Line defect classification method and line defect classification system

Abstract

To automatically classify a stripe-like defect included in a printed image so as to obtain a result close to visual classification.SOLUTION: At least a peak mean brightness value and a variation in a mean brightness value around a line defect are used as feature values, and a learning model is caused to learn a relationship between the feature values and a sorting destination, thereby generating a sorting model for sorting the line defect (Step S150). Thereafter, when a line defect is detected, at least a peak average brightness value and a variation are obtained as feature values representing features of the line defect (Step S180). Then, by inputting the feature amount to the sorting model, a sorting destination according to the feature of the detected line defect is determined (Step S190).SELECTED DRAWING: Figure 8

Description

[Technical Field]

[0001] The present invention relates to a technique for classifying streak-like defects that can occur in a printed image when an inkjet printing device has a nozzle that is in a discharge-failure state (hereinafter referred to as a "discharge-failure nozzle"). [Background technology]

[0002] Inkjet printing devices are widely known for printing by ejecting ink onto a print medium such as printing paper or film. In inkjet printing devices, as the ejection interval increases, problems such as drying of the ink due to evaporation of solvent near the nozzle, the intrusion of air bubbles into the nozzle, and the adhesion of dust to the nozzle can occur during printing. As a result, ejection defects, such as ink not being ejected from the nozzle or ink droplets ejected from the nozzle landing at positions other than their intended positions, can occur. Such ejection defects can result in poor print quality. For example, a streak-like defect (hereinafter referred to as a "streak defect" or simply "streak") corresponding to a nozzle that is not ejecting ink can appear in the printed image. Suppressing the occurrence of such streak-like defects is an important issue for inkjet printing devices. When a streak-like defect occurs, maintenance procedures (such as wiping, purging, and flushing) are performed to restore the functionality of the nozzle that is not ejecting ink.

[0003] In order to prevent streak defects from occurring in printed images of actual printed products, streak defects are detected based on captured images (captured data) obtained by capturing a printed image of a test chart with an imaging device at appropriate times. A method for detecting streak defects is disclosed, for example, in Japanese Patent Application Laid-Open No. 2017-181094. According to the method disclosed in Japanese Patent Application Laid-Open No. 2017-181094, a captured image obtained by capturing a printed image of an image with uniform density across the entire surface (a so-called "solid image") is divided into multiple local regions in the direction in which the streaks extend. Then, an average value of signal values ​​within the local regions is calculated for each pixel position in a direction perpendicular to the direction in which the streaks extend, and the streak defects are detected based on the results of comparing each value of the profile with two thresholds (a first threshold and a second threshold). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Laid-Open No. 2017-181094 Summary of the Invention [Problem to be solved by the invention]

[0005] Conventional methods for detecting streak defects can detect streaks (streak defects) based on a captured image of a printed image such as a test chart, but cannot grasp the degree of defect of each streak (the degree to which the defect affects the quality of the printed product). In other words, it is not possible to classify streaks according to the degree of defect. In this regard, for example, after an inkjet printing device is delivered from a manufacturer to a customer (e.g., a printing company), the customer quantitatively evaluates the print condition, and in this case, not only the number of streaks but also the degree of defect of each streak must be determined. Below is an example of criteria for classifying streaks into three levels: "large, medium, and small." Large: Defects in which the base is clearly visible from one end of the printed area to the other. Small: A defect that is not noticeable macroscopically, but is small enough that the base material appears and disappears in a linear pattern microscopically. Medium: Defects between "large" and "small"

[0006] If streaks are classified by the degree of defect, it becomes possible to take measures according to the degree of defect of the streaks that have occurred. For example, it becomes possible to take measures such as "if streaks classified as 'large' are present, perform cleaning; if streaks classified as 'large' are not present but streaks classified as 'medium' are present, perform flushing; if streaks classified as 'large' and 'medium' are present but streaks classified as 'small' are present, continue printing as is." However, as described above, conventional methods cannot grasp the degree of defect of each streak.

[0007] It is also conceivable to automatically classify streaks based on the results of comparing average luminance value data (corresponding to the profile described in JP 2017-181094 A) obtained from a captured image of a printed image of a test chart with a pre-prepared threshold value. In this regard, for example, as shown in FIG. 31 , it is conceivable to classify streaks based on the results of comparing the average luminance value data with three threshold values ​​(a first threshold value 91, a second threshold value 92, and a third threshold value 93). In the example shown in FIG. 31 , the degree of defect of the streak corresponding to the average luminance value at the position marked with the reference numeral 90 is judged to be “medium.” However, there is a large difference between the classification results obtained by such a method of automatically classifying streaks based on the average luminance value and the classification results obtained by visual inspection by an expert.

[0008] Therefore, an object of the present invention is to automatically classify streak defects contained in printed images so as to obtain results similar to those obtained by visual classification. [Means for solving the problem]

[0009] A first invention is a streak defect classification method for classifying streak defects contained in a printed image, comprising: a test chart printing step of printing a test chart for detecting streak defects; an imaging step of imaging a print image obtained in the test chart printing step; an average luminance value calculation step of calculating, for each of a plurality of pixel positions consecutive in a second direction perpendicular to a first direction in which the streak defect extends, an average luminance value which is an average value of luminance values ​​of a plurality of pixels at the same pixel position in the second direction, based on imaging data consisting of a plurality of luminance values ​​obtained in the imaging step; a feature value calculation step of calculating a feature value representing a feature of the streak defect based on the imaging data or the average luminance value for each of the plurality of pixel positions; a classification destination determination step of determining a classification destination according to the feature amounts by inputting the feature amounts into a trained learning model for classifying the streak-like defects; Including, The feature amount calculation step includes: a maximum value extraction step of extracting a maximum value corresponding to the streak defect from the average luminance values ​​for each of the plurality of pixel positions; a variation calculation step of calculating a variation of the average brightness value around the streak defect; Including, In the classification destination determination step, the maximum value and the variance are input as the feature amounts to the trained learning model.

[0010] The second invention is the first invention, the feature amount calculation step further includes a streak width calculation step of calculating a streak width, which is a width of the streak-like defect, In the classification destination determination step, the streak width is further input as the feature amount to the trained learning model.

[0011] The third invention is the second invention, In the streak width calculation step, two pixel positions corresponding to two minimum values ​​sandwiching the maximum value corresponding to the streak defect among a plurality of minimum values ​​extracted from the average luminance value for each of the plurality of pixel positions are obtained, and a value proportional to the standard deviation of an approximation curve obtained by fitting the relationship between the pixel positions and the average luminance value to a Gaussian function based on the plurality of average luminance values ​​between the two pixel positions is calculated as the streak width.

[0012] The fourth invention is the first invention, The variation calculation step is characterized in that a first predetermined distance is set to a distance less than a second predetermined distance, and the standard deviation of the average brightness value in a range from the center of the streak defect in the second direction to the first predetermined distance or more and the second predetermined distance or less is calculated as the variation.

[0013] The fifth invention is the first invention, The trained learning model is characterized in that it is a support vector machine.

[0014] The sixth invention is any one of the first to fifth inventions, The streak defect classification method further includes a streak detection step of detecting the streak defect based on the image data or the average brightness value for each of the plurality of pixel positions.

[0015] A seventh aspect of the present invention is the sixth aspect of the present invention, In the streak detection step, an image of a pixel position corresponding to a maximum value equal to or greater than a predetermined threshold value among a plurality of maximum values ​​extracted from the average brightness value for each of the plurality of pixel positions is detected as the streak defect.

