MONITORING DIGITAL PRINTING PROCESSES USING A NEURAL NETWORK

Description

DESCRIPTION Background of the present invention Field of the present invention The present invention relates generally to the field of analysis of 5 image, and more specifically the use of image analysis for monitoring of digital printing processes. Background of related technique By digital printing process we mean a process that allows you to 10 print a digital image directly onto a substrate without the use of plates of printing. Digital printing processes can be used in a wide variety of sectors applications, using a variety of different substrates, such as paper, cardboard, fabrics, wood, metal, glass. Quality control is of strategic importance in the printing sector 15 digital, and has as its final objective the prevention of the placing on the market of final products that do not meet specific quality standards. Product quality final (from now on, print product or simply print) depends strongly influenced by the presence / absence of printing defects that may have formed during printing operations. Identify the presence of defects as early as possible 20 printing in the print product is particularly important to ensure the possibility to stop the printing process (and trace the problem that gave rise to the defects (printing) quickly enough to reduce waste as much as possible energy, inks, and substrate. Inspection of print products to identify printing defects 25 is a difficult and time-consuming procedure. By doing so, example reference to the sector of digital printing on textile substrates, in which images digital are printed on substrates including fabrics, an inspection of this type may require the intervention of highly qualified personnel to perform complicated and detailed analyses on even very large portions of printed fabrics. 30 Thanks to the ever-increasing diffusion of neural networks in industrial sectors most disparate, in recent times we have begun to evaluate the possibility of exploiting the potential of such neural networks to support or even replace staff qualified when performing inspections on print products. Neural networks, in fact, are a computer tool that finds use increasingly in all those sectors that can benefit from analytical capabilities 5 automatic image processing offered by this tool. In particular, a system of monitoring that uses a neural network model trained to detect the presence of potential print defects in images of print products collected during the printing operations could advantageously and efficiently replace the long and complex manual inspection work by qualified personnel, and 10 provide real-time results during printing operations. Summary of the present invention The Applicant found that training a neural network from use in a digital printing system monitoring system to detect 15 The presence of printing defects is a very critical operation. Indeed, in order to effectively train a robust neural network in able to detect the presence of printing defects in a digital printing process, is It is necessary to set up a comprehensive training database, including a large quantity of training images of different types, concerning different types of 20 printing, different types of substrate and different types of printing defects. However, Generating a training database of this kind is time-consuming and inefficient manual operations for acquisition (e.g., using a camera) of images of print process results. Such manual operations would require an amount of time, and a corresponding consumption of electricity and ink, 25 excessive, since, given the rarity of the occurrence of printing defects in real cases, it would be necessary to collect countless images acquiring print results of printing processes continued for very long intervals of time. In view of the above, the Applicant has devised a method and a systems that solve the above-mentioned problem. 30 In general terms, the present invention is based on the idea of ​​training the network neural exploiting (also) artifactual training images through manipulation digital so as to introduce printing defects into them. In particular, one aspect of the present invention relates to a method for train a neural network model for use in monitoring applications printing defects in a substrate printing process. The method includes, under the control of a computing system, provide the computing system with a 5 plurality of digital images to be printed. The method further comprises providing the computing system one or more digital images of substrate each representative of a corresponding substrate. The method comprises generating, from part of the computing system, a plurality of raw training images representative of defect-free prints. Said to generate the first images 10 training comprises combining each of a group of said images digital to be printed with one of the digital substrate images. The method comprises furthermore, generate, by the computing system, a plurality of second training images representing prints each containing at least one printing defect. Said generating the second training images includes 15 perform the following operations a), b) on each of a group of said images digital to print: a) digitally manipulate said digital image to print to introduce in it at least one graphic alteration representing the corresponding at least one printing defect; b) combine the digital image to be printed digitally manipulated with one of the digital substrate images. The method 20 includes training, by the computing system, said network model neural based on the said first training images and the said second ones training images to optimize the capacity of such a neural network model to classify the first training images as defect-free, and to classify the second training images as containing at least one defect. 25 Thanks to this method of generating training images, it is advantageously possible to quickly build a training database complete and efficient, which allows the training of one or more robust neural networks without having to first carry out long and inefficient manual operations that would be necessary in case you want to build such a training database 30 using only digital images obtained from the acquisition of real results printing processes. Through digital manipulation it is advantageously possible to generate “synthetically” (i.e. introducing artefacts performed by exploiting manipulations digital) printing defects of various types, consequently generating a database of training that guarantees excellent coverage of the numerous cases that may occur during digital printing operations. 5 In accordance with one embodiment of the present invention, said digitally manipulate said digital image to be printed to introduce into it at least one graphic alteration representing the corresponding at least one defect printing includes superimposing on said digital image to be printed at least one representative image of a stain. In this way, it is advantageously 10 it is possible to faithfully replicate a stain defect. In accordance with one embodiment of the present invention, said digitally manipulate said digital image to be printed to introduce into it at least one graphic alteration representing the corresponding at least one defect printing involves blurring said digital image using a blurring algorithm 15 blur. In this way, it is advantageously possible to faithfully replicate a blur defect. In accordance with one embodiment of the present invention, said digitally manipulate said digital image to be printed to introduce into it at least one graphic alteration representing the corresponding at least one defect 20 printing involves dividing said digital image into a first portion and a second portion, and move the first portion so that it overlaps at least partially to the second portion, or in such a way as to create a gap between the first portion and the second portion. In this way, it is advantageously possible faithfully replicate a step defect. 25 In accordance with one embodiment of the present invention, said digital images to print include repetitive digital images to print each comprising a plurality of copies of the same single image. In accordance with one embodiment of the present invention, said digitally manipulate said digital image to be printed to introduce into it 30 at least one graphic alteration comprises altering, in each of said images repetitive digital copies to be printed, the position of at least one of said copies with respect to the position of the other copies. In this way, it is advantageously possible to replicate faithfully a defect of incorrect repetition. In accordance with one embodiment of the present invention, the method further includes, under the control of the computing system, extracting from each digital image to be printed corresponding color channels. 5 In accordance with one embodiment of the present invention, said operation a) comprises digitally manipulating each color channel of said digital image to be printed to introduce at least one graphic alteration into it representing the corresponding at least one printing defect. In accordance with one embodiment of the present invention, said 10 operation a) involves combining the digitally manipulated color channels together to obtain the corresponding digitally manipulated digital image to be printed. In accordance with one embodiment of the present invention, said digitally manipulate said digital image to be printed to introduce into it at least one graphic alteration comprises deletion from at least one color channel of 15 said digital image to be printed a corresponding set of pixels. In this way, it is advantageously possible to faithfully replicate a line defect. In accordance with one embodiment of the present invention, said digitally manipulate said digital image to be printed to introduce into it at least one graphic alteration involves translating one or more of the color channels of 20 said digital image to be printed with respect to the other color channels of said image so as to create a misalignment between the said one or more of the color channels and said other color channels. In this way, it is advantageously possible to replicate faithfully a color misalignment defect. In accordance with one embodiment of the present invention, said 25 method comprises, under the control of the computing system, scaling said digital images to be printed at a first resolution corresponding to a print resolution used for printing said digital images to be printed on substrate. In accordance with one embodiment of the present invention, said 30 method comprises, under the control of the computing system, generating said first training images and said second training images using said digital images to be printed scaled to said first resolution. This way, the synthetic generation of printing defects is carried out at the same resolution used for actual printing, increasing its fidelity. In accordance with one embodiment of the present invention, said the method comprises, under the control of the computing system, scaling said first 5 training images and said second training images generated at a second resolution, said training said neural network model being performed by the computing system on the basis of the said first images of training and of said second training images to the second resolution. 