Image recognition-based digital printing ink break and fly ink flaw real-time identification system

Abstract

The application relates to the technical field of image recognition, and particularly discloses a digital printing ink break and ink flying flaw real-time identification system based on image recognition. The system collects a printing image sequence and extracts synchronous jet ejection waveform data, generates a theoretical printing image based on ejection parameters. The actual image and the theoretical image are compared pixel by pixel to obtain a residual image sequence, time domain characteristic parameters of the gray residuals of each pixel are extracted, the ink break flaw is determined according to the linear regression slope, and the ink flying flaw is determined according to the frequency domain correlation of the main frequency component and the ejection frequency. Further, the residual image is used to reverse the actual ejection state of the jet orifice, the generation parameters of the subsequent theoretical image are dynamically corrected, and online accurate detection under a fixed template-free scene is realized.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically to a real-time identification system for ink breakage and ink splatter defects in digital printing based on image recognition. Background Technology

[0002] In digital printing, ink skipping and ink splatter are two common printing defects. Ink skipping manifests as missing lines or insufficient ink volume due to interruption of the ink path, while ink splatter manifests as random ink dots or streaks formed when ink droplets deviate from their intended landing point during flight. In personalized printing scenarios, the patterns on the substrate surface contain a large number of discontinuous design elements, such as brushstrokes, intentional breaks, and irregular ink dots. The grayscale distribution of these design elements at the pixel level is highly similar to that of ink skipping and ink splatter defects.

[0003] Existing image recognition-based defect detection methods typically compare the acquired actual printed image with a standard template image, locating defect areas through grayscale differences. However, in personalized printing, there are no fixed standard template images; the printed content itself is unique and non-repeatable, making it impossible to distinguish design elements from actual defects through template comparison. Some solutions attempt to directly identify defect areas through image segmentation or feature extraction, but since design elements and defects are essentially the same in terms of grayscale values, edge shapes, and texture features, the algorithm cannot define the normal design boundaries of the pattern. This easily leads to misjudging design elements such as ink splatter and breaks as ink skipping defects, resulting in batch missed detections and false alarms. Summary of the Invention

[0004] The purpose of this invention is to provide a real-time identification system for ink breakage and ink splatter defects in digital printing based on image recognition, thereby solving the above-mentioned technical problems.

[0005] The objective of this invention can be achieved through the following technical solutions: A real-time digital printing ink skipping and splashing defect recognition system based on image recognition includes: The acquisition module acquires a sequence of printed images from the surface of the substrate during the online production process of the digital printing press. The frame rate of the acquired printed image sequence is synchronized with the running speed of the digital printing press. Extract the nozzle ejection waveform data corresponding to each frame of the printed image. The nozzle ejection waveform data includes the theoretical ejection state of each nozzle at the moment of print image acquisition. Generation module: Generates a theoretical printing image sequence synchronized with the printing image sequence based on the nozzle jet waveform data; Feature extraction module: Performs pixel-level grayscale value subtraction between the printed image and the theoretical printed image at the same time to obtain the residual image, and the residual images at all times form a residual image sequence; Extract the temporal feature parameters of the grayscale residual value of each pixel position in the residual image sequence as a function of time; Defect identification module: Determines ink breakage defects and ink splatter defects based on time-domain feature parameters.

[0006] As a further aspect of the present invention: the process of extracting the nozzle jet waveform data is as follows: A timestamp mapping table is established between the printing image acquisition time and the printhead ejection waveform data output time in the printhead control system. The timestamp mapping table records the storage address of the printhead ejection waveform data output at the corresponding time, using the printing image acquisition time as an index. A one-to-one correspondence is established between each frame of printed image and the printhead ejection waveform data through the timestamp mapping table. The printhead ejection waveform data is read from the storage address. The printhead ejection waveform data is stored in the form of a binary bitmap. Each binary bit in the binary bitmap corresponds to a nozzle of the printhead. A binary bit value of 1 indicates that the corresponding nozzle is in the ejection state at the printing image acquisition time, and a binary bit value of 0 indicates that the corresponding nozzle is in the non-ejection state at the printing image acquisition time. For nozzles with a binary bit value of 1, the ejection voltage amplitude, ejection duration, and ejection droplet volume of the corresponding nozzle at the printing image acquisition time are read from the printhead control system.