[0016] An eighth aspect of the present invention is a streak defect classification method for classifying streak defects contained in a printed image, comprising the steps of: a first test chart printing step of printing a first test chart for detecting streak defects; a first imaging step of imaging a print image obtained in the first test chart printing step; a first average luminance value calculation step of calculating, for each of a plurality of pixel positions consecutive in a second direction perpendicular to a first direction in which the streak defect extends, a first average luminance value which is an average value of luminance values ​​of a plurality of pixels at the same pixel position in the second direction, based on first imaging data consisting of a plurality of luminance values ​​obtained in the first imaging step; a first streak detection step of detecting the streak defect based on the first imaging data or the first average luminance value for each of the plurality of pixel positions; a first maximum value extraction step of extracting, as a first maximum value, a maximum value corresponding to the streak defect detected in the first streak detection step from the first average luminance values ​​for each of the plurality of pixel positions; a first variation calculation step of calculating a variation in the first average brightness value around the streak defect detected in the first streak detection step as a first variation; a classification destination designation step in which an operator designates a classification destination corresponding to a combination of the first maximum value and the first variation; a learning step of using the first maximum value, the first variation, and the classification destination designated in the classification destination designation step as learning data to cause a learning model to learn the relationship between the combination of the first maximum value and the first variation and the classification destination; a second test chart printing step of printing a second test chart for detecting streak defects; a second imaging step of imaging the print image obtained in the second test chart printing step; a second average luminance value calculation step of calculating, for each of the plurality of pixel positions, a second average luminance value that is an average value of luminance values ​​of a plurality of pixels that are at the same pixel position in the second direction, based on second imaging data consisting of a plurality of luminance values ​​obtained in the second imaging step; a second streak detection step of detecting the streak defect based on the second imaging data or the second average luminance value for each of the plurality of pixel positions; a second maximum value extraction step of extracting, as a second maximum value, a maximum value corresponding to the streak defect detected in the second streak detection step from the second average luminance values ​​for each of the plurality of pixel positions; a second variation calculation step of calculating a variation in the second average brightness value around the streak defect detected in the second streak detection step as a second variation; a classification destination determination step of inputting the second maximum value and the second variation into the learning model that has been trained by the learning step, and determining a classification destination according to a combination of the second maximum value and the second variation; The present invention is characterized by comprising:

[0017] A ninth aspect of the present invention is a streak defect classification system for classifying streak defects contained in a printed image, the system comprising: an average luminance value calculation unit that calculates an average luminance value, which is an average value of luminance values ​​of a plurality of pixels at the same pixel position in a second direction perpendicular to a first direction in which the streak defect extends, for each of a plurality of pixel positions that are continuous in the second direction, based on imaging data consisting of a plurality of luminance values ​​obtained by imaging a printed image of a test chart for detecting streak defects; a feature amount calculation unit that calculates a feature amount representing a feature of the streak defect based on the imaging data or the average luminance value for each of the plurality of pixel positions; a classification destination determination unit that determines a classification destination according to the feature amount by inputting the feature amount into a trained learning model for classifying the streak-like defect; Equipped with The feature amount calculation unit a maximum value extracting unit that extracts a maximum value corresponding to the streak defect from the average luminance value for each of the plurality of pixel positions; a variation calculation unit that calculates the variation of the average brightness value around the streak defect; Including, The trained learning model is characterized in that the maximum value and the variance are input as the feature quantities.

[0018] Furthermore, modifications that can be understood by referring to the embodiments and drawings of the ninth invention are considered as means for solving the problems. [Effects of the Invention]

[0019] According to the first aspect of the present invention, when a printed image of a test chart for detecting streak defects contains a streak defect, a feature quantity representing the characteristics of the streak defect is obtained. The feature quantities include the maximum value of the average luminance value corresponding to the streak defect and the variation in the average luminance value around the streak defect. Then, by inputting the feature quantities into a trained learning model for classifying streak defects, a classification corresponding to the streak defect's characteristics is obtained. According to human visual perception, the greater the variation in brightness (luminance value) around a streak, the less likely the defect is to be perceived. In other words, the degree of defect perceived by humans regarding a streak in a printed image depends on the variation in luminance value around the streak. In this regard, according to the first aspect of the present invention, the streak defect is classified taking into account the variation in the average luminance value around the streak defect, thereby achieving results similar to those obtained by visual inspection. Furthermore, since a trained learning model is used for classification, streak defects in printed images are automatically classified. As a result, it is possible to automatically classify streak defects contained in a printed image so as to obtain results similar to those obtained by visual classification.

[0020] According to the second aspect of the present invention, the streak width (width of the streak defect) is also taken into consideration when classifying the streak defects, so that a result closer to classification by visual inspection can be obtained.

[0021] According to the third aspect of the invention, the same effects as those of the second aspect of the invention can be obtained.

[0022] According to the fourth aspect of the present invention, data on the portion where the streak defect occurs is excluded when calculating the variation in the average brightness value, so that the variation in the average brightness value around the streak defect can be found with high accuracy.

[0023] According to the fifth aspect of the present invention, even if the number of pieces of learning data used in learning by the learning model is relatively small, it is possible to suitably classify streak defects.

[0024] According to the sixth aspect of the present invention, streak defects can be detected with an accuracy close to that of visual detection.

[0025] According to the seventh aspect of the invention, the same effects as those of the sixth aspect of the invention can be obtained.

[0026] According to the eighth aspect of the invention, the same effects as those of the first aspect of the invention can be obtained.

[0027] According to the ninth to fifteenth aspects of the invention, the same effects as those of the first to seventh aspects of the invention can be obtained, respectively. [Brief explanation of the drawings]

[0028] [Figure 1] 1 is a diagram illustrating the overall configuration of a printing system according to an embodiment of the present invention. [Figure 2] FIG. 2 is a schematic diagram illustrating an example of the configuration of an inkjet printing apparatus according to the embodiment. [Figure 3] FIG. 2 is a plan view showing an example of the configuration of a recording unit in the embodiment. [Figure 4] 5A to 5C are diagrams for explaining the arrangement of nozzles in an ink ejection head in the embodiment. [Figure 5] FIG. 2 is a block diagram showing a hardware configuration of the print control device in the embodiment. [Figure 6] FIG. 2 is a block diagram showing a functional configuration of a control unit in the embodiment. [Figure 7] FIG. 2 is a block diagram showing a detailed functional configuration of a feature amount calculation unit in the embodiment. [Figure 8] 10 is a flowchart showing a general procedure of a series of processes for classifying streak defects in the embodiment. [Figure 9]FIG. 10 is a diagram showing an example of average luminance values ​​for each of a plurality of pixel positions in the embodiment. [Figure 10] 10 is a flowchart showing a detailed procedure of pre-processing in the embodiment. [Figure 11] FIG. 3 is a diagram schematically showing a test chart in the embodiment. [Figure 12] FIG. 10 is a diagram for explaining a process of removing noise using a median filter in the embodiment. [Figure 13] 10A to 10C are diagrams for explaining a process of correcting uneven brightness using a median filter in the embodiment. [Figure 14] 8A to 8C are diagrams for explaining detection of a streak defect in the embodiment. [Figure 15] FIG. 10 is a diagram showing an example of a stepped chart in the embodiment. [Figure 16] 10A to 10C are diagrams for explaining detection of a streak defect using a stepped chart in the embodiment. [Figure 17] 10 is a flowchart showing a detailed procedure for calculating a feature amount in the embodiment. [Figure 18] 10A to 10C are diagrams for explaining how to obtain a peak average luminance value in the embodiment. [Figure 19] 10A and 10B are diagrams for explaining a calculation process for fitting the relationship between pixel positions and average luminance values ​​to a Gaussian function in the embodiment. [Figure 20] FIG. 10 is a diagram for explaining a streak width in the embodiment. [Figure 21] 8A to 8C are diagrams for explaining calculation of the variation in average brightness value around a streak defect in the embodiment. [Figure 22] FIG. 10 is a diagram for explaining a discrimination plane obtained by learning using a support vector machine in the embodiment. [Figure 23]FIG. 10 is a diagram for explaining learning by a learning unit (learning of the relationship between a combination of three feature amounts and a classification destination) in the embodiment. [Figure 24] FIG. 10 is a diagram for explaining classification by a classification unit in the embodiment. [Figure 25] FIG. 10 is a diagram for explaining an experiment for confirming the effects of the above embodiment. [Figure 26] FIG. 10 is a diagram for explaining learning by a learning unit (learning of the relationship between a combination of two feature amounts and a classification destination) in a first modified example of the embodiment. [Figure 27] FIG. 10 is a diagram for explaining classification by a classification unit in the first modified example. [Figure 28] FIG. 10 is a diagram showing an example of the structure of a neural network used in a second modified example of the embodiment. [Figure 29] FIG. 10 is a diagram for explaining a process during learning using a neural network in the second modified example. [Figure 30] FIG. 10 is a diagram for explaining a process during learning using a neural network in the second modified example. [Figure 31] FIG. 10 is a diagram for explaining classification of streaks based on the results of comparing data of average luminance values ​​with three threshold values. DETAILED DESCRIPTION OF THE INVENTION

[0029] Hereinafter, an embodiment of the present invention will be described with reference to the accompanying drawings.

