If the structure is an open or closed cell, the center point (geometric center of gravity) of the cell is then determined. This is done using well-known, state-of-the-art algorithms.

[0161] After finding the geometric center of gravity, the shape of the cup is assessed.

[0162] In this case, the following procedures are combined: 1. Area factor - measure of the area ratios as well as length, width 2. Convexity Defects according to https: / / theailearner.com / 2020 / 11 / 09 / convexity-defectsopencv / 3. Convex Hull according to https: / / de.wikipedia.org / wiki / Konvexe_Hülle 4. Match Shapes using HU Moments https: / / de.wikipedia.org / wiki / Konvexe_Hülle 5. TemplateDistanceRefAngle: Contour distances "contour polygon" and reference template at different angles e.g. 0, 45 degrees, 90 degrees etc. - see above 6. Testing for web breakage and web deformations in open cells 7. Fractal dimension - see above

[0163] The methods provide values ​​for the deviation from specified target values. This deviation is expressed here as a percentage. Furthermore, the results of the methods contribute to the final result to varying degrees (see Table 1).

[0164] The measured image is then passed to the third neural network. This network is based on the PaDiM framework. It uses a pre-trained convolutional neural network. The third neural network detects structures that are not cups and classifies them as defects.

[0165] The third neural network delivers an anomaly defect score as a quantification of the defect extent. This score is calculated by applying an "ellipse fitting" to the defect contour detected by PaDiM using OpenCV (according to equation (1)). The result is that the defect area is identified with a rectangle and a direction vector. Area, maximum distance between two points, or direction vector are suitable quality measures here.

[0166] The results of the bowl evaluations and the anomaly defect value are included in the feature vector with the weights given in the following table: Tab. 1: Elements of the feature vector as well as weights of inclusion in these and tolerances of the individual values Nr. Proceedings Result weighting Tolerance / threshold 1 Area deviation / area factor 15% 10% 2 Convexity defects 25% 15% 3 Convex Hull 10% 15% 4 Match Shapes using HU Moments 5% 10% 5 TemplateDistanceRefAngles 35% 5% 6 Fractal dimension 10% 5% sum 100% 7 Deviation of a cell position 100% 10% 8 Deviation in the cell pattern 100% 20% 9 Anomaly defect value 100% 25%

[0167] Once the complete feature vector is available, it is classified using a support vector machine. The SVM considers a space with a dimension equal to the number of elements in the feature vector. Within this space, the feature vector defines a point. Based on experience and human knowledge, a subspace was defined within this space. If the point belonging to a structure lies within this subspace, the structure is considered to be OK or non-critical. If all structures in all measurement images lie within this subspace, the gravure roller is considered to be defect-free. Otherwise, the location of the defect is determined from the coordinates of the defective structure.

[0168] The SVM provides a corresponding output that includes at least "error-free" or "errored". Reference symbol

[0169] 1Print roller 2Blue LED 3Lens 4Beam splitter 5Objective lens 6Deflection mirror 7Camera 8Resolver for detecting the position of the image on the roller 9Data processing system DDefect BDew of the defect LLength of the defect AFarea of ​​the defect Amaxempirical value from measurements F1, F2Vertices of the enveloping ellipse αAngle to the cylinder axis of the gravure cylinder Cited non-patent literature