[0007] As a further aspect of the present invention: the process of obtaining the theoretical printing image sequence is as follows: The radius of the base of the spherical cap is calculated by inversely using the formulas for the volume of the ejected ink droplet and the volume of the spherical cap. Twice the radius of the base is taken as the diameter of the ink dot formed by the ink droplet on the substrate surface. The ejection kinetic energy of the ink droplet is obtained based on the ejection voltage amplitude and ejection duration. The maximum spreading radius during the ink droplet spreading process is calculated based on the ejection kinetic energy, the viscosity parameters and surface tension parameters of the ink droplet. The difference between the maximum spreading radius and the ink dot diameter is divided by the ink dot diameter to obtain the edge diffusion gradient. A circular area is defined according to the ink dot diameter, with the projection position of the nozzle with a binary bit value of 1 on the substrate surface as the center point. Within the circular area, pixel gray values ​​are assigned according to the edge diffusion gradient. The gray value of the pixel at the center point of the circular area is equal to the theoretical gray value. The theoretical gray value is the gray value of the center point of the ink dot formed on the substrate surface when the nozzle ejects ink droplets according to the ejected ink droplet volume. The pixel gray values ​​within the circular area decrease pixel by pixel along the edge diffusion gradient from the center point to the edge. The gray values ​​of all pixels within the circular area are filled into a blank image according to their pixel coordinates to obtain the theoretical printed image. The above process is repeated for each printed image acquisition time to obtain a theoretical printed image sequence.

[0008] As a further aspect of the present invention, the process of obtaining the residual image is as follows: The mapping relationship between the pixel coordinate system of the industrial camera and the physical coordinate system of the printhead is obtained. Based on the mapping relationship, the correspondence between the pixel coordinates of each pixel in the printed image and the corresponding pixel coordinates of the same physical position in the theoretical printed image is established. All pixel coordinates in the actual printed image are traversed, and the corresponding pixel coordinates of the same physical position in the theoretical printed image are found according to the correspondence. The gray value of the pixel in the printed image is subtracted from the gray value of the corresponding pixel in the theoretical printed image to obtain the gray value residual. The gray value residuals of all pixel positions are arranged according to the pixel coordinates of the actual printed image to form a residual image.

[0009] As a further aspect of the present invention, the process of extracting time-domain feature parameters is as follows: The grayscale residual values ​​of the same pixel coordinate position in multiple consecutive residual images are extracted from the residual image sequence to form a time series. The time series is fitted with linear regression to obtain the linear regression slope. The time series is then subjected to Fourier transform to obtain the frequency domain amplitude spectrum. The dominant frequency component and the amplitude corresponding to the dominant frequency component are extracted from the frequency domain amplitude spectrum. The linear regression slope, the dominant frequency component, and the amplitude corresponding to the dominant frequency component constitute the time domain feature parameters.

[0010] As a further aspect of the present invention: the process for determining ink breakage defects is as follows: The linear regression slope is extracted from the temporal feature parameters. The linear regression slope is compared with a preset slope threshold. When the linear regression slope is greater than the preset slope threshold, the pixel position corresponding to the linear regression slope is marked as a candidate pixel for ink breakage. Connectivity analysis is performed on the candidate pixels to obtain the candidate connected regions for ink breakage. The average linear regression slope of all pixels in the candidate connected regions for ink breakage is extracted. When the average value is greater than the preset connected region slope threshold, the candidate connected regions for ink breakage are judged as ink breakage defects.