[0030] <1. Overall configuration of the printing system> FIG. 1 is a diagram illustrating the overall configuration of a printing system according to an embodiment of the present invention. This printing system comprises an inkjet printing device 10 and a print data generating device 30. The inkjet printing device 10 and the print data generating device 30 are connected to each other via a LAN 4. The print data generating device 30 generates print data by performing rasterization processing on input data such as a PDF file. This print data is not subjected to halftoning; halftoning is performed by a print control device 100 within the inkjet printing device 10, as described below. The print data generated by the print data generating device 30 is sent to the inkjet printing device 10 via the LAN 4. The inkjet printing device 10 comprises a printing press main unit 200 and a print control device 100 that controls the operation of the printing press main unit 200. The inkjet printing device 10 outputs a print image on printing paper as a printing medium based on the print data transmitted from the print data generating device 30, without using printing plates. Note that the present invention can also be applied when a printing medium other than printing paper (e.g., film) is used. In this embodiment, the print control device 100 realizes a streak defect classification system.

[0031] <2. Configuration of the printer body of the inkjet printing device> 2 is a schematic diagram showing an example of the configuration of the inkjet printing apparatus 10. As described above, the inkjet printing apparatus 10 is made up of the print control device 100 and the printing machine main body 200.

[0032] The printing machine main body 200 includes a paper feed section 202 that supplies printing paper 5 to the printing mechanism 201, the printing mechanism 201 that prints on the printing paper 5, and a paper winding section 208 that winds up the printing paper 5 into a roll after printing.

[0033] The printing mechanism 201 includes a first drive roller 203 for transporting the printing paper 5 inside, a plurality of support rollers 204 for transporting the printing paper 5 inside the printing mechanism 201, a recording unit 205 for recording a print image on the printing paper 5, a drying mechanism 206 for drying the printing paper 5 on which the print image has been recorded, and a second drive roller 207 for outputting the printing paper 5 from inside the printing mechanism 201. The recording unit 205 is composed of a K head unit 25K that ejects K (black) ink, a C head unit 25C that ejects C (cyan) ink, an M head unit 25M that ejects M (magenta) ink, and a Y head unit 25Y that ejects Y (yellow) ink. The printing mechanism 201 also includes a contact image sensor (CIS) 40 as an imaging device that captures the print image recorded on the printing paper 5 by the recording unit 205. The imaging data (captured image) obtained by the contact image sensor 40 capturing the print image is sent to the print control device 100. In the following, when the color of ink ejected from the head unit is not to be distinguished, the head unit will be denoted by the reference symbol 25.

[0034] FIG. 3 is a plan view showing an example of the configuration of the recording unit 205. As shown in FIG. 3, the recording unit 205 is composed of a K-color head unit 25K, a C-color head unit 25C, an M-color head unit 25M, and a Y-color head unit 25Y, which are arranged in a row in the transport direction of the printing paper 5. Each head unit 25 is composed of a plurality of ink ejection heads (print heads) 251 arranged in a staggered pattern. Each ink ejection head 251 includes a large number of nozzles (not shown in FIG. 3) that eject ink. Each nozzle of the ink ejection head 251 included in the K-color head unit 25K ejects K-color ink, each nozzle of the ink ejection head 251 included in the C-color head unit 25C ejects C-color ink, each nozzle of the ink ejection head 251 included in the M-color head unit 25M ejects M-color ink, and each nozzle of the ink ejection head 251 included in the Y-color head unit 25Y ejects Y-color ink.

[0035] FIG. 4 is a diagram illustrating the arrangement of nozzles in the ink ejection head 251. Typically, the ink ejection head 251 includes multiple rows of nozzle groups, each consisting of multiple nozzles arranged in the paper width direction. In the example shown in FIG. 4, the ink ejection head 251 includes four rows of nozzle groups. The portion marked with reference numeral 41 in FIG. 4 schematically shows the landing positions on the printing paper 5 of ink ejected from each nozzle. The multiple nozzles in the ink ejection head 251 are arranged so that the landing positions of ink ejected from the nozzles in the first row of nozzle group, the landing positions of ink ejected from the nozzles in the second row of nozzle group, the landing positions of ink ejected from the nozzles in the third row of nozzle group, and the landing positions of ink ejected from the nozzles in the fourth row of nozzle group are all different from one another. For example, the landing positions of ink ejected from the nozzles in the first row of nozzle group are between the landing positions of ink ejected from the nozzles in the third row of nozzle group and the landing positions of ink ejected from the nozzles in the fourth row of nozzle group. In the example shown in Figure 4, the landing position 42 of ink ejected from the nozzle marked with the symbol 252(p) is a position between the landing position 43 of ink ejected from the nozzle marked with the symbol 252(q) and the landing position 44 of ink ejected from the nozzle marked with the symbol 252(r).

[0036] The configurations shown in FIGS. 2 to 4 are merely examples, and the specific configurations of the printing mechanism 201, the recording unit 205, and the ink ejection head 251 are not particularly limited.

[0037] <3. Hardware configuration of print control device> FIG. 5 is a block diagram showing the hardware configuration of the print control device 100. As shown in FIG. 5, the print control device 100 includes a main body 110, an auxiliary storage device 121, an optical disk drive 122, a display unit 123, a keyboard 124, and a mouse 125. The main body 110 includes a CPU 111, a memory 112, a first disk interface unit 113, a second disk interface unit 114, a display control unit 115, an input interface unit 116, and a communication interface unit 117. The CPU 111, the memory 112, the first disk interface unit 113, the second disk interface unit 114, the display control unit 115, the input interface unit 116, and the communication interface unit 117 are connected to one another via a system bus. The auxiliary storage device 121 is connected to the first disk interface unit 113. The optical disk drive 122 is connected to the second disk interface unit 114. The display control unit 115 is connected to a display unit (display device) 123. A keyboard 124 and a mouse 125 are connected to the input interface unit 116. The printing machine main body 200 is connected to the communication interface unit 117 via a communication cable. The communication interface unit 117 is also connected to the LAN 4. The auxiliary storage device 121 is a magnetic disk device or the like. An optical disk 19, which is a computer-readable recording medium such as a CD-ROM or DVD-ROM, is inserted into the optical disk drive 122. The display unit 123 is a liquid crystal display or the like. The display unit 123 is used to display information desired by the operator. The keyboard 124 and mouse 125 are used by the operator to input instructions to this printing control device 100.

[0038] The auxiliary storage device 121 stores a print control program 13 (a program for controlling the execution of a print process by the printing press main body 200). In this embodiment, the print control program 13 includes, as a subprogram, a streak defect classification program for classifying streak defects contained in a printed image. The CPU 111 reads the print control program 13 stored in the auxiliary storage device 121 into the memory 112 and executes it to realize various functions of the print control device 100. The memory 112 includes RAM (Random Access Memory) and ROM (Read Only Memory). The memory 112 functions as a work area for the CPU 111 to execute the print control program 13 stored in the auxiliary storage device 121. The print control program 13 is provided by being stored on the computer-readable recording medium (non-transitory recording medium). That is, for example, a user purchases an optical disc 19 as a recording medium for the print control program 13, inserts it into the optical disc drive 122, reads the print control program 13 from the optical disc 19, and installs it in the auxiliary storage device 121.

[0039] In the example shown in FIG. 5, the print control device 100 is provided with only one CPU 111 as a processor, but this is not limited to this. A configuration using multiple processors, such as a configuration using multiple CPUs, can also be adopted. In addition to the CPU 111, other processors such as an MPU (Micro Processing Unit), a GPU (Graphics Processing Unit), or a DSP (Digital Signal Processor) can also be adopted. A combination of multiple types of processors can also be used. For example, with regard to the functional components within the print control device 100 (see FIG. 6), some components and the remaining components can be implemented by different processors. Furthermore, a configuration including an FPGA (Field-Programmable Gate Array) or an ASIC (Application Specific Integrated Circuit) can also be adopted.

[0040] <4. Functional configuration of print control device> 6 is a block diagram showing the functional configuration of the control unit 50 that is realized by executing the print control program 13 on the print control device 100. The control unit 50 includes a print data storage unit 510, a halftone processing unit 512, an ink discharge control unit 514, a conveyance control unit 520, a drying control unit 530, an imaging control unit 540, an average brightness value calculation unit 550, a streak detection unit 560, a feature amount calculation unit 570, a learning unit 580, and a classification destination determination unit 590.

[0041] The print data storage unit 510 stores the RIP-processed print data 60 sent from the print data generation device 30. The halftone processing unit 512 performs halftone processing on the RIP-processed print data 60 to generate halftone image data 61, which includes information indicating the ink dot size corresponding to each pixel. The specific halftone processing method is not particularly limited, and known methods such as error diffusion and dithering can be used. The ink ejection control unit 514 controls the amount of ink ejected from each nozzle included in each ink ejection head 251 constituting the recording unit 205 based on the halftone image data 61. Note that the print data storage unit 510 stores test chart data representing a test chart for detecting streak defects as print data 60 related to the present invention. The ink ejection control unit 514 controls the amount of ink ejected from each nozzle based on the halftone image data 61 generated by halftone processing based on the test chart data, thereby forming a print image of the test chart on the printing paper 5.