10 In accordance with one embodiment of the present invention, said second resolution corresponds to the resolution of images acquired by a printing process monitoring system during printing operations. In This way, the neural network used by the monitoring system will be advantageously trained using training images having the same 15 resolution of the images that the neural network will have to classify during its operation. In accordance with one embodiment of the present invention, said at least one printing defect includes a single type of printing defect. In accordance with one embodiment of the present invention, said 20 first training images include a first number of first images of training. In accordance with one embodiment of the present invention, said second training images include a second number of seconds training images representing prints each containing at least one 25 printing defect of the said single type of printing defect. In accordance with one embodiment of the present invention, said first number corresponds to 40%-60% of a total number of images of training used by the computing system for training said neural network model. 30 In accordance with one embodiment of the present invention, said at least one printing defect comprises a plurality of different types of printing defects press. In accordance with one embodiment of the present invention, said first training images include a first number of first images of training. In accordance with one embodiment of the present invention, said 5 second training images include for each type of print defect of said plurality a corresponding second number of second images of training representative of prints each containing at least one defect printing of said type of printing defect. In accordance with one embodiment of the present invention, said 10 first number and said second numbers are substantially equal. In accordance with one embodiment of the present invention, said train, by the computing system, the so-called neural network model further includes including in said first training images digital images corresponding to acquisitions made using a digital camera 15 images of prints free from printing defects. In accordance with one embodiment of the present invention, said train, by the computing system, the so-called neural network model further includes including in said second training images digital images corresponding to acquisitions made using a digital camera 20 images of prints containing at least one print defect. An additional aspect involves a computer program to implement the method for training. A further aspect includes a corresponding program product for computer. 25 A further aspect involves a computational system to implement the method for training. Another aspect of the present invention relates to a method for monitoring printing defects in a substrate printing process. The method includes: acquiring using one or more image acquisition sensors an image of a product of 30 print obtained by a substrate printing process; provide said image acquired by a monitoring computing system; use, by the system of tracking computation, a neural network model trained to classify said acquired image; evaluate, by the computing system of monitoring, the presence of printing defects in the printed product based on a classification of said acquired image performed by said network model neural; provide an output of the said evaluation. 5 In accordance with one embodiment of the present invention, said providing an output of the outcome of said evaluation includes displaying a message in based on the classification of the acquired image. In accordance with one embodiment of the present invention, said providing an outcome of such assessment includes issuing an indicative warning 10 of the presence of printing defects if the classification of the acquired image corresponds to the presence of a printing defect. In accordance with one embodiment of the present invention, said providing an output of the said evaluation includes piloting the process shutdown of printing if the classification of the acquired image corresponds to the presence of 15 a printing defect. An additional aspect involves a computer program to implement the method for monitoring. A further aspect includes a corresponding program product for computer. 20 A further aspect involves a computational system to implement the method for monitoring. One or more aspects of the present invention are set forth in the claims. independent, with advantageous features of the same invention which are indicated in the dependent claims, the wording of which is included here verbatim for 25 reference (with any advantageous feature being provided with reference to a specific aspect of the present that applies mutatis mutandis to any other aspect of the same). Brief description of the drawings 30 These and other features and advantages of the present invention will appear more clearly by reading the following detailed description of embodiments exemplifying and not limiting the same. For its better intelligibility, the The following description should be read with reference to the attached drawings, in which: Figure 1 is a schematic view of a digital printing system in which concepts in accordance with embodiments of the present invention may 5 be applied; Figure 2 shows units included in a computing system of monitoring and in a configuration computing system in accordance with a embodiment of the present invention; Figure 3 illustrates major software components that can be 10 used for a printing defect monitoring method performed by the system monitoring computation in a printing process in accordance with a form of embodiment of the present invention; Figure 4 illustrates major software components that can be used for a training method performed by the computing system 15 configuration for training a neural network used by the system monitoring computation in accordance with one embodiment of the present invention; Figure 5 is a block diagram illustrating the operations performed by the monitoring computing system relating to a monitoring method of 20 printing defects in a printing process according to an embodiment of the present invention; Figures 6A-6D show block diagrams illustrating procedures and operations performed by the configuration computing system relating to a training method of the neural network used by the computing system 25 monitoring in accordance with an embodiment of the present invention; Figure 7 illustrates some images used to experimentally verify the accuracy of a neural network trained using a training method in accordance with embodiments of the present invention. 30 Detailed description of exemplary and non-limiting embodiments of the present invention Referring to the drawings, Figure 1 is a schematic view of a system of 100 digital printing in which concepts in accordance with embodiments hereof invention can be applied. The digital printing system 100 illustrated in Figure 1, and described in 5 detail in the following of this description, is a digital printing system for textile applications configured to perform digital printing on substrates including fabrics. In any case, it is emphasized that the concepts of the present invention can can also be applied to other types of digital printing systems, in which the prints digital are performed on different types of substrates, such as those comprising 10 paper, cardboard, wood, plastic, metal, PVC, film. In accordance with one embodiment of the present invention, the system 100 digital printing is a digital printing system of the so-called “multi-pass” type or “scanner”, where printing is done incrementally through a sequence of printing steps in which at each step a respective portion of a 15 digital image on a corresponding portion of the substrate. The generic step of printing involves a portion of a digital image being printed using one or more multiple movable print heads that print by moving along a first direction in correspondence of a respective portion (usually a strip) of the substrate. A once that portion has been printed, the substrate is translated along a second 20 direction perpendicular to the first direction. The next printing step is then made by printing an additional portion of the digital image in correspondence of a respective further portion of the substrate (typically, adjacent to the portion printed in the previous print step). In accordance with one embodiment of the present invention, the system 25 digital printing press 100 includes a conveyor belt 102 that defines a flat support structure for carrying a substrate 115 (e.g., an element textile) on which to perform digital printing. In accordance with one embodiment of the present invention, the tape conveyor 102 is configured to be moved along a first direction X 30 (using appropriate handling systems not illustrated) so as to allow a controlled translation of the substrate 115 along the first X direction. In accordance with one embodiment of the present invention, the system digital printing 100 comprises a printing module 120 equipped with one or multiple print heads (not shown) capable of printing portions of a digital image on respective portions of the substrate 115 which are gradually found in correspondence of the printing module 120 following the translation of the substrate 115 5 along the first X direction. Without going into implementation details well known to skilled in the art, in accordance with an embodiment of the present invention Each print head comprises a plurality of nozzles (e.g., of diameter of a few tens of microns) electronically controlled for the precise expulsion of drops of ink, and is controlled to translate (back and forth) along a second 10 Y direction perpendicular to the first X direction. In accordance with one embodiment of the present invention, the area of ​​the substrate 115 is conceptually divided into a grid of small elementary cells (e.g. example, square) each of which is potentially destined to receive a respective drop of ink from one of the print head nozzles. 15 In accordance with one embodiment of the present invention, the module 120 print includes a processing system configured to process the digital image to be printed in order to identify the elementary cells to be hit with ink to print the digital image. In accordance with a form of embodiment of the present invention, the printing module 120 comprises a 20 control and actuation system configured to translate the print heads along the second Y direction being a few millimeters above the substrate 115 (without contact the latter directly) and to control the selective dispensing of drops of ink from the nozzles towards selected elementary cells of the substrate in agreement with the processed digital image. In this way, a portion (strip) of the 25 substrate 115 is selectively soaked with the dispensed ink, thus printing a corresponding portion of the digital image. At this point, the substrate 115 is translated by the conveyor belt 102 along the first X direction to allow the printing an additional portion of the digital image onto a new portion of the substrate 115. 30 In accordance with a further embodiment of the present invention, the processing of the digital image to be printed can be done by a system of computing distinct (e.g., remote) from the printing module 120. In this case, the printing module 120 has only the task of moving the print heads and control the selective delivery of ink from the printhead nozzles according to digital image processing. While the 100 digital printing system described is a printing system of the 5 “multi-step” type, it is emphasized that the concepts of the present invention can however, can also be applied to single-pass 100 digital printing systems, in where the digital image is printed onto the substrate in a single pass. In case of color printing, the digital image comprises a plurality of Overlapping color channels. Each color channel is a version of the image. 