[0011] As a further aspect of the present invention: the process for determining ink splatter defects is as follows: Extract the dominant frequency component from the time-domain feature parameters, obtain the jet frequency of the corresponding nozzle in the nozzle jet waveform data, calculate the frequency ratio of the dominant frequency component to the jet frequency, and mark the pixel position corresponding to the dominant frequency component as a candidate pixel for ink spraying when the frequency ratio is within the preset ratio range and the amplitude corresponding to the dominant frequency component is greater than the preset amplitude threshold. Perform connected component analysis on the candidate pixels to obtain the candidate connected regions for ink spraying, and determine the candidate connected regions for ink spraying as ink spraying defects.

[0012] As a further aspect of the present invention: after obtaining a preset number of residual images, the following steps are performed to obtain the theoretical printing image: Based on the projection position of the nozzle on the substrate surface, the grayscale residual values ​​of the corresponding region in each frame of the residual image sequence are extracted from the residual image sequence. The grayscale residual values ​​of the corresponding region in multiple consecutive frames of the residual image are then combined to form a grayscale residual time series corresponding to the nozzle. The average value of all grayscale residual values ​​in the grayscale residual time series is calculated. Multiplying the average value by the ejected ink droplet volume of the nozzle and then dividing by the theoretical grayscale value yields the effective ejected ink droplet volume correction. The geometric center coordinates of the grayscale residual values ​​within the region corresponding to the nozzle are extracted from each frame of the residual image. The ink droplet landing position offset is obtained by subtracting the average value of the geometric center coordinates in the multi-frame residual images from the projection position of the nozzle on the substrate surface. The gray-level residual gradient sequence from the center point to the edge in the region corresponding to the nozzle is extracted from each frame of the residual image. The average value of the gray-level residual gradient sequence in the consecutive frames of residual images is subtracted from the edge diffusion gradient to obtain the ink droplet edge spreading correction. The printhead ejection waveform data is corrected based on the effective ejected ink droplet volume correction, the ink droplet landing position offset, and the ink droplet edge spreading correction to obtain the theoretical printed image.

[0013] The beneficial effects of this invention are: This invention extracts printhead jet waveform data synchronized with the printing image acquisition time from the printhead control system. Based on physical parameters such as the jet voltage amplitude, jet duration, and jet droplet volume of each nozzle, a theoretical printed image is generated. This theoretical printed image serves as a comparison benchmark for dynamic matching with the printed content, enabling pixel-by-pixel comparison between the actual printed image and the theoretical expected image in scenarios without a fixed template. By calculating the residual image between the actual and theoretical printed images, the temporal feature parameters of the grayscale residual at each pixel location are extracted. Utilizing the physical law that the grayscale residual of ink breakage defects exhibits a monotonically increasing trend during continuous printing, and the physical law that the oscillation frequency of the grayscale residual of ink splatter defects has a frequency domain correlation with the printhead jet waveform frequency, these temporal feature parameters are used as criteria to distinguish defect types from design elements. This invention does not rely on a preset standard template image and can adapt to online detection of any personalized printed content. By introducing physical parameters from the printhead drive end as prior information, it fundamentally solves the technical problem of the difficulty in distinguishing design elements from defects at the image level. Meanwhile, by inverting the residual image, the effective jet droplet volume correction, droplet landing position offset, and droplet edge spreading correction are obtained. These corrections are then fed back into the generation process of the theoretical printed image, enabling the dynamic updating of the theoretical printed image as the actual jetting state of the nozzle, thus improving the real-time accuracy of the comparison benchmark. Attached Figure Description

[0014] The invention will now be further described with reference to the accompanying drawings.