[0042] The conveyance control unit 520 controls the speed (conveyance speed) at which the conveyance mechanism 29 conveys the printing paper 5. The conveyance mechanism 29 is realized by the paper delivery unit 202, first drive roller 203, multiple support rollers 204, second drive roller 207, and paper winding unit 208 (see FIG. 2). The drying control unit 530 controls the temperature (drying temperature) at which the drying mechanism 206 dries the printing paper 5 after printing. The imaging control unit 540 controls the timing at which the contact image sensor 40 captures the printed image.

[0043] The average brightness value calculation unit 550, the streak detection unit 560, the feature calculation unit 570, the learning unit 580, and the classification destination determination unit 590 are functional components directly involved in the process of classifying streak defects based on the imaging data 62 consisting of a plurality of brightness values ​​obtained by the contact image sensor 40 capturing an image of the printed image of the test chart.

[0044] The average brightness value calculation unit 550 calculates an average brightness value 63, which is the average value of the brightness values ​​of multiple pixels at the same pixel position in the paper width direction, for each of multiple consecutive pixel positions in the paper width direction, which is a direction perpendicular to the transport direction of the printing medium (the direction in which the streak defects extend), based on the imaging data 62. The transport direction of the printing paper corresponds to the first direction, and the paper width direction corresponds to the second direction.

[0045] The streak detection unit 560 detects streak defects based on the average luminance value (average luminance value for each of the multiple pixel positions) 63 calculated by the average luminance value calculation unit 550. How streak defects are detected will be described in more detail later. The streak detection unit 560 outputs streak information 64 that identifies the position of the streak defect.

[0046] The feature amount calculation unit 570 calculates a feature amount 65 representing the feature of the streak defect based on the streak information 64 and the average luminance value 63 (average luminance value for each of the multiple pixel positions) calculated by the average luminance value calculation unit 550. In this embodiment, as shown in FIG. 7 , the feature amount calculation unit 570 includes a peak average luminance value extraction unit 572, a streak width calculation unit 574, and a variation calculation unit 576. The peak average luminance value extraction unit 572 extracts the "peak value of the average luminance value" (i.e., the maximum value) corresponding to the streak defect from the average luminance values ​​63 for each of the multiple pixel positions based on the streak information 64 as a peak average luminance value 652. The streak width calculation unit 574 calculates a streak width 654, which is the width of the streak defect, based on the streak information 64 and the average luminance value 63 calculated by the average luminance value calculation unit 550. The variation calculation unit 576 calculates the variation 656 of the average brightness value 63 around each streak defect based on the streak information 64 and the average brightness value 63 calculated by the average brightness value calculation unit 550. This variation 656 represents the granularity around the streak defect. As described above, in this embodiment, the feature calculation unit 570 calculates the peak average brightness value 652, the streak width 654, and the variation (the variation of the average brightness value 63 around the streak defect) 656 as feature quantities representing the characteristics of the streak defect. A more detailed explanation of how these three feature quantities 65 are calculated will be given later. Note that it is also possible to employ a configuration in which the feature calculation unit 570 calculates at least one feature quantity based on the imaging data 62 rather than the average brightness value 63.

[0047] The learning unit 580 performs learning to classify streak defects. In this embodiment, streak defects are classified into four levels: "large, medium, small, and non-streak." Note that "non-streak" means that the detection by the streak detection unit 560 was a false detection (misdetection). To enable learning by the learning unit 580, an expert (operator) specifies in advance the classification destination corresponding to the combination of the three feature amounts 65 (peak average luminance value 652, streak width 654, and variation 656). Under this premise, the learning unit 580 uses the three feature amounts 65 and labels (teaching data) 66 indicating the classification destination specified by the expert as learning data to train a learning model on the relationship between the combination of the three feature amounts 65 and the classification destination. Learning using a sufficient amount of learning data in the learning unit 580 generates a trained learning model with optimized parameters. The trained learning model thus generated is held in the classification destination determination unit 590 as a classification model 592 for classifying streak defects.

[0048] The classification destination determination unit 590 includes the above-mentioned classification model (a trained learning model for classifying streak defects) 592, and by inputting the feature 65 calculated by the feature calculation unit 570 into the classification model 592, a classification destination 67 corresponding to the feature 65 is calculated.

[0049] <5. Classification of streak defects> Next, a method for classifying streak defects contained in a printed image will be described.

[0050] <5.1 Outline of procedure> A summary of the steps of a series of processes for classifying streak defects will be described with reference to the flowchart shown in FIG. 8. Since the classification of streak defects is performed using a machine learning technique, the series of processes is divided into a learning phase and a classification phase (inference phase) as shown in FIG. 8. In this regard, the learning phase and the classification phase do not necessarily have to be executed by the same device (print control device 100). For example, the learning phase may be executed by a device of the manufacturer of the inkjet printing device 10, and then the classification phase may be executed by a device of the user of the inkjet printing device 10.

[0051] In the learning phase, first, preprocessing is performed to obtain data (hereinafter referred to as "learning source data") that will be used for learning by the learning unit 580 (step S110). Details will be described later, but this preprocessing ultimately generates data of average luminance values ​​63 (average values ​​of luminance values ​​of multiple pixels at the same pixel position in the paper width direction) for each of multiple pixel positions that are consecutive in the paper width direction as learning source data. The average luminance values ​​63 for each of the multiple pixel positions are expressed, for example, as shown in FIG. 9.

[0052] After the pre-processing (step S110) is completed, the learning source data is used to detect streak defects (step S120). Then, feature quantities 65 representing the characteristics of the streak defects detected in step S120 are calculated based on the learning source data (step S130). Details of the process for detecting streak defects and the process for calculating feature quantities 65 will be described later.

[0053] Next, an expert (operator) specifies a classification destination corresponding to the feature amounts 65 calculated in step S130 (so-called "labeling work") (step S140). As described above, in this embodiment, learning is performed using three feature amounts 65 (peak average luminance value 652, streak width 654, and variation 656). Therefore, in step S140, an expert specifies a classification destination corresponding to the combination of these three feature amounts 65.

[0054] Thereafter, the three feature quantities 65 and the labels 66 indicating the classification destinations specified in step S140 are used as training data, and a process is performed in which the training model learns the relationship between the combinations of the three feature quantities 65 and the classification destinations (step S150). As a result, a trained training model (the above-mentioned classification model 592) with optimized parameters is generated. This marks the end of the training phase.

[0055] In the classification phase, first, preprocessing is performed to obtain inspection data that is the target for detecting and classifying streak defects (step S160). The preprocessing in step S160 is the same as the preprocessing in step S110. Therefore, the inspection data obtained in step S160 is data of the average brightness value 63 for each of a plurality of pixel positions that are consecutive in the paper width direction (see FIG. 9).

[0056] After the pre-processing (step S160) is completed, the inspection data is used to detect streak defects (step S170). Then, feature quantities 65 representing the characteristics of the streak defects detected in step S170 are calculated based on the inspection data (step S180).

[0057] Thereafter, the feature quantities 65 (specifically, the peak average luminance value 652, the streak width 654, and the variation 656) calculated in step S180 are input to the classification model 592, which is a trained learning model generated in step S150, to determine a classification destination 67 according to the feature quantities 65 (step S190). This ends the classification phase.

[0058] In this embodiment, step S120 implements a first streak detection step, step S140 implements a classification destination designation step, step S150 implements a learning step, step S170 implements a streak detection step and a second streak detection step, step S180 implements a feature calculation step, and step S190 implements a classification destination determination step.

[0059] 5.2 Pretreatment The detailed procedure of the preprocessing (the processing of steps S110 and S160 in FIG. 8) will be described with reference to the flowchart shown in Fig. 10. This preprocessing is performed in both the learning phase and the classification phase (inference phase).

[0060] After the pre-processing starts, a test chart for detecting streak defects is printed (step S210). Specifically, halftone processing is performed on the test chart data stored in the print data storage unit 510 to generate halftone image data 61, and the ink discharge control unit 514 controls the amount of ink discharged from each nozzle based on the halftone image data 61 to form a print image of the test chart on the printing paper 5 (see FIG. 6).