10 digital including only the color channel's contribution to the image final digital. Each color channel corresponds to one of the primary colors of the print module 120 (for example, using the CMYK color scheme, one channel of cyan color, a magenta color channel, a yellow color channel, and a (black in color). In this case, the overall image is printed by performing multiple 15 successive overlapping prints of each color channel, each made using the corresponding color ink. The 100 digital printing system is a very complex system, which must manage a large number of variables and parameters very precisely. Consequently, printing defects that may affect the print results can 20 can be of very varied types and nature. An illustrative and non-exhaustive list of such defects of print is reported below. Line defect: This printing defect involves the formation of one or more unwanted lines extending along the X-direction. Such lines are due to the failure to deliver ink from one or more nozzles of one or more printheads 25 print, for example because these nozzles are clogged. Color Mismatch Defect: This defect involves inconsistencies on the edges of the shapes represented in the printed matter. These inconsistencies are due to a incorrect overlap of color channels. Stain Defect: This defect involves the presence of unwanted stains, 30 caused by dripping substances (e.g. water or ink) onto the substrate. Blur defect: This defect involves the printed design or at least part of it is blurred, for example due to an inappropriate chemical treatment of the substrate. Incorrect repetition defect: This defect involves that in a drawing including the repetition of copies of the same single image, the position of at least one of the copies is not repeated correctly (for example, with a ratio or 5 a different relative position than the other copies). Pitch defect: This defect, typical of “multi- step”, foresees the presence of overlapping portions of the image or the presence of unprinted lines along the Y direction. This defect can be for example caused by incorrect movement (for example, insufficient extension) 10 or excessive) of the conveyor belt that transports the media along the X direction. In accordance with one embodiment of the present invention, the system digital printing unit 100 includes a monitoring unit 130 configured to monitor the results of the prints made by the 120 printing module by evaluating the presence (or absence) of printing defects in the part of the substrate 115 on which the module 15 of 120 printing performed printing operations. In accordance with one embodiment of the present invention, the unit of monitoring 130 includes one or more imaging sensors 135(i) (e.g. example, cameras) located near the conveyor belt 102 downstream of the printing module 120 so as to acquire images (of portions) of the substrate 115 20 on which the printing module 120 has performed printing operations. In accordance with one embodiment of the present invention, the sensors of image acquisition 135(i) (four, in the figure) are installed on a structure of bridge support 140 which crosses the conveyor belt 102 along the Y direction, allowing each 135(i) image acquisition sensor to frame from above 25 (with respect to a Z direction perpendicular to the X and Y directions) a respective portion 145(i) of the substrate 115 on which the printing module 120 has performed printing operations. The concepts of the present invention can however be directly applied also to cases where the image acquisition sensors 135(i) they are installed in different positions, for example located on the sides of the belt 30 transporter 102. In accordance with one embodiment of the present invention, the unit of monitoring 130 is advantageously equipped with one or more monitoring devices 150 illumination (only one shown in the figure) configured to illuminate the portions 145(i) of the substrate 115 framed by the image acquisition sensors. In accordance with one embodiment of the present invention, the unit of monitoring 130 also includes one or more acquisition control modules 160 5 including control electronics for the image acquisition sensors 135(i). In accordance with one embodiment of the present invention, the unit of monitoring 130 further includes a monitoring computing system 170 configured to evaluate the presence of print defects in the substrate part 115 on which the printing module 120 performed printing operations based on images 10 P(i) acquired by the 135(i) image acquisition sensors. In accordance with one embodiment of the present invention, the system of monitoring computation 170 is configured to use a model of machine learning, and in particular a neural network model, for classify the images P(i) acquired by the image acquisition sensors 135(i) and 15 evaluate the presence of printing defects on the substrate 115 based on this classification. In accordance with one embodiment of the present invention, and as will be described in more detail below, the neural network used by the system monitoring computation 170 is configured to classify each of the images P(i) acquired in a selected from a set of predefined classes C(j). In 20 In accordance with an embodiment of the present invention, said set of predefined classes C(j) includes a class C(0) indicating an absence of defects print and one or more classes C(1), C(2), … each indicative of the presence of a corresponding type of printing defect. A non-limiting example of predefined classes C(j) can be the following: 25 C(0): no printing defects; C(1): presence of at least one line defect; C(2): presence of at least one color misalignment defect; C(3): presence of at least one stain defect; C(4): presence of at least one blur defect; 30 C(5): presence of at least one incorrect repetition defect; C(6): presence of at least one pitch defect. In accordance with one embodiment of the present invention, the system of monitoring computing 170 is configured to report, for example via a respective message displayed via a system display unit of monitoring computation 170, an acoustic signal, and / or the automatic sending of an alarm notification to a remote terminal, the presence of printing defects on the 5 substrate 115 printed if at least one of the images P(i) acquired by the sensors Image acquisition 135(i) has been classified as belonging to one of the classes default C(j) corresponding to a printing defect. In this way, as soon as the monitoring system 170 has identified the presence of at least one print defect since a P(i) image has been classified 10 as belonging to one of the predefined classes C(j) corresponding to a defect of printing, it is possible to act promptly, for example by stopping the printing system digital 100, thus avoiding further waste of materials (substrate and inks) and electricity. Once the 100 digital printing system has been stopped, it is it is therefore possible to carry out an inspection of the components of the digital printing system 15 100 and intervene to fix what caused the printing defect. Identifying the cause of the printing defect is advantageously done taking into account the specific class (and therefore the specific printing defect) in which image P(i) has been classified. In accordance with one embodiment of the present invention, the same monitoring computing system 170 can be 20 configured to automatically shut down the digital printing system 100 following of the identification of at least one printing defect. In accordance with one embodiment of the present invention, the configuration of the neural network used by the computing system monitoring 170 is determined by corresponding CD configuration data which are 25 states generated by a 180-configuration computing system. In accordance with an embodiment of the present invention, the computing system of configuration 180 is distinct from monitoring computing system 170. Referring to Figure 2, each of the computing systems of monitoring 170 and configuration 180 includes several units that are connected 30 between them through a bus structure 210. In particular, a microprocessor 220, or more, provides a logical capability of computing systems monitoring / configuration 170, 180. A non-volatile memory (ROM) 230 stores basic code for bootstrapping monitoring computing systems / configuration 170, 180 and a volatile memory (RAM) 240 is used as memory working from the 220 microprocessor. The computing system of monitoring / configuration 170, 180 is equipped with a mass memory 250 5 for storing programs and data, for example, a solid-state drive (SSD). Additionally, the monitoring / configuration computing system 170, 180 It includes 260 controllers for peripheral units, or Input / Output units, such as keyboards, display devices, network adapters, drives for Read / write removable data storage devices. The monitoring system 10 170 is also equipped with appropriate control units for the control unit modules monitoring 130. Referring to Figure 3, software components are shown main ones that can be used for a defect monitoring method print performed by the monitoring computing system 170 in a process of 15 printing in accordance with an embodiment of the present invention. All software components (programs and data) are identified in the complex with reference 300. The software components 300 are typically stored in the mass memory of the monitoring computing system 170 and loaded (at least in part) in the working memory of the monitoring computing system 170 20 when programs are running, along with an operating system and other application programs not directly relevant to the solution of this disclosure (and therefore not represented in the figure for simplicity and clarity). programs are initially installed in the mass storage, for example by removable storage devices or from the network. Each program can be either 25 comprise a module, segment or portion of code, which comprises one or more executable instructions to implement the specified logical function. In accordance with one embodiment of the present invention, a image acquirer 305 drives acquisition control modules 160 of the imaging unit 130 monitoring (see Figure 1) to control the acquisition of P(i) images 30 of the substrate 115 by the image acquisition sensors 135(i) during the printing operations. In accordance with one embodiment of the present invention, the image acquirer 305 saves an image in a memory variable 310 P(i) (defined by a matrix of pixels) which is acquired during the operations of press. In accordance with a further embodiment of the present invention 5 (not shown), the imager 305 may additionally have write access to a database of acquired images to store in that database one or more of the (for example example, more recent) acquired P(i) images. In accordance with a form of embodiment of the present invention, the acquired image database 310 has a entry for each acquired P(i) image. 