[0015] Figure 1This is a flowchart illustrating the real-time identification system for ink breakage and ink splatter defects in digital printing based on image recognition, as per the present invention. Figure 2 This is a schematic diagram of the process for obtaining theoretical printing images according to the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figures 1-2 As shown, this invention is a real-time identification system for ink skipping and ink splatter defects in digital printing based on image recognition, comprising: The acquisition module acquires a sequence of printed images from the surface of the substrate during the online production process of the digital printing press. The frame rate of the acquired printed image sequence is synchronized with the running speed of the digital printing press. Specifically, an industrial camera is mounted above the conveyor belt of the digital printing press, transporting the printing substrate. The trigger signal input of the industrial camera is connected to the rotary encoder of the printing press. Each pulse output by the rotary encoder represents a 0.1 mm movement of the printing substrate. The industrial camera exposes once every 20 pulses received, and continuous exposures form a sequence of printed images. In the printed image sequence, two adjacent frames correspond to a 2 mm movement of the printing substrate. The exposure time of the industrial camera is set to 0.5 milliseconds. When the printing press operates at a speed of 200 mm / s, the frame rate of the printed image sequence is 100 frames per second. The industrial camera transmits the printed image sequence via a gigabit Ethernet interface and simultaneously connects to the printhead controller via a serial interface to receive printhead jet waveform data synchronized with the printed image set.

[0018] Extract the nozzle ejection waveform data corresponding to each frame of the printed image. The nozzle ejection waveform data includes the theoretical ejection state of each nozzle at the moment of print image acquisition. In a preferred embodiment of the present invention, the process of extracting the nozzle spray waveform data is as follows: A circular buffer is set up inside the nozzle controller, divided into 512 storage units, each with a capacity of 4 kilobytes. At the end of each spray control cycle, the nozzle controller writes the nozzle spray waveform data of the current cycle into the circular buffer, and simultaneously stores a record composed of the timestamp of the writing time and the storage address in a mapping table. At the exposure time, the industrial camera simultaneously sends a synchronization pulse signal to the nozzle controller. When the nozzle controller receives the synchronization pulse signal, it reads the current system clock value and uses it as the index position for writing the printed image acquisition time into the mapping table. The industrial camera and the nozzle controller are connected via a PCIe bus. After receiving a frame of printed image from the industrial camera, the acquisition time corresponding to that frame is extracted from the additional information field of the printed image. This acquisition time is used as an index to look up the corresponding storage address in the mapping table. For example, when the acquisition time is 12345678 microseconds, the record with index value 12345678 is located in the mapping table, and the storage address 300 is read from this record. The nozzle spray waveform data is then read from the 300th storage unit in the circular buffer. The printhead ejection waveform data consists of 1280 binary bits, each corresponding to a nozzle of the printhead. The printhead model is a specific type of piezoelectric printhead, with nozzle numbers ranging from 1 to 1280. The binary bit arrangement is consistent with the nozzle numbering order; the first binary bit corresponds to nozzle number 1, the second to nozzle number 2, and so on. A binary bit value of 1 indicates that the corresponding nozzle is in ejection mode at the time of image acquisition, while a binary bit value of 0 indicates that the corresponding nozzle is in non-ejection mode. For nozzles with a binary bit value of 1, the ejection parameters for that nozzle at the time of image acquisition are further read from the parameter register of the printhead controller. The parameter register stores parameters according to the nozzle number. For example, for nozzle number 15 with a binary bit value of 1, the storage area corresponding to nozzle number 15 in the parameter register is accessed to read the ejection voltage amplitude as 24 volts, the ejection duration as 35 microseconds, and the ejected ink droplet volume as 6 picoliters. The nozzle number, jet voltage amplitude, jet duration, and jet droplet volume corresponding to all the read nozzles with a binary bit value of 1 are combined with the binary bit map to form complete printhead jet waveform data, which is stored in memory and associated with the printed image of the corresponding frame.