[0061] FIG. 11 is a diagram schematically illustrating a test chart 70. The test chart 70 includes three types of constant density regions for each ink color. More specifically, as shown in FIG. 11, the test chart 70 includes a 100% solid image region, an 80% solid image region, and a 60% solid image region for each of the K, C, M, and Y colors. Note that in FIG. 11, the 80% solid image region for the C color, for example, is denoted by the symbol C(80).

[0062] After the test chart 70 described above is printed, the printed image of the test chart 70 is imaged by the contact image sensor 40 (step S220). This results in image data 62 consisting of a plurality of brightness values. Each of the plurality of brightness values ​​constituting the image data 62 corresponds to a combination of coordinates in the transport direction and coordinates in the paper width direction of the printing paper 5. In other words, the image data 62 contains one brightness value for each combination of coordinates in the transport direction and coordinates in the paper width direction of the printing paper 5. In other words, the image data 62 contains brightness value data corresponding to each coordinate on a two-dimensional plane.

[0063] Next, a process is performed to extract data of a processing target region from the image data 62 obtained in step S220 (step S230). Detection and classification of streak defects are performed for each ink color and each density. Therefore, in step S230, a region corresponding to one ink color and one density is treated as one processing target region. That is, data is extracted from the image data 62 for each region corresponding to one ink color and one density. As can be seen from FIG. 11, in this embodiment, 12 regions are sequentially treated as processing target regions.

[0064] Next, data of a predetermined color channel is extracted from the data of the processing target area (step S240). In this regard, the print image is formed by ejecting K, C, M, and Y inks from the recording unit 205 onto the printing paper 5, and the data of any pixel in the captured image obtained by capturing the print image with the contact image sensor 40 is composed of a grayscale luminance value, an R (red) luminance value, a G (green) luminance value, and a Y (yellow) luminance value. In this embodiment, if the processing target area is a solid image area for an ink color other than K, luminance value data of a color that is complementary to the ink color is extracted from the data acquired in step S230. More specifically, if the processing target region is a solid image region for C color, R luminance value data is extracted from the data acquired in step S230, if the processing target region is a solid image region for M color, G luminance value data is extracted from the data acquired in step S230, and if the processing target region is a solid image region for Y color, B luminance value data is extracted from the data acquired in step S230. If the processing target region is a solid image region for K color, grayscale luminance value data is extracted from the data acquired in step S230.

[0065] Next, the data acquired in step S240 is subjected to a process of removing noise using a median filter (step S250). In this regard, a median filter is a filter that replaces the value of the pixel at the center of a predetermined region with the median value among the values ​​of the pixels that make up the predetermined region. For example, if we focus on the values ​​of the nine pixels within the dotted line labeled 712 in part A of FIG. 12, the fifth largest value from the top is 20. Therefore, when the median filter is applied, the value of the pixel in the thick line portion labeled 710 in part A of FIG. 12 is replaced with 20, as shown in the thick line portion labeled 714 in part B of FIG. 12. In step S250, such a median filter is applied to the entire region to be processed, thereby removing noise from the data acquired in step S240.

[0066] Next, for the data (brightness value data) obtained in step S250, the average value of multiple brightness values ​​at the same pixel position in the transport direction of the printing paper 5 is calculated for each of multiple consecutive pixel positions in the paper width direction (step S260).

[0067] Finally, a median filter is applied to the data (average luminance data) obtained in step S260 to correct brightness variations (step S270). The data obtained in step S260 is spatially one-dimensional. Therefore, in step S270, the value of the central pixel among a plurality of consecutive pixels in the paper width direction is replaced with the median value among the values ​​of the plurality of pixels. For example, among the five pixel values ​​within the dotted line labeled 722 in part A of FIG. 13, the third largest value from the top is 18. Therefore, when the median filter is applied, the pixel value in the bold line portion labeled 720 in part A of FIG. 13 is replaced with 18, as shown in the bold line portion labeled 724 in part B of FIG. 13. In step S270, this median filter is applied to the average luminance data obtained in step S260, thereby obtaining average luminance data from which brightness variations have been removed. This completes the preprocessing.

[0068] In this embodiment, the first test chart printing step is realized by step S210 performed in the learning phase, the test chart printing step and the second test chart printing step are realized by step S210 performed in the classification phase, the first imaging step is realized by step S220 performed in the learning phase, the imaging step and the second imaging step are realized by step S220 performed in the classification phase, the first average brightness value calculation step is realized by steps S230 to S270 performed in the learning phase, and the average brightness value calculation step and the second average brightness value calculation step are realized by steps S230 to S270 performed in the classification phase.

[0069] 5.3 Detection of streak defects The process of detecting streak defects (the processes of steps S120 and S170 in FIG. 8) will now be described in detail. In order to detect streak defects based on data generated by preprocessing (data on the average luminance value 63 for each of a plurality of pixel positions consecutive in the paper width direction), in this embodiment, a threshold value to be compared with the average luminance value 63 is determined in advance. Then, of a plurality of maximum values ​​extracted from the average luminance values ​​63 for each of the plurality of pixel positions, an image at a pixel position corresponding to a maximum value equal to or greater than the threshold value is detected as a streak defect. In the case where the average luminance values ​​63 for each of the plurality of pixel positions are expressed as in FIG. 9, a threshold value 730 is determined, for example, as shown in FIG. 14. In this case, an image at a pixel position corresponding to a maximum value indicated by arrows with reference numerals 731 to 733 in FIG. 14 is detected as a streak defect.

[0070] It should be noted that pixel positions where streak defects occur can also be detected using a chart other than the test chart 70 shown in FIG. 11. For example, pixel positions where streak defects occur can be detected based on imaging data (captured image) obtained by capturing a printed image of a stepped chart 74 as shown in FIG. 15. In FIG. 15, the black-shaded areas indicate areas where ink should be applied by ejecting ink from the nozzles included in the ink ejection head 251. As can be seen from FIG. 15, the stepped chart 74 is composed of numerous linear patterns. Each linear pattern is formed by ejecting ink from a corresponding nozzle. Now, assume that printing is actually performed based on data for forming the stepped chart 74 shown in FIG. 15, and a printed image such as that shown in FIG. 16 is obtained. In the dotted line area labeled 741 in FIG. 16, no ink is applied at all in the area where ink should be applied. This indicates that a nozzle that should apply ink to the dotted line area labeled 741 in FIG. 16 is experiencing an ejection failure, which could result in a streak defect at the pixel position corresponding to that nozzle.

[0071] Furthermore, in the learning phase, density value data that causes streak defects may be included in the test chart data to intentionally cause streak defects in the printed image of the test chart. In this case, the pixel positions where streak defects occur are identified in advance, eliminating the need to detect streak defects based on data generated by preprocessing. In other words, it is possible to omit the process of step S120 in FIG. 8.

[0072] 5.4 Calculation of feature quantities The detailed procedure of the process for calculating the feature amount (the processes of steps S130 and S180 in FIG. 8) will be described with reference to the flowchart shown in FIG.

[0073] First, a peak average luminance value 652 is calculated (step S310). As described above, after the average luminance value 63 for each of a plurality of pixel positions consecutive in the paper width direction is calculated by preprocessing (steps S110 and S160 in FIG. 8), an image of a pixel position corresponding to a maximum value equal to or greater than a predetermined threshold value among a plurality of maximum values ​​extracted from the average luminance values ​​63 for each of the plurality of pixel positions is detected as a streak defect (steps S120 and S170 in FIG. 8). For a streak defect detected in this manner, the corresponding maximum value is determined as the peak average luminance value. In other words, in a graph showing the relationship between pixel positions and average luminance values ​​(for example, the graph shown in FIG. 9), the maximum value of the average luminance value in the section from the start position of the monotonous increase to the end position of the monotonous decrease (provided that this maximum value is equal to or greater than the threshold value) is determined as the peak average luminance value of the corresponding streak defect. In the example shown in FIG. 18, the maximum value 750 of the average brightness value in the section 755 from the pixel position (pixel position in the paper width direction) 753 corresponding to the point 751 to the pixel position 754 corresponding to the point 752 is the peak average brightness value of the streak defect present at this position.

[0074] Next, the streak width is calculated (step S320). In step S320, first, an approximation curve of a graph representing the relationship between pixel positions and average luminance values ​​is obtained for the portion where the streak defect is detected. In the example shown in FIG. 18, a calculation process is performed to fit the relationship between pixel positions and average luminance values ​​to a Gaussian function based on a plurality of "combinations of pixel positions and average luminance values" in an interval 755 from pixel position 753 corresponding to point 751 to pixel position 754 corresponding to point 752. That is, two pixel positions corresponding to two minimum values ​​sandwiching a maximum value corresponding to the streak defect among a plurality of minimum values ​​extracted from the average luminance values ​​for each of a plurality of pixel positions in the paper width direction are obtained, and then a calculation process is performed to fit the relationship between pixel positions and average luminance values ​​to a Gaussian function based on a plurality of average luminance values ​​between the two pixel positions.