[0170] Ming-Kuei Hu, "Visual Pattern Recognition by Moment Invariants", doi:10.1109 / TIT.1962.1057692 Shaoqing Ren u. a. "Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks". In: CoRR abs / 1506.01497 (2016). arXiv: 1506.01497. url: http: / / arxiv.org / abs / 1506.01497 Tsung-Yi Lin u. a. "Focal Loss for Dense Object Detection". In: CoRR abs / 1708.02002 (2017). arXiv: 1708.02002. url: http: / / arxiv.org / abs / 1708.02002 Alexey Bochkovskiy, Chien-Yao Wang und Hong-Yuan Mark Liao. "YOLOv4: Optimal Speed and Accuracy of Object Detection". In: CoRR abs / 2004.10934 (2020). arXiv: 2004.10934. url: https: / / arxiv.org / abs / 2004.10934 https: / / de.wikipedia.su / wiki / Fraktale_Dimension#Boxcounting-Dimension (Stand, 04.04.2022) https: / / theailearner.com / 2020 / 11 / 09 / convexity-defects-opencv / https: / / de.wikipedia.org / wiki / Konvexe_Hülle https: / / cvexplained.wordpress.com / 2020 / 07 / 21 / 10-4-hu-moments / https: / / www.researchgate.net / figure / Improved-YOLOv4-training-loss-function-curve_ fig1_350756829 Ross Girshick. Fast R-CNN. Techn. Ber. arXiv: 1504.08083v2. URL: https: / / gith ub.com / rbgirshick / . Ross Girshick u. a. Rich feature hierarchies for accurate object detection and semantic segmentation Tech report (v5). Techn. Ber. arXiv: 1311.2524v5. Joseph Redmon u. a. You only look once: Unified, reaktime object detection. Techn. Ber. 2016, S. 779-788. DOI: 10.1109 / CVPR.2016.91. arXiv: 1506.02640v5. URL: http: / / pjreddie.com / yolo / . Shaoqing Ren u. a. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. Techn. Ber. 6. 2017, S. 1137-1149. DOI: 10.1109 / TPAMI.2016 .2577031. arXiv: 1506.01497. Thomas Defard u. a. "PaDiM: a Patch DistributionModeling Framework for Anomaly Detection and Localization. In: CoRR abs / 2011.08785 (2020). arXiv: 2011.08785. URL: https: / / arxiv.org / abs / 2011.08785. https: / / github.com / xiahaifeng1995 / PaDiM-Anomaly-Detection-Localization-master

Claims

1. Method for visually inspecting the surfaces of gravure cylinders and gravure plates, comprising at least the following steps: a) Place the gravure cylinder or gravure plate into an inspection device, b) Scanning optical detection of the surface of the gravure cylinder or the gravure plate by means of an image capture unit in one or more measurement images and detection of the position of the measurement images relative to the surface of the gravure cylinder or the gravure plate c) Detection of cells by scanning the first measurement image, d) Comparison of a series of predetermined properties with the given target values assigned to these properties and recording the deviation of each property from the respective target value as a value in a feature vector, e) Detection of defects in the areas of the first measurement image in which no cells were detected, f) Determination of an anomaly defect value of a detected defect, g) Classifying the feature vector of each cell and the anomaly defect value of each defect based on predetermined result weights and allowable tolerance values of each property of the feature vector using a pre-trained support vector machine (SVM) and, h) Repeat steps c) to g) for each additional measurement image, i) Output of an evaluation of the gravure cylinder or gravure plate, whereby the evaluation includes at least the information "fault-free" or "faulty".

2. Method for visually inspecting the surfaces of gravure cylinders and gravure plates according to claim 1, characterized in that after step c) an overall image of the surface of the gravure cylinder or the gravure plate is assembled from the measurement images and the method is carried out with this assembled overall image as a measurement image.

3. Method for visually inspecting the surfaces of gravure cylinders and gravure plates according to claim 1 or 2, characterized in that step c) is carried out by means of i. Pattern recognition (template matching) with determination of geometric center of gravity and contour of the recognized cell, and / or ii. a first pre-trained neural network that recognizes and marks the cells and subsequent determination of the geometric center of gravity and contour of the recognized cell,4. Method for visually inspecting the surfaces of gravure cylinders and gravure plates according to one of the preceding claims, characterized in that steps c) and d) are carried out by means of a second pre-trained neural network which recognizes and evaluates the cells and transfers a confidence value as a value into the feature vector.

5. Method according to one of the preceding claims, characterized in that step e) is carried out by means of i. Threshold method for object detection, and / or ii. a pre-trained third neural network6. Method for visually inspecting the surfaces of gravure cylinders and gravure plates according to one of the preceding claims, characterized in that, on the basis of the known images engraved on the gravure cylinder or the gravure plate, it is checked whether there is a deviation of a cell position from the image data, whereas a value being determined which is included in the feature vector.