[0019] Generation module: Generates a theoretical printing image sequence synchronized with the printing image sequence based on the nozzle jet waveform data; In a preferred embodiment of the present invention, the process of obtaining the theoretical printing image sequence is as follows: Extract the ejected ink droplet volume V corresponding to the nozzle with a binary bit value of 1 from the printhead ejection waveform data. The unit of the ejected ink droplet volume V is picoliter. Obtain the contact angle θ between the ink droplet and the substrate surface. The contact angle θ is determined by the surface energy of the ink and the substrate material; in this embodiment, θ is set to 60 degrees. Calculate the diameter D of the ink droplet formed on the substrate surface using the spherical cap volume formula and the contact angle θ. The spherical cap volume formula is: ; Understandably, after an ink droplet lands on the substrate, it contracts into a spherical shape due to surface tension. The spherical cap is a portion of a sphere truncated by a plane, and its volume is uniquely determined by its height and base radius. This formula originates from the integral calculation of the volume of a solid of revolution in solid geometry, essentially treating the ink droplet as a stable form of an ideal incompressible fluid reaching static equilibrium on a solid surface; where h is the height of the spherical cap, and r is the base radius. According to the geometric relationship of the contact angle θ, the height h and the base radius r of the spherical cap satisfy: ; Understandably, the contact angle θ is the angle formed at the contact point between the solid-liquid interface and the gas-liquid interface. For a spherical droplet, the angle between the line connecting the center of the droplet to the edge of the bottom surface and the vertical direction is equal to θ. Through geometric derivation, the ratio of the height h of the droplet to the radius r of the bottom surface is tangent to the half-contact angle. This relationship links the macroscopically measurable contact angle with the microscopic geometric dimensions of the droplet.

[0020] Substituting θ = 60 degrees into the equation, tan(30°) ≈ 0.577, we get h = 0.577r. Substituting h = 0.577r into the formula for the volume of a spherical cap, we get: ; Therefore, the solution is... The solution r is taken as the base radius of the spherical cap, and twice the base radius r is taken as the diameter D of the ink dot, that is, D=2r; For example, when the volume of the ejected ink droplet V is equal to 6 picoliters, the calculated r is approximately 1.82 picoliters. When converting picoliters to millimeters and multiplying by a preset conversion factor such as 0.01, r is approximately equal to 0.0182 millimeters, and the ink droplet diameter D is equal to 0.0364 millimeters.

[0021] Extract the jet voltage amplitude U and jet duration T corresponding to the nozzle with a binary bit value of 1 from the printhead jet waveform data. Obtain the electromechanical coupling coefficient K of the piezoelectric ceramic, which is determined by the piezoelectric ceramic material and is measured in Newtons per volt. Multiply the jet voltage amplitude U by the electromechanical coupling coefficient K to obtain the thrust F generated by the piezoelectric ceramic during jetting (the piezoelectric ceramic deforms under voltage, and the deformation is linearly related to the applied voltage, with the proportionality constant being the piezoelectric constant. The thrust of the piezoelectric ceramic is equal to the deformation multiplied by the elastic modulus of the ceramic material, which is then summed to the electromechanical coupling coefficient K multiplied by the voltage U). Multiply the thrust F by the jet duration T to obtain the impulse I exerted by the piezoelectric ceramic on the ink droplet (when the thrust F remains constant during the jet duration T, the impulse is equal to the force multiplied by the time. This formula comes from Newton's second law). The integral form of the law describes the measure of the change in momentum of an object by force. The mass *m* of the ink droplet is obtained by multiplying the volume *V* of the ejected droplet by the ink density *p*, where *p* is 1.05 grams per cubic centimeter (density *p* is the mass per unit volume; the droplet mass equals density multiplied by volume. This formula is the definition of density and applies to uniform fluids). According to the impulse theorem, dividing the impulse *I* by the droplet mass *m* yields the ejection velocity *v*. According to the kinetic energy formula, multiplying the droplet mass *m* by the square of the ejection velocity *v* and then dividing by 2 yields the ejection kinetic energy *E* of the droplet. k (Kinetic energy is defined as the energy possessed by an object due to its motion, and is equal to half the product of its mass and the square of its velocity. This formula is the fundamental definition of kinetic energy in classical mechanics and is derived from Newtonian mechanics.)