[0075] However, in practice, the average luminance value of a pixel position (hereinafter referred to as an "additional pixel position" for convenience) outside the streak defect, one of two pixel positions adjacent to the pixel position corresponding to the larger of the two minimum values, is considered to be equal to the smaller of the two minimum values, and the above calculation process is performed taking into account the data of the additional pixel position. Therefore, if the relationship between pixel positions and average luminance values ​​in the portion where a streak defect is detected is the relationship shown in Figure 18, as shown in Figure 19, the average luminance value of pixel position 756 adjacent to pixel position 754 corresponding to point 752 is considered to be equal to the average luminance value of pixel position 753 (see points 751 and 757). The above calculation process is then performed based on a plurality of average luminance values ​​for a section 758 between pixel positions 753 and 756.

[0076] By performing the calculation process as described above, an approximate curve such as the curve indicated by reference numeral 772 in Fig. 20 can be obtained. Note that in Fig. 20, reference numeral 771 is attached to the graph representing the "relationship between pixel position and average luminance value" from which the approximate curve was generated.

[0077] Incidentally, in the calculation process for obtaining an approximate curve, the standard deviation of the approximate curve is obtained. In this embodiment, a value twice the standard deviation is set as the streak width. In the example shown in FIG. 20, the value corresponding to the length of the arrow labeled with reference numeral 773 is the standard deviation of the approximate curve, and the value corresponding to the length of the arrow labeled with reference numeral 774 is twice the standard deviation. Note that the streak width does not necessarily have to be twice the standard deviation of the approximate curve; a value proportional to the standard deviation of the approximate curve may be set as the streak width.

[0078] After calculating the streak width, a variation 656 of the average luminance value 63 around the streak defect is calculated (step S330). In this embodiment, the standard deviation of the average luminance value 63 in a range within 20 pixels from the center of the streak defect, excluding a range within 5 pixels from the center of the streak defect, is calculated as the variation 656 of the average luminance value 63 around the streak defect. FIG. 21 shows a graph representing the "relationship between pixel position and average luminance value" in a range within 20 pixels from the center 781 of a streak defect. In the example shown in FIG. 21, the standard deviation of the average luminance value 63 in a range excluding the range represented by the arrows ... The first predetermined distance is set to a distance less than the second predetermined distance, and the standard deviation of the average luminance value 63 in the range from the center of the streak defect in the paper width direction that is equal to or greater than the first predetermined distance and equal to or less than the second predetermined distance can be calculated as the variation 656. In the above example, the distance of 5 pixels wide corresponds to the first predetermined distance, and the distance of 20 pixels wide corresponds to the second predetermined distance. Once the variation 656 is calculated, the process of calculating the feature amount ends.

[0079] In this embodiment, step S310 performed in the learning phase realizes the first maximum value extraction step, step S310 performed in the classification phase realizes the maximum value extraction step and the second maximum value extraction step, step S320 performed in the classification phase realizes the streak width calculation step, step S330 performed in the learning phase realizes the first variation calculation step, and step S330 performed in the classification phase realizes the variation calculation step and the second variation calculation step.

[0080] 5.5 Learning The learning performed in step S150 of FIG. 8 will be described in detail. In this embodiment, the relationship between the combination of the three feature amounts 65 (peak average luminance value 652, streak width 654, and variance 656) determined in step S130 and the classification destination specified in step S140 is learned. A method using a support vector machine is adopted as the learning method. A support vector machine is designed to separate two classes, and when learning is performed using the support vector machine based on three feature amounts, a discrimination plane separating the two classes is obtained. In this embodiment, as shown schematically in FIG. 22, a discrimination plane 79 separating the two classes is obtained in a three-dimensional space consisting of an axis of peak average luminance value, an axis of streak width, and an axis of variance. In step S190, a classification destination class for the data to be classified (target of classification) is determined based on this discrimination plane.

[0081] In this embodiment, four classes are defined to classify streak-like defects into four levels: "large, medium, small, and non-streak." In the following, the class corresponding to "large" will be referred to as "first class," the class corresponding to "medium" as "second class," the class corresponding to "small" as "third class," and the class corresponding to "non-streak" as "fourth class."

[0082] As described above, the support vector machine is used to separate two classes, but in this embodiment, it is necessary to classify streak defects into four classes. Therefore, in step S150, a discrimination plane separating the first class from the second class, a discrimination plane separating the first class from the third class, a discrimination plane separating the first class from the fourth class, a discrimination plane separating the second class from the third class, a discrimination plane separating the second class from the third class, a discrimination plane separating the second class from the fourth class, and a discrimination plane separating the third class from the fourth class are calculated. In other words, six discrimination planes are calculated.

[0083] If the peak average luminance value axis is the x-axis, the streak width axis is the y-axis, and the variation axis is the z-axis, each discrimination plane is expressed by the following equation (1). ax+by+cz+d=0 (1) In the above equation (1), a, b, c, and d are parameters.

[0084] In step S150, the values ​​of the four parameters (a, b, c, and d) included in the above formula (1) are calculated for each of the six discrimination planes. In this way, calculating the values ​​of the four parameters for each of the six discrimination planes corresponds to generating classification model 592 as a trained learning model.

[0085] During learning, as shown in FIG. 23, three feature quantities 65 (peak average luminance value 652, streak width 654, and variance 656) and a label 66 indicating a classification destination are input to a learning unit 580 as learning data. Based on the learning data, the learning unit 580 uses a support vector machine to learn the relationship between the combination of the three feature quantities 65 and the classification destination. Six parameter sets PR(1) to PR(6) for specifying the six discrimination planes described above are then output, thereby generating a classification model 592. Each parameter set PR consists of the four parameters (a, b, c, and d) described above.

[0086] 22 shows a discrimination plane 79 formed in a three-dimensional space consisting of an axis of peak average luminance, an axis of streak width, and an axis of variation, but it is also possible to employ a method (kernel method) in which the discrimination plane is formed in a new high-dimensional feature space obtained by converting the three feature quantities 65 using a kernel function. By employing the kernel method, feature quantity data that cannot be separated linearly can be converted into data that can be separated linearly.

[0087] <5.6 Determining classification> The process of determining the classification destination (the process of step S190 in FIG. 8) will now be described in detail. As described above, the streak defect to be classified is detected in step S170, and the feature quantities 65 for the streak defect to be classified are obtained in step S180. Under these assumptions, in step S190, as shown in FIG. 24, a peak average brightness value 652, a streak width 654, and a variation 656 of the average brightness value 63 around the streak defect are input to a classification model 592, which is a trained learning model, as feature quantities 65 for the streak defect to be classified. As a result, data indicating a classification destination 67 corresponding to the combination of these three feature quantities 65 is output from classification model 592.

[0088] In this embodiment, streak defects are classified into four classes. That is, so-called "multi-class classification" is performed. Methods for performing multi-class classification using a support vector machine include a one-to-one method and a one-to-many method. Although there are no particular limitations on the method that can actually be adopted, the one-to-one method is adopted in this embodiment. Therefore, in order to achieve classification into four classes, six discrimination planes are obtained by learning using a support vector machine as described above (step S150 in FIG. 8).

[0089] On the premise that six discriminant planes are required, when the feature quantity 65 of the streak defect to be classified is input to the classification model 592, it is determined whether the streak defect should be classified into the first class or the second class based on the discriminant plane separating the first class from the second class, it is determined whether the streak defect should be classified into the first class or the third class based on the discriminant plane separating the first class from the third class, and it is determined whether the streak defect should be classified into the first class or the third class based on the discriminant plane separating the first class from the fourth class. Based on the discriminant plane separating the second and third classes, it is determined whether the streak defect should be classified into the second or fourth class. Based on the discriminant plane separating the second and fourth classes, it is determined whether the streak defect should be classified into the second or fourth class. Based on the discriminant plane separating the second and fourth classes, it is determined whether the streak defect should be classified into the third or fourth class. Based on the six determination results obtained in this manner, the classification destination is determined by majority vote. Note that if the classification destination cannot be determined by majority vote, it can be determined by a method called the maximum discriminant function method, in which linear discriminant functions equal to the number of classes (four linear discriminant functions in this embodiment) are prepared and the defect is classified into the class with the largest discriminant function value. As described above, a known method may be used as a specific method for multi-class classification.