10 In accordance with one embodiment of the present invention, a machine learning model is used to classify the images P(i) into corresponding predefined classes C(j) by applying techniques of machine learning. In short, machine learning is used to perform a specific task (in this case, image classification) without 15 using explicit instructions but automatically deducing how to do it from examples (exploiting a corresponding model that has been learned from them). In particular The implementation in question applies a deep learning technique, which is a branch of machine learning based on neural networks. According to a form In carrying out the present invention, the machine learning model is 20 a neural network 320. The 320 neural network is a data processing system that comes close to the functioning of the human brain. The 320 neural network includes elements of basic processing (neurons), which perform operations based on corresponding weights. Neurons are connected via one-way channels (synapses), which transfer data 25 among them. Neurons are organized into layers that perform different operations. In accordance with one embodiment of the present invention, the network neural network 320 includes at least one input layer for receiving network input neural network 320 in the form of data representing an acquired image P(i). The network neural 320 further includes an output layer for providing the output of the neural network 30 320 in the form of data representing a class selected from the predefined classes C(j). In accordance with one embodiment of the present invention, the output of the neural network 320 is a CA classification vector that provides an indication of a class selected among the predefined classes C(j). For example the vector of CA classification provides for each of the predefined classes C(j) a corresponding class probability value indicative of the probability that the image P(i) represents a reproduction of a portion of the substrate free from printing defects (for 5 class C(0)) or including at least one printing defect of a specific type (for the further classes C(1), C(2), …). In accordance with one embodiment of the present invention, the network neural 320 is a convolutional neural network, that is, a type of neural network comprising one or more convolutional layers that perform (cross) operations 10 convolution. Each convolutional layer performs a convolution operation. (in sequence, on portions of the data) through a convolution matrix (called filter or kernel) defined by corresponding weights. The convolution operation is in typically followed by further operations, including for example a normalization to adjust the mean and variance of the data and the application of a function 15 activation for the introduction of a nonlinearity factor. In the case considered, the weights can for example represent a particular visual feature to to search. In accordance with one embodiment of the present invention, in the network neural 320 one or more of the convolutional layers may be followed by a 20 corresponding max-pooling layer configured to perform a sub-procedure sampling in order to allow a certain degree of translation invariance and to reduce the computational load for subsequent layers. In accordance with one embodiment of the present invention, the network neural 320 further comprises final layers of the “fully connected” type, i.e. 25 non-convolutional layers where each output value of a layer's output is function of all the input values ​​of that layer's input. These final layers act as final classifiers having a number of output neurons equal to the number of possible predefined classes C(j), so that each output neuron is associated with a specific one among the predefined classes C(j). 30 In accordance with one embodiment of the present invention, the network neural 320 is configured by reading access to a 330 configuration database containing the CD configuration data generated by the computing system configuration 180. This CD configuration data defines one or more neural network configurations 320. The configuration database 330 has an entry for each configuration of the neural network 320. The generic entry stores a corresponding configuration defined by a corresponding specific set of 5 values ​​of neural network weights 320. In accordance with one embodiment of the present invention, the network neural 320 (configured via CD configuration data stored in the database of configuration 330) accesses the memory variable 310 (or the image database acquired, if present) to recover the acquired image P(i) (or, the image 10 most recently acquired in the acquired image database, if any), ranking the acquired image P(i) generating a corresponding classification vector CA and save the CA classification vector in a memory variable 340. In accordance with a further embodiment of the present invention (not shown), the neural network 320 may additionally have write access to a database of 15 classification to store the CA classification vector in this database. In accordance with one embodiment of the present invention, the database of classification has one entry for each acquired image P(i) stored in the database of acquired images. For example, each entry in the database of classification 340 stores a link to an entry in the acquired image database 20 310 where an acquired image P(i) and a corresponding vector of CA classification identifier of the class in which this acquired image P(i) was assigned by the 320 neural network. In accordance with one embodiment of the present invention, the software components 300 may further include a viewer 350 that accesses 25 the memory variable 340 (or the classification database, if present), and which drives a display of the monitoring computing system 170 for displaying a message based on the class that has been assigned to the acquired image P(i) retrieving the CA classification vector from memory variable 340 (or from the classification database, if any). For example, if the acquired image P(i) is 30 was classified by the neural network 320 into one of the corresponding classes C(1), C(2), … in case of a printing defect, the display 350 can drive the display to show a message indicating the presence of printing defects on the printed substrate. Advantageously, the display 350 can also indicate in the message the type specific printing defect detected, identified on the basis of the specific class selected. If the acquired image P(i) has been classified into class C(0), and then the printed substrate is assessed to be free of print defects, the 350 viewer 5 can drive the display to show a message indicating the absence of defects of printing on the printed substrate, or alternatively to not display any message. In accordance with one embodiment of the present invention, the software components 300 may further include a 360 alert generator that 10 accesses memory variable 340 (or the classification database, if present), and that drives one or more peripheral units of the monitoring computing system 170 for issue a warning indicating the presence of printing defects if the class that has been assigned to the acquired image P(i) – identified by reading the vector of CA classification retrieved from memory variable 340 (or from the database) 15 classification, if present) - is one of the classes C(1), C(2), … corresponding to a printing defect. For example, the 360 ​​alert generator can drive a device for generating an acoustic warning signal, or can drive a device communication to send a remote alert message. In accordance with one embodiment of the present invention, the 20 software components 300 may further include a print controller 370 that accesses memory variable 340 (or the classification database, if present), and that automatically controls the stop of the conveyor belt 102 and the printing module 120 if the class that was assigned to the acquired image P(i) – identified reading the CA classification vector retrieved from memory variable 340 (or 25 from the classification database, if any) - is one of the classes C(1), C(2), … corresponding to a printing defect. Referring to Figure 4, software components are shown main ones that can be used for a training method performed by the 180-configuration computing system for neural network training 30 320 used by the monitoring computing system 170 and the generation of corresponding CD configuration data for such neural network 320 in accordance with a embodiment of the present invention. All software components (programs and data) are identified in the complex with reference 400. The software components 400 are typically stored in the mass storage of the computing system of configuration 180 and loaded (at least in part) in the working memory of the computing system 5 configuration 180 when programs are running, along with a system operating and other application programs not directly relevant to the solution of this disclosure (and therefore not represented in the figure for simplicity and clarity). Programs are initially installed in the mass memory, e.g. for example, via removable storage drives or from the network. Each 10 program may be or comprise a module, segment or portion of code, which includes one or more executable instructions to implement the logical function specified. In accordance with one embodiment of the present invention, a training database 405 stores AP(k) training images for 15 training the neural network 320. In accordance with one embodiment of the present invention, the training database 405 has an entry for each AP(k) training image. The generic training database entry 405 stores a training image AP(k), defined by a matrix of pixel, and an indication of the class C(j) to which that image belongs. 20 In accordance with one embodiment of the present invention, an engine training 410 trains a copy of the neural network used by the system monitoring 170, identified with the same numerical reference 320. During a neural network training procedure 320, training engine 410 read access to the training database 405 to retrieve the images of 25 AP(k) training and generate corresponding CD configuration data for the network neural 320 (in particular, comprising one or more specific sets of weight values of the neural network 320). The training engine 410 writes to a copy of the configuration database used by the monitoring system 170, identified with the same numerical reference 330, to store the CD configuration data 30 generated. In accordance with one embodiment of the present invention, the AP(k) training images stored in training database 405 and used by training engine 410 to train neural network 320 include AP(k) training images of two classes, generated in two ways different, and in particular: - non-artifactual AP(k) training images (hereinafter referred to as 5 as APN(k)), obtained through direct acquisitions (e.g. through camera) of digital images of substrate print results; - artifactual AP(k) training images (hereinafter referred to as as APA(k)), obtained from digital images of substrates obtained through acquisitions (e.g. via camera) and images 10 digital drawings to print (for example obtained by creating the drawings using graphic editors or camera captures) by applying appropriate digital combination and / or manipulation operations. In accordance with one embodiment of the present invention, a Image Acquisition 415 controls the acquisition of training images 15 non-artifactual APN(k) and write access to the training database 405 for store the APN(k) non-artifactual training images in this database acquired. In accordance with one embodiment of the present invention, a 420 image acquirer controls the acquisition of digital images of substrate 20 PS(l) each obtained from an acquisition (for example through camera) of one or more portions of a respective substrate (e.g., a particular type of fabric, a particular type of paper, ...). In accordance with one embodiment of the present invention, The 420 image acquirer writes to a database of digital images. 