[0022] For example, when the jet voltage amplitude U equals 24 volts, the jet duration T equals 35 microseconds, and the jet droplet volume V equals 6 picoliters, the piezoelectric electromechanical coupling coefficient K is 0.05 N / V. The calculated thrust F equals 24 multiplied by 0.05 equals 1.2 N, the impulse I equals 1.2 multiplied by 0.000035 equals 0.000042 N / s, the droplet mass m equals 1.05 multiplied by 6 multiplied by 10⁻⁹ approximately equals 6.3 multiplied by 10⁻⁹ kg, the jet velocity v equals 0.000042 divided by 6.3 multiplied by 10⁻⁹ approximately equals 6667 m / s, and the jet kinetic energy E... k It equals 6.3 multiplied by 10 to the power of -9 multiplied by 6667 squared and divided by 2, which is approximately 0.14 millijoules.

[0023] Obtain the viscosity parameter *u* and surface tension parameter *y* of the ink droplet. The viscosity parameter *u* is set to 0.01 Pa·s, and the surface tension parameter *y* is set to 0.03 N·m. Calculate the maximum spreading radius *Rmax* during the ink droplet spreading process based on the jet kinetic energy *Ek*, viscosity parameter *u*, and surface tension parameter *y*. Assume that the kinetic energy of the ink droplet after impacting the substrate surface is converted into the work done to overcome viscous resistance and the work done to increase surface energy. The work done to overcome viscous resistance is equal to the integral term *K* of the spreading velocity. vMultiply by the viscosity parameter u, multiply by π, multiply by the square of the maximum spreading radius Rmax, and the integral term K for the spreading rate. v Taking an empirical value of 0.5, the work done by increasing surface energy is equal to the surface tension parameter y multiplied by n multiplied by the square of the maximum spreading radius Rmax (in brackets) minus the square of the ink dot radius r (in brackets). The maximum spreading radius Rmax is obtained by iteratively solving the energy balance equation during the spreading process. The energy balance equation is: ; Where K v It is worth noting that after an ink droplet impacts the substrate surface, part of its initial kinetic energy is dissipated by viscosity, and part is converted into an increase in surface energy. The viscous dissipation power is proportional to the viscosity *u* and the rate of change of the spreading area, and approximates approximately 0.5 after integrating over time. , where k v For the integral term of the spreading speed, an empirical value is taken as 0.5 in this embodiment. The increase in surface energy is equal to the surface tension y multiplied by the area increment, where the area increases from the initial base area r of the spherical cap. 2 The final spread area becomes πRmax 2 This equation is based on the principle of energy conservation.

[0024] Dividing the difference between the maximum spreading radius Rmax and the radius r corresponding to the ink dot diameter D by r yields the edge diffusion gradient G, i.e.: ; It is important to note that the edge diffusion gradient is defined as the growth rate of the maximum radius of the ink droplet after it spreads relative to the original radius. This value quantifies the degree of blurring at the edge of the ink dot and is used for subsequent grayscale allocation.

[0025] For example, when the ink dot radius r is 0.0182 mm and the jet kinetic energy Ek is 0.14 mJ, the maximum spreading radius Rmax is 0.052 mm, and the edge diffusion gradient G is approximately 1.86 (0.052 minus 0.0182 divided by 0.0182).

[0026] The theoretical grayscale value L is calculated based on the jet droplet volume V. Calibration data is pre-stored, recording the center grayscale values ​​of ink dots corresponding to different jet droplet volumes. The calibration process uses the same ink and substrate as in production, and the actual center grayscale values ​​of ink dots are measured using an optical microscope. The theoretical grayscale value L is obtained by linear interpolation of the current jet droplet volume V within the calibration data. For example, in the calibration data, the center grayscale value of an ink dot is 120 when the jet droplet volume is 5 picoliters, and 180 when the jet droplet volume is 7 picoliters. When the current jet droplet volume V is 6 picoliters, the theoretical grayscale value L is calculated to be 150 through linear interpolation.