[0090] <6. Effects> According to this embodiment, when a streak defect is included in a printed image of a test chart 70 for detecting streak defects, three feature quantities 65 (peak average luminance value 652, streak width 654, and variation 656 of average luminance value 63 around the streak defect) that represent the characteristics of the streak defect are calculated. Then, by inputting these three feature quantities 65 into a classification model 592, which is a trained learning model for classifying streak defects, a classification destination 67 according to the characteristics of the streak defect is obtained. According to human visual perception, the greater the variation in brightness (luminance value) around a streak, the less likely the streak is perceived. In other words, the degree of defect perceived by humans regarding a streak included in a printed image depends on the variation in luminance value around the streak. In this regard, this embodiment classifies the streak defect taking into account the variation 656 of average luminance value 63 around the streak defect, thereby achieving results similar to those obtained by visual classification. Furthermore, since a trained learning model is used for classification, streak defects contained in printed images are automatically classified. As described above, according to this embodiment, it is possible to automatically classify streak defects contained in printed images so as to obtain results close to those obtained by visual classification. Furthermore, since it is possible to take appropriate measures when streak defects are detected, for example, wasteful consumption of printing paper 5 and ink due to unnecessary reprinting can be reduced. In this way, it is possible to contribute to the achievement of the SDGs (Sustainable Development Goals).

[0091] Here, we will explain the results of an experiment comparing a conventional method with the method of this embodiment. In this experiment, streak defects were classified into three levels: "medium, small, and non-streak." The conventional method classified streak defects based on the results of comparing the peak average luminance value with two thresholds. In this experiment, the inspection objects were visually classified by an expert, and the relationship between the peak average luminance value and the frequency for each classification was as shown in FIG. 25. In FIG. 25, the thick solid line labeled 801 represents the relationship between the peak average luminance value and the frequency for inspection objects classified as "large," the thick dotted line labeled 802 represents the relationship between the peak average luminance value and the frequency for inspection objects classified as "medium," and the solid line labeled 803 represents the relationship between the peak average luminance value and the frequency for inspection objects classified as "non-streak." For the conventional method, two thresholds for classification were determined based on the above relationships to maximize the accuracy rate. Specifically, two values ​​corresponding to the two dotted lines denoted by reference numerals 811 and 812 in FIG. 25 were adopted as threshold values ​​for classification.

[0092] The results of the experiment are shown in Table 1. In this experiment, data was collected for five different days, and cross-validation was performed by dividing the data into four parts. [Table 1]

[0093] According to Table 1, the accuracy rate for the method of this embodiment is higher than that of the conventional method for all ink colors. Furthermore, the accuracy rate for all gradations is higher for the method of this embodiment than for the conventional method. The overall accuracy rate for the method of this embodiment is 4.1% higher than that of the conventional method. In this way, by adopting the method of this embodiment, results closer to visual classification can be obtained than with conventional methods.

[0094] <7. Variations> Modifications of the above embodiment will now be described.

[0095] <7.1 First Modification> In the above embodiment, the peak average luminance value 652, the streak width 654, and the variation 656 of the average luminance value 63 around the streak defect are used as feature quantities 65 for classifying the streak defect. However, the present invention is not limited to this, and a configuration that does not use the streak width 654 may also be employed. In other words, a configuration may be employed in which the streak defect is classified based on two feature quantities 65 (the peak average luminance value 652 and the variation 656 of the average luminance value 63 around the streak defect).

[0096] In this modification, in step S150 of FIG. 8, as shown in FIG. 26, two feature quantities 65 (peak average luminance value 652 and variance 656) and a label 66 indicating a classification destination are input as training data to the training unit 580. Then, based on the training data, the training unit 580 uses a support vector machine to train the relationship between the combination of the two feature quantities 65 and the classification destination. As a result, as in the above embodiment, six parameter sets PR(1) to PR(6) for specifying six discrimination planes are output from the training unit 580, and a classification model 592 is generated. Also, in step S190 of FIG. 8, two feature quantities 65 are input to the classification model 592 as shown in FIG. 27. Then, data indicating a classification destination 67 corresponding to the combination of the two feature quantities 65 is output from the classification model 592.

[0097] <7.2 Second Modification> In the above embodiment, a method using a support vector machine is adopted as the machine learning method. However, the present invention is not limited to this. For example, a method using a neural network (including a convolutional neural network) or a method using a random forest can also be adopted. Therefore, a method using a general feedforward neural network will be described as a second modified example.

[0098] FIG. 28 is a diagram showing an example of the structure of a neural network 81 used in this modification. This neural network 81 is a general forward propagation neural network and is composed of an input layer, a hidden layer (intermediate layer), and an output layer. The input layer is composed of units (i.e., three units) equal in number to the number of feature quantities 65. The number of units in the hidden layer is not particularly limited. In the example shown in FIG. 28, the number of hidden layers is one, but the number of hidden layers may be two or more. The output layer is composed of four units that output four pieces of probability data 82(1) to 82(4) representing the probability that each streak defect will be classified into each of the four classes (first to fourth classes) described above. The connections between the input layer and the hidden layer and between the hidden layer and the output layer are fully connected. For example, a sigmoid function is used as the activation function of the hidden layer. For the activation function of the output layer, a softmax function is used to make the sum of the probability data 82(1) to 82(4) output from this neural network 81 equal to one.

[0099] During learning using this neural network 81, a peak average brightness value 652, a streak width 654, and a variation 656 of the average brightness value 63 around the streak defect are provided to the input layer as feature quantities 65. This causes forward propagation processing within the neural network 81, and a cross-entropy error, for example, is calculated based on probability data 82(1) to 82(4) and correct answer data 83(1) to 83(4) output from the output layer (see FIG. 29). For example, a cross-entropy error is calculated based on the probability data and correct answer data as shown in FIG. 30. As shown in FIG. 30, the probability data 82(1) to 82(4) output from the output layer are data between 0 and 1. Furthermore, as shown in FIG. 30, the correct answer data 83(1) to 83(4) are data of 1 or 0. For example, when learning is performed using data on streak defects that should be classified into the third class, only correct answer data 83(3) is 1, and correct answer data 83(1) to 83(2) and 83(4) are 0. The cross-entropy error is calculated in the above manner, and the parameters (weighting coefficients, bias) of neural network 81 are updated so that the cross-entropy error is minimized. By repeating learning in the above manner, the parameters are optimized.

[0100] When classifying streak defects using this neural network 81, a peak average brightness value 652, a streak width 654, and a variation 656 of the average brightness value 63 around the streak defect are provided to the input layer as feature quantities 65 of the streak defect to be classified. Then, a forward propagation process is performed within the neural network 81, and probability data 82(1) to 82(4) are output from the output layer. The class corresponding to the maximum value of these probability data 82(1) to 82(4) is determined as the classification destination for the streak defect to be classified.

[0101] <8.Other> In the above embodiment (including the modified example), an inkjet printing apparatus 10 that performs color printing is used. However, the present invention is not limited to this, and an inkjet printing apparatus that performs monochrome printing may also be used. In this case, however, the test chart for detecting streak defects is composed of, for example, a 100% solid image area, an 80% solid image area, and a 60% solid image area for only the color K.

[0102] Furthermore, in the above-described embodiment (including the modified examples), an inkjet printing apparatus 10 using aqueous ink was employed. However, the present invention is not limited to this, and an inkjet printing apparatus using UV ink (ultraviolet-curable ink), such as an inkjet printing apparatus for label printing, may also be employed. In this case, an ultraviolet irradiation mechanism that cures the UV ink on the printing paper 5 by irradiating it with ultraviolet light is provided inside the printing mechanism 201 (see FIG. 2) instead of the drying mechanism 206.

[0103] <9. Notes> From the above disclosure, a streak defect classification system having the following configuration can also be considered.