25 substrate 425 to store in this database the digital images of substrate PS(l) acquired. In accordance with one embodiment of the present invention, the Substrate 425 digital image database has an entry for each image digital substrate PS(l), which is defined by a matrix of pixels. In accordance with one embodiment of the present invention, a 30 image collector 430 collects digital images of drawings to be printed PD(n) each obtained by creating (and processing) the drawings using graphic editors or by camera captures. In accordance with one embodiment of the present invention, the image collector 430 write accesses a database of digital images drawings to print 435 to store in this database the digital images of drawings to be printed PD(n) acquired. In accordance with one embodiment of the 5 present invention, the database of digital images of printable designs 435 has a entry for each digital drawing image to be printed PD(n), which is defined by a matrix of pixels. In accordance with one embodiment of the present invention, and as will be described in more detail below, an image generator 10 training artifacts 440 read access databases 425 and 435 to recover digital images of PS(l) substrate and digital images of PD(n) printable designs and generate APA(k) artifactual training images each obtained by combining a respective pair of digital image of PS(l) substrate and digital image of drawing to be printed PD(n) and possibly applying manipulation operations 15 digital. In accordance with one embodiment of the present invention, the 440 artifact training image generator write access to the database training 405 to store the training images in this database APA(k) artifacts generated. 20 Figure 5 is a block diagram illustrating the operations performed by the 170 monitoring computing system relating to a monitoring method of printing defects in a printing process in accordance with an embodiment of the present invention. Each block may correspond to one or more instructions or executable procedures to implement one or more specific logical functions on 25 components of the 170 monitoring computing system illustrated in Figure 3. The method starts when the digital printing system 100 is activated and the module of printing 120 begins to print on the substrate 115. In accordance with a form of embodiment of the present invention, the image acquirer 305 begins to acquire an image P(i) of the printed substrate (block 505), saving such 30 acquired image P(i) in memory variable 310. Optionally, the image acquired P(i) can also be stored in the acquired images database, if present (block 508). In accordance with one embodiment of the present invention, the network neural 320, configured with the weights specified in the CD configuration data stored in the configuration database 330, classifies the acquired image P(i) in a class selected among the predefined classes C(j) generating a corresponding 5 CA classification vector (block 509) and saving that CA classification vector in memory variable 340. Optionally, the CA classification vector can also be stored in the classification database, if present (block 510). In accordance with one embodiment of the present invention, the viewer 350 accesses memory variable 340 (or the classification database, 10 if present) to retrieve the CA classification vector corresponding to the acquired image P(i) and drive a display of the computing system monitoring 170 to display a message based on the class it was assigned to the acquired image P(i) (block 520), for example a message that specifies the class C(j) into which the acquired image P(i) has been classified. 15 In accordance with one embodiment of the present invention, the alert generator 360 accesses memory variable 340 (or the database of classification, if present) to retrieve the CA classification vector corresponding to the acquired image P(i) and drive one or more peripheral units of the 170 monitoring computing system to issue an indicative warning of the 20 presence of printing defects if the class that was assigned to the acquired image P(i) – identified by reading the CA classification vector retrieved from the variable of memory 340 (or from the classification database, if present) - is one of the C(1) classes, C(2), … corresponding to the presence of a printing defect (block 525). In accordance with one embodiment of the present invention, the 25 print controller 370 accesses memory variable 340 (or the database of classification, if present) to retrieve the CA classification vector corresponding to the acquired image P(i) and control the conveyor belt stop 102 and the print module 120 if the class that has been assigned to the image acquired P(i) – identified by reading the CA classification vector retrieved from the 30 memory variable 340 (or from the classification database, if present) - is one of the classes C(1), C(2), … corresponding to a printing defect (block 530). In accordance with one embodiment of the present invention, the flow of operations then returns to block 505, with the acquisition of a new image P(i) of the printed substrate. It is emphasized that the concepts of the present invention also apply in case where only a subset of the group comprising the 350 viewer, the 5 alert generator 360 and print controller 370 access the variable memory 340 to perform the respective operations 520, 525 and 530 (e.g., only the viewer 350). Figures 6A-6D show block diagrams illustrating procedures and operations performed by the configuration computing system 180 relating to a 10 neural network training method 320 (i.e. a configuration method of the network by setting its weights) to optimize the network capacity neural 320 to classify the acquired images P(i) according to a form of embodiment of the present invention. Each block may correspond to one or multiple statements or procedures executable to implement one or more logical functions 15 specifications on 180-configuration computing system components illustrated in Figure 4. Referring to Figure 6A, the network training method neural 320 in accordance with an embodiment of the present invention comprises a procedure 602 directed to the generation of training images 20 AP(k), followed by a 604 procedure aimed at optimizing the network weights neural 320 (and hence to the generation of corresponding CD configuration data), where this optimization is performed by exploiting the training images AP(k) created in the first procedure. In accordance with one embodiment of the present invention, the 25 procedure 602 provides for the generation of non-artifactual training images APN(k) (subprocedure 608), the generation of artifactual training images APA(k) (sub-procedure 609), and the storage of training images APN(k), APA(k) in training database 405 (block 610). The subprocedure 608 is performed by the image acquirer 415, while the sub-procedure 609 is 30 performed by the image acquirer 420 and the image collector 430 and the 440 artifactual training image generator of the computing system configuration 180. In accordance with a further embodiment of the present invention (not shown), sub-procedure 608 may not be executed. In this case, the AP(k) training images consist only of training images APA(k) artifacts. In other words, the concepts of the present invention also apply 5 to the case where the neural network 320 is trained using only images of artificial training. In accordance with one embodiment of the present invention, the procedure 604 is executed by the training engine 410 of the computing system of configuration 180 by setting the weights of the neural network 320 in the following way. 10 The training engine 410 reads the training database 405 to retrieve a stored AP(k) training image (which can be an APN(k) non-artifactual training image (if present) or a APA(k) artifact training and the corresponding indication of the class C(j) of such image (block 612). In this regard, in accordance with one embodiment 15 exemplifying and not limiting the present invention, such indication of the class C(j) is in the form of a classification vector CT having the probability value of class corresponding to the actual class C(j) of the image AP(k) which is equal to 1, and the probability values ​​corresponding to the other classes C(j) which are equal to 0. The training engine 410 provides the training image AP(k) 20 retrieved to the neural network 320, configured with weights derived from a current version of CD configuration data stored in the 330 configuration database (see Figure 4), and this image is classified by the neural network 320 with the generation of a corresponding CA classification vector (block 613). The training engine 410 compares (e.g., by subtraction) 25 the CA classification vector (image classification identifier training AP(k) performed by the neural network 320 in training) with the vector of CT classification (indicative of the actual class of the training image AP(k)) and calculates (e.g., using an error function) a corresponding ER error value that quantifies the classification error committed by the network 30 neural 320 (block 614). The training engine 410 then updates the weights of the neural network 320 in based on the calculated ER error value (e.g., by a step-down method) gradient), thus modifying the CD configuration data stored in the training database 405 (block 615). The sequence of operations corresponding to blocks 612 – 615 is therefore repeated several times, selecting 405 from the training database each time 5 different AP(k) training images. In accordance with the embodiment of the invention just described, the weights of the neural network 320 are updated every time a new image training AP(k) was classified by the 320 neural network. However, the concepts of the present invention also apply to cases where the weights update 10 executed at block 615 is executed only after the operations corresponding to the blocks 612 – 614 are performed for a plurality of training images AP(k) (e.g. example, for all training images AP(k) contained in the database training 405 or for a portion of them). In accordance with one embodiment of the present invention, for the purpose 15 to ensure sufficient generalization of the neural network 320, i.e. to ensure correct functioning of the trained 320 neural network classification of generic P(i) images different from training images AP(k) used in training, in addition to procedure 604, a Generalization testing procedure (not illustrated). For example, the images 20 AP(k) training can be divided into two groups, and procedure 604 can be performed using only the AP(k) training images from the first group. The training images AP(k) of the second group can then be used from the verification procedure to verify the level of generalization achieved by the trained neural network 320. If the verification has not given a positive result, that is, if 25 the neural network 320 was not found to be sufficiently generalized, it is It is appropriate to repeat the training procedure 604 using different conditions. It should be emphasized that the procedure 604 described is only one example of how the weights of the 320 neural network can be optimized by exploiting the images AP(k) training system, and that the concepts of the present invention may be applied 30 even using different procedures. Sub-procedure 608 in accordance with one embodiment of the present invention for the generation of non-artifactual training images APN(k) is illustrated in Figure 6B. The first phase of sub-procedure 608 in accordance with a form of realization of the present invention involves the acquisition by of the image acquirer 415 of digital images of substrate print results 5 (block 620). These images are obtained by direct acquisitions via imaging sensors (e.g., cameras) of (portions of) substrates that have been subjected to digital printing operations. For example, such images are acquired immediately after the execution of the printing operations by means of image acquisition sensors placed downstream of the printing module that performed the 10 digital printing, such as the acquisition sensors 135(i) illustrated in Figure 1. In accordance with one embodiment of the present invention, the digital images acquired by image capture device 415 are selected manually, that is with the intervention of one or more operators, to ensure that the a sufficient variety of cases, involving different types of substrate and different 15 images to print, and considering both print results free of printing defects either print results including a sufficient variety of print defects different. In accordance with one embodiment of the present invention, the APN(k) non-artifactual training images are generated by the image acquisition system. 