[0027] Obtain the pixel resolution of the industrial camera, which is 0.02 mm per pixel. Divide the ink dot diameter D by the pixel resolution to obtain the number of pixels corresponding to ink dot diameter D. Using the projection position of the nozzle with a binary bit value of 1 on the substrate surface as the center point, delineate a circular region according to the ink dot diameter D. The boundary of the circular region is determined by adding or subtracting the number of pixels corresponding to the ink dot radius from the center point coordinates. Set the grayscale value of the pixel at the center point within the circular region to equal the theoretical grayscale value L. Set the grayscale value of the pixels within the circular region to decrease pixel by pixel along the edge diffusion gradient G from the center point to the edge. For a pixel within the circular region at a distance d pixels from the center point, calculate the grayscale value of that pixel as: ; Where R is the pixel radius corresponding to the ink dot diameter D, and d ranges from 0 to R. When d equals R, the grayscale value is 0. It should be noted that it is assumed that the grayscale value at the center of the ink dot is the highest, L, and the grayscale value decreases linearly along the radial direction, with the deceleration rate controlled by the edge diffusion gradient G. When the distance from the center d = R, the grayscale value drops to 0. This formula is a simplified linear grayscale model used to simulate the optical response of the ink distribution in an actual ink dot decreasing as the radius increases.

[0028] The grayscale values ​​of all pixels within the circular area are filled into a blank image according to their pixel coordinates to obtain the theoretical printing image. This process is repeated for each printing image acquisition time, and all theoretical printing images are arranged in chronological order to obtain a theoretical printing image sequence.

[0029] It is important to note that during digital printing, ink droplets ejected from the nozzle undergo three physical stages on the substrate surface: deposition, spreading, and solidification. The final shape and grayscale distribution of the resulting ink droplets are uniquely determined by the nozzle's driving parameters and the ink's physical properties. By reading the volume of the ejected ink droplets from the printhead control system, the radius of the droplet's base and the diameter of the droplet can be calculated based on a spherical cap geometry model. This is because, under the influence of surface tension, the droplet spontaneously contracts into a spherical cap shape, and there is a definite geometric relationship between the base radius and volume of the cap. The ejection kinetic energy acquired by the droplet can be derived from the ejection voltage amplitude and ejection duration. This is because the electromechanical coupling of the piezoelectric ceramic converts electrical energy into mechanical energy. The initial kinetic energy acquired by the droplet after being ejected from the nozzle determines its velocity upon impacting the substrate surface, and the impact kinetic energy further determines the degree of spread of the droplet on the substrate surface, including the maximum spreading radius and the edge diffusion gradient. When the ink droplet spreading radius exceeds the original ink dot radius, the ink volume gradually decreases from the center to the edge, forming an optical distribution where grayscale decreases from the center to the edge. The edge diffusion gradient is the key parameter for quantifying this decreasing rate of decrease. The theoretical printed image generated according to the above physical relationship is not based on an arbitrary template or statistical model, but completely reproduces the ink dot shape and grayscale distribution that the printhead should produce under the current driving parameters. By comparing this theoretical image with the actual acquired printed image pixel by pixel, the temporal variation characteristics of grayscale deviation in the residual image can truly reflect whether there is ink interruption or ink splattering abnormality in the nozzle. This is because ink interruption will cause the ink volume to continuously decrease, resulting in a monotonically increasing grayscale residual, while ink splattering will cause the ink droplet landing point to randomly shift, resulting in the grayscale residual exhibiting oscillations related to the printhead driving frequency. Therefore, this method of generating theoretical printed images based on physical deposition simulation fundamentally solves the problem of indistinguishable design elements from real defects in scenarios without a fixed template. It makes defect identification no longer dependent on the manual design of image features, but directly uses the prior physical information of the printhead driving end as the judgment benchmark.