[0104] 1. A streak defect classification system for classifying streak defects contained in a printed image, comprising: a computer having a processor; Memory for storing programs Equipped with When the program stored in the memory is executed by the processor, the program causes the processor to perform the following operations (A), (B), and (C): (A) Based on image data consisting of a plurality of brightness values ​​obtained by capturing an image of a printed image of a test chart for detecting streak defects, an average brightness value is calculated for each of a plurality of consecutive pixel positions in a second direction perpendicular to the first direction in which the streak defects extend, which is the average value of the brightness values ​​of a plurality of pixels at the same pixel position in the second direction. (B) A feature quantity representing a feature of the streak defect is obtained based on the image data or the average brightness value for each of the plurality of pixel positions. (C) The feature amounts are input to a trained learning model for classifying the streak defects, thereby determining a classification destination according to the feature amounts. a maximum value of the average luminance values ​​corresponding to the streak defect and a variation of the average luminance values ​​around the streak defect are obtained as the feature amounts; the local maximum value and the variation are input as the feature amounts to the trained learning model. [Explanation of symbols]

[0105] 10...Inkjet printing device 40...Contact image sensor (CIS) 50...Control unit 65...Features 67…Category destination 70...Test chart 100...printing control device 200...printing machine body 201...Printing mechanism 205...Recording Department 550...Average brightness value calculation unit 560...Streak detection unit 570...Feature calculation unit 572...Peak average luminance value extraction unit 574...Strip width calculation section 576...Variation calculation unit 580…Study Department 590...Classification destination determination section 592...Classification Model 652...Peak average brightness value 654…Strip width 656...Variation (variation of average brightness value around streak defects)

Claims

1. A streak defect classification method for classifying streak defects contained in a printed image, comprising: a test chart printing step of printing a test chart for detecting streak defects; an imaging step of imaging a print image obtained in the test chart printing step; an average luminance value calculation step of calculating, for each of a plurality of pixel positions consecutive in a second direction perpendicular to the first direction in which the streak defect extends, an average luminance value which is an average value of luminance values ​​of a plurality of pixels at the same pixel position in the second direction, based on imaging data consisting of a plurality of luminance values ​​obtained in the imaging step; a feature value calculation step of calculating a feature value representing a feature of the streak defect based on the imaging data or the average luminance value for each of the plurality of pixel positions; a classification destination determination step of determining a classification destination according to the feature amounts by inputting the feature amounts into a trained learning model for classifying the streak-like defects; Including, The feature amount calculation step includes: a maximum value extraction step of extracting a maximum value corresponding to the streak defect from the average luminance values ​​for each of the plurality of pixel positions; a variation calculation step of calculating a variation in the average brightness value around the streak defect; Including, a classification step of determining a classification destination for each of the streak defects, the classification destination determining step including inputting the maximum value and the variation as the feature amounts to the trained learning model;

2. the feature amount calculation step further includes a streak width calculation step of calculating a streak width, which is a width of the streak-like defect, 2. The streak defect classification method according to claim 1, wherein in the classification destination determination step, the streak width is further input as the feature amount to the trained learning model.

3. 3. The streak defect classification method according to claim 2, wherein in the streak width calculating step, two pixel positions corresponding to two minimum values ​​sandwiching the maximum value corresponding to the streak defect among a plurality of minimum values ​​extracted from the average luminance value for each of the plurality of pixel positions are obtained, and a value proportional to a standard deviation of an approximation curve obtained by fitting a relationship between pixel positions and average luminance values ​​to a Gaussian function based on the plurality of average luminance values ​​between the two pixel positions is calculated as the streak width.

4. 2. The streak defect classification method according to claim 1, wherein in the variation calculation step, a first predetermined distance is set to a distance less than a second predetermined distance, and a standard deviation of the average brightness values ​​in a range equal to or greater than the first predetermined distance and equal to or less than the second predetermined distance from the center of the streak defect in the second direction is calculated as the variation.

5. The streak defect classification method according to claim 1 , wherein the trained learning model is a support vector machine.

6. 6. The streak defect classification method according to claim 1, further comprising a streak detection step of detecting the streak defect based on the imaging data or the average brightness value for each of the plurality of pixel positions.

7. 7. The streak defect classification method according to claim 6, wherein in the streak detection step, an image at a pixel position corresponding to a maximum value equal to or greater than a predetermined threshold value among a plurality of maximum values ​​extracted from the average brightness value for each of the plurality of pixel positions is detected as the streak defect.

8. A streak defect classification method for classifying streak defects contained in a printed image, comprising: a first test chart printing step of printing a first test chart for detecting streak defects; a first imaging step of imaging a print image obtained in the first test chart printing step; a first average luminance value calculation step of calculating, for each of a plurality of pixel positions consecutive in a second direction perpendicular to a first direction in which the streak defect extends, a first average luminance value which is an average value of luminance values ​​of a plurality of pixels at the same pixel position in the second direction, based on first imaging data consisting of a plurality of luminance values ​​obtained in the first imaging step; a first streak detection step of detecting the streak defect based on the first imaging data or the first average luminance value for each of the plurality of pixel positions; a first maximum value extraction step of extracting, as a first maximum value, a maximum value corresponding to the streak defect detected in the first streak detection step from the first average luminance values ​​for each of the plurality of pixel positions; a first variation calculation step of calculating a variation in the first average brightness value around the streak defect detected in the first streak detection step as a first variation; a classification destination designation step in which an operator designates a classification destination corresponding to a combination of the first maximum value and the first variation; a learning step of using the first maximum value, the first variation, and the classification destination designated in the classification destination designation step as learning data to cause a learning model to learn the relationship between the combination of the first maximum value and the first variation and the classification destination; a second test chart printing step of printing a second test chart for detecting streak defects; a second imaging step of imaging a print image obtained in the second test chart printing step; a second average luminance value calculation step of calculating, for each of the plurality of pixel positions, a second average luminance value that is an average value of luminance values ​​of a plurality of pixels that are at the same pixel position in the second direction, based on second imaging data consisting of a plurality of luminance values ​​obtained in the second imaging step; a second streak detection step of detecting the streak defect based on the second imaging data or the second average luminance value for each of the plurality of pixel positions; a second maximum value extraction step of extracting, as a second maximum value, a maximum value corresponding to the streak defect detected in the second streak detection step from the second average luminance values ​​for each of the plurality of pixel positions; a second variation calculation step of calculating a variation of the second average brightness value around the streak defect detected in the second streak detection step as a second variation; a classification destination determination step of inputting the second maximum value and the second variation into the learning model that has been learned by the learning step, and determining a classification destination according to the combination of the second maximum value and the second variation; A method for classifying streak defects, comprising:

9. 1. A streak defect classification system for classifying streak defects contained in a printed image, comprising: an average luminance value calculation unit that calculates an average luminance value, which is an average value of luminance values ​​of a plurality of pixels at the same pixel position in a second direction perpendicular to a first direction in which the streak defect extends, for each of a plurality of pixel positions that are continuous in the second direction, based on imaging data consisting of a plurality of luminance values ​​obtained by imaging a printed image of a test chart for detecting streak defects; a feature amount calculation unit that calculates a feature amount representing a feature of the streak defect based on the imaging data or the average luminance value for each of the plurality of pixel positions; a classification destination determination unit that determines a classification destination according to the feature amount by inputting the feature amount into a trained learning model for classifying the streak-like defect; Equipped with The feature amount calculation unit a maximum value extracting unit that extracts a maximum value corresponding to the streak defect from the average luminance value for each of the plurality of pixel positions; a variation calculation unit that calculates the variation of the average brightness value around the streak defect; Including, a learning model that has been trained and that receives as input the maximum value and the variation as the feature quantities;

10. the feature amount calculation unit includes a streak width calculation unit that calculates a streak width that is a width of the streak-like defect, 10. The streak defect classification system according to claim 9, wherein the streak width is further input as the feature amount to the trained learning model.

11. 11. The streak defect classification system according to claim 10, wherein the streak width calculation unit determines two pixel positions corresponding to two minimum values ​​sandwiching the maximum value corresponding to the streak defect among a plurality of minimum values ​​extracted from the average luminance values ​​for each of the plurality of pixel positions, and calculates, as the streak width, a value proportional to a standard deviation of an approximation curve obtained by fitting a relationship between pixel positions and average luminance values ​​to a Gaussian function based on the plurality of average luminance values ​​between the two pixel positions.

12. 10. The streak defect classification system according to claim 9, wherein the variation calculation unit calculates, as the variation, a standard deviation of the average brightness values ​​in a range equal to or greater than the first predetermined distance and equal to or less than the second predetermined distance from the center of the streak defect in the second direction, with a first predetermined distance being a distance less than a second predetermined distance.

13. The streak defect classification system according to claim 9 , wherein the trained learning model is a support vector machine.

14. 14. The streak defect classification system according to claim 9, further comprising a streak detection unit that detects the streak defect based on the imaging data or the average brightness value for each of the plurality of pixel positions.

15. 15. The streak defect classification system according to claim 14, wherein the streak detection unit detects, as the streak defect, an image of a pixel position corresponding to a maximum value equal to or greater than a predetermined threshold value among a plurality of maximum values ​​extracted from the average brightness value for each of the plurality of pixel positions.

Images