20 images 415 applying (block 622) processing algorithms on each of the digital images acquired to adapt the format (size, resolution) to a standard format chosen for all training images to be stored in the training database, and to introduce variations (e.g., rotations and duplications) aimed at making the training database larger. 25 In accordance with one embodiment of the present invention, the image acquirer 415 performs (block 624), under the manual control of one or more operators, a classification of each of the training images does not artifacts APN(k) into a corresponding one of the predefined classes C(j). To be executed, sub-procedure 608 requires an expenditure of time not 30 indifferent, not only because of the manual classification of images by operators, but also due to the difficulty in finding authentic images of real results of prints on substrates containing specific printing defects. Consequently, in accordance with one embodiment of the present invention, the images non-artifactual training APN(k) generated with sub-procedure 608 are advantageously only a smaller portion of the total imagery training AP(k) stored in the training database 405. For example, 5 APN(k) non-artifactual training images can only represent 10%, 5%, or less (as previously reported, the training images did not APN(k) artifacts may not even be present at all) of the totality of images AP(k) training values ​​stored in training database 405. Sub-procedure 609 in accordance with one embodiment of the 10 present invention for the generation of APA(k) artifactual training images is illustrated in Figure 6C. In accordance with one embodiment of the present invention, the sub- procedure 609 provides for the generation of a collection of digital images of PS(l) substrate (which are stored in the substrate digital image database 15 425), generating a collection of digital images of drawings to be printed PD(n) (which are stored in the database of digital images of drawings to be printed 435), and the generation of the APA(k) artifactual training images from a combination of the PS(l) and PD(n) images. In accordance with one embodiment of the present invention, the 20 generation of digital images of PS(l) substrate involves the following sequence of operations, globally identified with the reference 626. In accordance with one embodiment of the present invention, Image acquirer 420 acquires (block 627) an image of a substrate (for example, a type of fabric, a type of wood panel, a type of paper, a type of 25 glass…), for example obtained using a camera. In accordance with one embodiment of the present invention, The imager 420 generates a digital image of PS(l) substrate from from the substrate image acquired by applying algorithms to the latter processing to apply noise and / or apply brightness and contrast variations 30 to increase visibility (block 628). In accordance with one embodiment of the present invention, The 420 image acquirer writes to the digital image database of substrate 425 for saving the generated PS(l) substrate digital image (block 629). The sequence of operations 626 is then repeated to generate a plurality of digital images of PS(l) substrate. In accordance with one embodiment of the present invention, the 5 generation of digital images of drawings to be printed PD(n) provides the following sequence of operations, globally identified with the reference 630. In accordance with one embodiment of the present invention, the image collector 430 collects a digital image to print, representing a sample of a design to be printed, i.e. to be applied by means of 10 digital printing, on a substrate (block 631). The digital image to be printed may include drawings generated using graphics editors or acquisitions from camera, and generally have any size and resolution. In accordance with one embodiment of the present invention, the digital image to be printed is subjected by the image collector 430 to 15 a scaling operation (block 633) aimed at scaling the image to a resolution compatible with the resolution used during printing operations digital. For example, given a dimension Dp =Dp(x)·Dp(y) of the area occupied by the print result on the substrate, where Dp(x) is the dimension (e.g., in inches) of that area along the X-direction, and Dp(y) is the size (e.g., in inches) of 20 such area along the Y direction, the digital image to be printed is scaled to a resolution corresponding to the print resolution R(print) used by the module print 120. The print resolution R(print) can for example be between 150 and 1440 dpi. For example, given an area of ​​the print result on a substrate having a dimension Dp(x) equal to 5 inches and a dimension Dp(y) equal to 50 inches, and a 25 print resolution R(print) equal to 600 dpi, the digital image to be printed is scaled to have a size of Dp(x) R(print) · Dp(y) R(print) = 3000 pixels · 30000 pixels. In “multi-pass” or “scanner” digital printing systems such as the digital printing 100 illustrated in Figure 1, the printing of digital images is done 30 incrementally where at each “step” a portion of the substrate 115 having a rectangular shape having a generally length (long (the Y direction) is significantly greater than the height (along the X direction). D i consequently, to acquire an image of the printed substrate portion in a step it is necessary to use a plurality of images P(i) by means of a plurality of 135(i) image acquisition sensors (aligned along the Y direction). The subdivision into a plurality of smaller P(i) images also allows to reduce the 5 amount of data to be processed, consequently reducing the computational load. For this reason, in accordance with one embodiment of this invention, the 430 image collector divides the digital image to be printed scaled into portions (tiles) by means of a “tessellation” operation (in English, “tiling”) (block 635). The size (in pixels) of the single tile Dt = Dt(x) Dt(y) 10 depends on the size (in pixels) A = A(x) A(y) of the training images APA(k) with which you want to train the neural network, and the ratio between the resolution of R(print) print used by the 120 printing module and the R(camera) resolution of the sensors of image acquisition 135(i), for example by means of the relation: ( ) ( ) = ( ) ( ) ( ) ( ) = ( ) ( ) The size A (in pixels) of the artifactual training images APA(k) with which you want to train the neural network can for example correspond to (or more in in general, depend on the) size of the images P(i) acquired by the sensors 20 image acquisition 135(i) in order to train the network neural using training images having a corresponding size to the size of the images that the neural network will actually have to classify while monitoring printing operations. It is emphasized, however, that the concepts of the present invention can however also be applied to cases in 25 which the size (in pixels) of the APA(k) artifact training images with which you want to train the neural network (and, therefore, the tile) is different from the size (in pixels) of images P(i) acquired by the image acquisition sensors 135(i). The resolution R(camera) of 135(i) image acquisition sensors used in the above report is expressed in dpi and is a function of the actual 30 resolution (in pixels) of the image acquisition sensors 135(i) and of the position relative (e.g., distance) between the image acquisition sensors 135(i) and the framed target. For example, given a print resolution R(print) equal to 600 dpi and a R(camera) resolution of the image acquisition sensors equal to 150 dpi, and 5 choosing to train the neural network with images having a dimension A(x) A(y) equal to 480 pixels by 480 pixels, The size (in pixels) of the single tile Dt would be 1920 pixels by 1920 pixels. In accordance with one embodiment of the present invention, the Image Collector 430 generates one or more digital images of drawings to be printed 10 PD(n) starting from each newly created tile by applying to each tile processing algorithms aimed at performing one or more of the following operations: rotation, flipping, color tone changing and similar operations (block 637). The application of such processing algorithms is aimed at enriching the diversity of PD(n) images that will be used to generate the artifactual training images 15 APA(k). In accordance with one embodiment of the present invention, the image collector 430 write access to the digital image database drawings to print 435 to save one or more digital images of drawings to print print generated PD(n) (block 638). 20 The sequence of operations 630 is then repeated to generate a plurality of digital images of drawings to be printed PD(n) using a plurality of scanned digital images to print. In accordance with one embodiment of the present invention, the generation of APA(k) artifactual training images from a 25 combination of the images PS(l) and PD(n) provides the following sequence of operations, globally identified with the reference 650. In accordance with one embodiment of the present invention, the 440 artifact training image generator read access to the database digital images of drawings to print 435 to select one of the images 30 digital drawings to print PD(n) (block 652). In accordance with one embodiment of the present invention the Artifact training image generator 440 splits (block 653) the digital drawing image to be printed PD(n) selected in multiple color channels (for example, four color channels in a CMYK color scheme). Each channel color includes a grayscale version of the digital drawing image to print PD(n) which reflects the contribution of the color channel to the digital image 5 of drawing to print PD(n). In accordance with one embodiment of the present invention, the artifact training image generator 440 processes (block 655) each color channel of the digital drawing image to be printed PD(n) using a dithering algorithm for creating color depth. Examples not 10 comprehensive examples of such algorithms include the Floyd-Steinberg algorithm, the Ordered algorithm, the Jarvis algorithm and the precomputed matrix algorithm. In accordance with one embodiment of the present invention, the Artifact Training Image Generator 440 generates a set of images APA(k) artifact training images including both training images 15 identifiers of print results free from printing defects – and therefore belonging to the class C(0) – both training images identifying print results containing printing defects – and therefore belonging to the other classes C(1), C(2), ... In accordance with one embodiment of the present invention, the subdivision of the artifactual training images APA(k) is uniform, with a 20 number of artifactual training images APA(k) which is essentially equal for each of the predefined classes C(j). For example, to train the neural network to identify a single type of printing defect corresponding to the predefined class C(1), the APA(k) artifactual training images generated by the generator Artifactual training images 440 may comprise approximately 50% (e.g. 25 example, 40-60%) of APA(k) artifactual training images free of defects print (and therefore corresponding to class C(0)) and approximately 50% (for example, 60-40%) of APA(k) artifact