7. Method for visually inspecting the surfaces of gravure cylinders and gravure plates according to one of the preceding claims, characterized in that a value is determined from the deviation of the position of individual cells from the regular pattern resulting from the plurality of cells, which is included in the feature vector.

8. Method for visually inspecting the surfaces of gravure cylinders and gravure plates according to one of the preceding claims, characterized in that the measurement images are recorded with overlapping areas.

9. Method for visually inspecting the surfaces of gravure cylinders and gravure plates according to one of the preceding claims, characterized in that the pattern recognition (template matching) is carried out based on the normalized cross-correlation.

10. Method for visually inspecting the surfaces of gravure cylinders and gravure plates according to one of the preceding claims, characterized in that the series of predetermined properties includes some or all the following properties: area size deviation, convexity error, convexity hull, shape matches by means of HU moments, contour distances, fractal dimension and a method for evaluating the contour deviation.

11. Method for visually inspecting the surfaces of gravure cylinders and gravure plates according to claim 10, characterized in that the method for evaluating the contour deviation includes determining the deviation of the found contour of a cell from the ideal contour of the cell at several, preferably 8, points distributed on the circumference of the found cell and incorporating it into the feature vector.

12. Method for visually inspecting the surfaces of gravure cylinders and gravure plates according to one of the preceding claims, characterized in that the first and the second neural network are each a YOLOv4 network.

13. Method for visually inspecting the surfaces of gravure cylinders and gravure plates according to one of the preceding claims, characterized in that the third neural network is a convolutional neural network using a patch distribution modeling framework.

14. Method for visually inspecting the surfaces of gravure cylinders and gravure plates according to one of the preceding claims, characterized in that the anomaly defect value is determined considering some or all of the following properties of the defect: the area, the dimensions in at least two different directions, the angle of progression.

15. Method for visually inspecting the surfaces of gravure cylinders and gravure plates according to one of the preceding claims, characterized in that the SVM is determined in its parameters for classification by correcting the parameters by human intervention until the evaluation of existing measurement images by humans and SVM agree to a predetermined extent.

16. Method for the visual inspection of the surfaces of gravure cylinders and gravure plates according to one of the preceding claims, characterized in that the neural networks are trained in a first training step with ideal shapes of the respective cup shapes, which are generated as 2D reference shapes, followed in a second step by training with known data which originate from inspections of the surfaces of gravure cylinders or gravure plates.

17. Gravure plates according to claim 16, characterized in that the data for training the first and second neural network in the second step contain, in addition to the images of the surfaces of gravure cylinders or gravure plates, information about whether image objects in the images are cells.

18. Method according to claim 16, characterized in that the data for training the second neural network in the second step further contains, in addition to the images of the surfaces of gravure cylinders or gravure plates and the information as to whether image objects in the images are cells, an evaluation parameter for the severity of a defect at a respective cell.

19. Method for visually inspecting the surfaces of gravure cylinders and gravure plates according to claim 16, characterized in that the data for training the second neural network in the second step contain, in addition to the images of the surfaces of gravure cylinders or gravure plates, information on whether image objects in the images are cells and whether these cells are to be assessed as damaged or usable.

20. Method for visually inspecting the surfaces of gravure cylinders and gravure plates according to claim 16, characterized in that the data for training the third neural network contain, in addition to the images of the surfaces of gravure cylinders or gravure plates, information as to whether objects in the images are defects.

21. Method for visually inspecting the surfaces of gravure cylinders and gravure plates according to one of claims 16 to 20, characterized in that the SVM generates training data for improving the detection of cells by the first neural network and / or for improving the detection and evaluation of cells by the second neural network and / or for improving the detection of defects by the third neural network from the evaluation of the feature vector of each cell and the anomaly defect value of each defect and the one or more measurement images, and the neural networks are trained with this training data in a third training step.