training images containing this type of printing defect (and therefore corresponding to the class C(1)). As another example, to train the network neural to identify three types of printing defects corresponding to predefined classes 30 C(1), C(2), C(3), the APA(k) artifactual training images generated by the Artifactual training image generator 440 may include a 25% approximately of APA(k) artifactual training images free of printing defects (and hence corresponding to the class C(0)), about 25% of the training images are artifactual APA(k) containing a first type of printing defect (corresponding to class C(1)), approximately 25% of APA(k) artifact training images containing a second type of printing defect (corresponding to class C(2)), and approximately 25% of images 5 APA(k) training artifacts containing a first type of print defect (corresponding to class C(3)). The concepts of the present invention can However, apply to cases where the division of the training images artifacts APA(k) is not uniform, i.e. with a number of images APA(k) artifact training that is not equal for each of the predefined classes 10 C(j). In view of this, in accordance with one embodiment of this invention, the artifact training image generator 440 determines whether the digital drawing image to be printed PD(n) is to be used to generate a APA(k) artifact training image containing a print defect or not 15 (decision block 657). In accordance with one embodiment of the present invention, if the digital drawing image to be printed PD(n) is to be used to generate a APA(k) artifact training image containing a printing defect, the 440 digitally manipulates artifactual training image generator (block 20 660) one or more of the color channels of the digital drawing image to be printed PD(n) through an appropriate digital manipulation procedure so as to introduce a specific printing defect into it. In accordance with a form of realization of the present invention, if instead the digital image of drawing from print PD(n) is to be used to generate an artifactual training image 25 APA(k) free from printing defects, this digital manipulation procedure is not performed. At this point, in accordance with an embodiment of this invention, the 440 artifact training image generator combines (block 662) color channels of the digital drawing image to be printed PD(n) 30 (possibly digitally manipulated to include a printing defect) for recompose the digital image of the drawing to be printed PD(n) in color. In accordance with In one embodiment of the present invention, the recombination is performed using one of the known channel mixing techniques. In the case of prints in black and white, this step can be skipped. In accordance with one embodiment hereof, the generator of 440 artifact training images apply to digital drawing image 5 print PD(n) (possibly digitally manipulated to include a defect in printing) additional processing algorithms aimed at varying brightness and contrast (block 664). Unlike the case where the algorithms of processing for the generation of digital images of drawings to be printed PD(n) (block 637), it is important that the processing algorithms used in this phase do not 10 provide for rotations, scaling or cropping in order not to lose information relating to to any printing defect that has been added or to avoid creating images containing printing defects that are not possible in reality. In accordance with one embodiment of the present invention, the 440 artifact training image generator read access to the database 15 digital images of substrate 425 to select (block 666) a specific image PS(l) substrate digital based on the type of digital printing you want to monitor (printing on a substrate comprising a particular fabric, printing on a substrate of a particular type of paper…). In accordance with one embodiment of the present invention, the 20 artifact training image generator 440 combines, for example, by overlay, the digital drawing image to be printed PD(n) (possibly digitally manipulated to include a printing defect) with the digital image of the selected PS(l) substrate (block 668). In this way, the resulting image from the combination (superposition) of the two images comes advantageously 25 rendered similar to an image that would have been obtained by acquiring (for example by camera) a portion of substrate on which a printing operation has been performed digital to print that specific digital drawing image to be printed PD(n). In accordance with one embodiment of the present invention, the Artifact training image generator 440 generates a corresponding 30 APA(k) artifact training image by scaling (block 669) the image resulting from the combination of the images PD(n), PS(l) in order to bring it to the size (in pixels) A = A(x) A(y), for example corresponding to the resolution of the 135(i) image acquisition sensors, so as to make such image compatible with the size of the images P(i) acquired by the image acquisition sensors 135(i). In accordance with one embodiment of the present invention, the 5 artifact training image generator 440 classification (block 670) automatically generates the artifactual training image APA(k) based on of the operations performed previously: - if the digital manipulation corresponding to block 660 has not been performed, and therefore no printing defect was introduced, the image of 10 APA(k) artifact training is classified in class C(0) corresponding to the absence of printing defects; - if the digital manipulation corresponding to block 660 has been performed to introduce a particular printing defect corresponding to a specify default class C(j), the artifact training image 15 APA(k) is classified in class C(j) corresponding to the presence of a printing defect of that type. In accordance with one embodiment of the present invention, the sequence of operations 650 is then repeated to generate further images of APA(k) artifact training, using a different digital image of 20 drawing to be printed PD(n), and / or introducing via digital manipulation a different type of printing defect, and / or combining the digital image of the drawing from print PD(n) with a different digital image of PS(l) substrate. The digital manipulation procedure performed at block 660 for the introduction of a printing defect into a digital drawing image to be printed 25 PD(n) (or color channel thereof) in accordance with one embodiment of the present invention is illustrated in Figure 6D. In accordance with one embodiment of the present invention, the artifact training image generator 440 selects (block 680) a type of printing defect to be introduced into the digital image of the drawing to be printed 30 PD(n) among the classifiable printing defects (i.e. among the printing defects identified in a of the predefined classes C(1), C(2), …). The choice of the particular type of defect of print to be introduced may depend on the type of 100 digital printing system you are using wants to monitor (for example, from the typology / technology of the 120 printing module), and / or the type of digital printing performed. At this point, in accordance with an embodiment of this invention, the 440 manipulates artifactual training image generator 5 the digital drawing image to be printed PD(n) (or only one of the color channels corresponding) to introduce the printing defect of the selected type (block 682). In accordance with one embodiment of the present invention, in order to introduce a line fault (corresponding to the default class C(1)) in the digital drawing image to print PD(n), the image generator 10 training artifacts 440 manipulates one or more color channels of the image digital drawing to print PD(n) to remove from each of them a corresponding set of pixels, for example a line along a direction corresponding to the X direction. In this way, when the image is recomposed by overlapping the color channels, a line is created in which a 15 color component is absent. In accordance with one embodiment of the present invention, in order to introduce a color misalignment defect (corresponding to the default class C(2)) in the digital drawing image to be printed PD(n), the image generator training artifacts 440 manipulates one or more color channels of the image 20 digital drawing to print PD(n) to translate them (by a quantity of pixels) with respect to to the other color channels. This way, when the image is recomposed By combining the color channels, an inconsistent effect is created on the edges of the shapes represented in the image due to the altered overlapping of the channels color. 25 In accordance with an embodiment of the present invention, in order to introduce a stain defect (corresponding to the default class C(3)) in the digital drawing image to print PD(n), the image generator training artifacts 440 manipulates one or more color channels of the image digital drawing to be printed PD(n) by superimposing one or more images on them 30 representing one or more spots. In accordance with one embodiment of the present invention, in order to introduce a blur defect (corresponding to the default class C(4)) in the digital drawing image to print PD(n), the image generator training artifacts 440 manipulates one or more color channels of the image digital drawing to be printed PD(n) using a blurring algorithm. In accordance with one embodiment of the present invention, in order to 5 introduce a wrong repetition fault (corresponding to the default class C(5)) in the digital image of the drawing to be printed PD(n) when such image includes the repetition of copies of the same single image, the generator of artifact training images 440 manipulates one or more color channels of the digital image of drawing to be printed PD(n) to alter the position of at least 10 one of such copies relative to the position of the other copies. In accordance with one embodiment of the present invention, in order to introduce a step defect (corresponding to the default class C(6)) in the digital drawing image to print PD(n), the image generator training artifacts 440 manipulates one or more color channels of the image 15 digital drawing to print PD(n) to divide it into a first portion and a second portion, and move the first portion so that it partially overlaps to the second portion or in such a way as to create a separation between the two portions. This particular type of printing defect is introduced only if the printing system digital 100 to monitor includes a 120 printing module that works in 20 “multi-step” modes. In accordance with one embodiment of the present invention, after having manipulated the digital image of drawing to be printed PD(n), the generator of artifactual training images 440 performs a validation process (block 684), specific for each type of printing defect that can be introduced, 25 to check whether this printing defect is actually visible or not. In this In this way, the identification of false positives due to a classification into one of the classes C(1), C(2), … of an image in which the defect print is not actually visible. In view of this, in accordance with one embodiment of this 30 invention, the artifact training image generator 440 determines whether the manipulated digital drawing image to be printed PD(n) is validated or not (decision block 686). In accordance with one embodiment of the present invention, if the artifact training image generator 440 determined that the image digital drawing to be printed PD(n) manipulated cannot be validated, the artifact training image generator 440 discard (block 688) such 5 digital drawing image to print PD(n) manipulated, and try at least one more aimed at manipulating the digital drawing image to be printed PD(n) (return to the block 682). In accordance with one embodiment of the present invention, if the artifact training image generator 440 determined that the image 10 digital drawing to be printed PD(n) manipulated is validated, the procedure of Digital manipulation 660 completed successfully. In accordance with one embodiment of the present invention, the sub- procedure 609 for generating artifactual training images APA(k) is used for training multiple 320 different neural networks, each used for 15 identify a respective printing defect (corresponding to a specific class C(j)) among the expected printing defects. In this case, this sub-procedure is applied to each digital drawing image to be printed PD(n) multiple times, each time manipulating it to introduce a different type of print defect corresponding to a different class C(j). 20 In accordance with another embodiment of the present invention, the sub-procedure 609 for generating APA(k) artifactual training images a single 320 neural network is used for training to identify different types of printing defects (corresponding to more than one class C(j)). In this case, for each new digital drawing image to be printed PD(n) to be manipulated 25 a randomly assigned type of printing defect will be introduced, so that create a homogeneous distribution of print defect types present in the database training 405. The APA(k) artifactual training images generated according to the embodiments of the present invention described permit the rapid 30 construction of a complete and efficient training database 405, which allows training one or more robust neural networks 320 without having to first perform long and inefficient manual operations that would be necessary in the event of wanted to build such a training database using only images digital images obtained from the acquisition of results from real printing processes. As already mentioned above, a manual process would indeed require a lot of time, and a corresponding waste of electricity and ink, excessive, in 5 how much, given the rarity of the occurrence of printing defects in real cases, would be necessary to collect countless images by acquiring print results of printing processes continued for very long intervals of time. The monitoring unit 130, and in particular the computing system of monitoring 170 using neural network 320 configured with data from 10 CD configurations obtained with the training procedure according to the shapes of the realization described, is particularly advantageous, as it allows perform very accurate and efficient monitoring of printing operations, but without requiring active intervention by specifically trained operators the monitoring of specific types of print production. 15 The Applicant has experimentally verified the accuracy of a network neural network trained using the training method according to the forms of achievements described here. Referring specifically to Figure 7, the Applicant has verified that a neural network that has been trained to recognize pitch printing defects 20 using AP(k) as training images: - of the APA(k) artifactual training images including (also) the APA(1) and APA(2) images (obtained by manipulating images to introduce a step defect), and - non-artifactual APN(k) training images including (also) 25 the APN(1) image (obtained by acquiring an image of a result of actual print including an actual pitch defect), was able to correctly classify the images P(1) and P(2) (obtained from the acquisition of actual print results) such as images including a defect in step. 30 It is emphasized that the images P(1) and P(2) which have been classified correctly from the neural network of this experimental example were not in any way way present among the AP(k) training images used for training of the neural network itself. In other words, the neural network was trained with a degree of generalization such as to allow for the correct classification of images as well very different from those used during training. Of course, in order to meet local and specific requirements, a person 5 a person skilled in the art can apply various modifications to the invention described above and logical and / or physical alterations. More specifically, although the present invention has been described with a certain degree of particularity with reference to its forms of favorite realization, it should be understood that various omissions are possible, substitutions and modifications in form and details, as well as other embodiments. 10 In particular, various embodiments of the invention may also be implemented without the specific details indicated in the previous description for provide a deeper understanding of it; on the contrary, well-known functions they may have been omitted or simplified so as not to weigh down the description with unnecessary details. 15 * * * * *

Claims

1. A method for training a neural network model (320) for use in print defect monitoring applications in a substrate printing process, said method comprising, under the control of a computing system (180): - providing (630) to the computing system a plurality of digital images to be printed; - providing (626) to the computing system one or more digital substrate images each representative of a corresponding substrate; - generating (650), by the computing system, a plurality of first training images representative of defect-free prints, said generating the first training images comprising combining (668) each of a group of said digital images to be printed with one of the digital substrate images;- generating (650), by the computing system, a plurality of second training images representative of prints each containing at least one printing defect, said generating the second training images comprising performing the following operations a), b) on each of a group of said digital images to be printed: a) digitally manipulating (660) said digital image to be printed to introduce therein at least one graphic alteration representing the corresponding at least one printing defect; b) combining (668) the digitally manipulated digital image to be printed with one of the substrate digital images;- training (604), by the computing system, said neural network model based on said first training images and said second training images to optimize the ability of said neural network model to classify the first training images as free of defects, and to classify the second training images as containing at least one defect.; 2. The method of claim 1, wherein said digitally manipulating (660) said digital image to be printed to introduce therein at least one graphic alteration representing the corresponding at least one printing defect comprises I24005-IT at least one of: - superimposing (682) on said digital image to be printed at least one image representing a stain; - blurring (682) said digital image by means of a blurring algorithm; - dividing (682) said digital image into a first portion and a second portion, and moving (682) the first portion so as to at least partially overlap it with the second portion, or so as to create a gap between the first portion and the second portion.

3. The method of claim 1 or claim 2, wherein said digital images to be printed comprise repetitive digital images to be printed each comprising a plurality of copies of the same single image, said digitally manipulating (660) said digital image to be printed to introduce therein at least one graphic alteration comprising: - altering (682), in each of said repetitive digital images to be printed, the position of at least one of said copies relative to the position of the other copies.

4. The method of any preceding claim, wherein: - the method further comprises, under the control of the computing system, extracting (633) from each digital image to be printed corresponding color channels; - said operation a) comprises: - digitally manipulating (660) each color channel of said digital image to be printed to introduce therein at least one graphic alteration representing the corresponding at least one printing defect; - combining (668) the digitally manipulated color channels with each other to obtain the corresponding digitally manipulated digital image to be printed.

5. The method of claim 4, wherein said digitally manipulating (660) said digital image to be printed to introduce therein at least one graphic alteration comprises at least one of: - erasing (682) from at least one color channel of said digital image to be printed a corresponding set of pixels; - translating (682) one or more of the color channels of said digital image to be printed with respect to the other color channels of said image so as to create a misalignment between said one or more of the color channels and said other color channels.

6. The method of any preceding claim, wherein said method comprises, under the control of the computing system: - scaling (633) said digital images to be printed to a first resolution corresponding to a printing resolution used for printing said digital images to be printed on substrate; - generating (650) said first training images and said second training images using said digital images to be printed scaled to said first resolution; - scaling (669) said first training images and said generated second training images to a second resolution, said training (604) said neural network model being performed, by the computing system, on the basis of said first training images and said second training images at the second resolution.

7. The method of any preceding claim, wherein: - said at least one print defect comprises a single type of print defect; - said first training images comprise a first number of first training images; - said second training images comprise a second number of second training images representative of prints each containing at least one print defect of said single type of print defect; - said first number corresponds to 40%-60% of a total number of training images used by the computing system for training said neural network model. I24005-IT 8. The method of any preceding claim, wherein: - said at least one print defect comprises a plurality of different types of print defects; - said first training images comprise a first number of first training images; - said second training images comprise for each type of print defect of said plurality a corresponding second number of second training images representative of prints each containing at least one print defect of said type of print defect; - said first number and said second numbers being substantially the same.

9. The method of any preceding claim, wherein said training (604), by the computing system, said neural network model further comprises: - including (608) in said first training images digital images corresponding to camera acquisitions of images of prints free of print defects; - including (608) in said second training images digital images corresponding to camera acquisitions of images of prints containing at least one print defect.

10. A method for monitoring printing defects in a substrate printing process, the method comprising: - acquiring (505) by one or more image acquisition sensors an image of a print product obtained by a substrate printing process; - providing said acquired image to a monitoring computing system (170); - using (509), by the monitoring computing system, a neural network model (320) trained in accordance with any of the preceding claims to classify said acquired image; I24005-IT - evaluating (509), by the monitoring computing system, the presence of printing defects in the print product on the basis of a classification of said acquired image carried out by said neural network model; - providing (520, 525, 530) as output an outcome of said evaluation.