[0030] Feature extraction module: Performs pixel-level grayscale value subtraction between the printed image and the theoretical printed image at the same time to obtain the residual image, and the residual images at all times form a residual image sequence; In a preferred embodiment of the present invention, the process of obtaining the residual image is as follows: Joint calibration of the printhead and camera is performed beforehand. A calibration board is placed on the substrate surface. The calibration board is printed with a grid of dots with a spacing of 5 mm. The printhead ejects ink droplets onto the calibration board to form an array of ink dots, while the industrial camera captures an image of the calibration board. The coordinates of each ink dot in the ink dot array in the printhead's physical coordinate system are extracted. For example, the physical coordinates of nozzle 1 are set to (0, 0), nozzle 2 to (0.5, 0), and so on. The pixel coordinates of each ink dot are extracted from the calibration board image. For example, the pixel coordinates of ink dot 1 in the image are (100, 200), and the pixel coordinates of ink dot 2 are (250, 200). The physical coordinates of each nozzle and the pixel coordinates of the corresponding ink dot are paired to form a mapping point. An affine transformation matrix is ​​fitted using the least squares method. The affine transformation matrix includes rotation, translation, and scaling components. The reciprocal of the scaling component is the pixel resolution of the industrial camera, 0.02 mm per pixel. The affine transformation matrix is ​​stored in memory as a mapping relationship between the pixel coordinate system of the industrial camera and the physical coordinate system of the nozzle.

[0031] The mapping relationship is read from memory. For each pixel coordinate in the printed image, the pixel coordinate is multiplied by the inverse of the affine transformation matrix to obtain the corresponding printhead physical coordinate. Then, based on the physical coordinate, the corresponding pixel coordinate at the same physical location is found in the theoretical printed image. For example, for a pixel with pixel coordinates (200, 150) in the printed image, its physical coordinates are calculated as (4, 3) by substituting it into the inverse of the affine transformation matrix. In the theoretical printed image, all pixel coordinates are traversed, and the corresponding physical coordinate is calculated for each pixel coordinate. The pixel coordinates (4, 3) are found, and assumed to be (400, 300). The pixel coordinates (200, 150) in the printed image are recorded as corresponding to the pixel coordinates (400, 300) in the theoretical printed image. This correspondence is established for each pixel coordinate in the printed image, and all correspondences are stored as a mapping table.

[0032] The process reads the printed image and the theoretical printed image from memory. It iterates through each pixel coordinate in the printed image, looking up the corresponding pixel coordinate in the theoretical printed image from a mapping table. The grayscale value of that pixel location in the printed image is read; for example, the grayscale value of pixel coordinates (200, 150) in the printed image is 80. The grayscale value of the corresponding pixel coordinate (400, 300) in the theoretical printed image is read; for example, the theoretical grayscale value is 150. The grayscale value of the theoretical image is subtracted from the grayscale value of 80 in the printed image to obtain a grayscale residual value of -70. The grayscale residual values ​​of all pixel locations are arranged according to the pixel coordinates of the printed image. A blank image of the same size as the printed image is created in memory, and the grayscale residual value of each pixel location is filled into the same coordinate position in the blank image, forming a residual image.

[0033] Extract the temporal feature parameters of the grayscale residual value of each pixel position in the residual image sequence as a function of time; In another preferred embodiment of the present invention, the process of extracting time-domain feature parameters is as follows: A residual image queue is maintained in memory, with a queue length of 50 frames. Each new residual image frame is pushed to the tail of the queue, and the image at the head of the queue is removed. A fixed pixel coordinate position is selected in the residual image, for example, pixel coordinates (200, 200). Gray-level residual values ​​are read from the same coordinate position in each frame of the residual image queue and arranged in chronological order to form a time series, for example, a time series of -70, -65, -60, -55, and -50 (5 values). Linear regression is performed on the time series, using the frame number as the independent variable and the gray-level residual value as the dependent variable, and the slope of the regression line is calculated; for example, the slope of the above sequence is 5. A Fourier transform is performed on the time series to convert the gray-level residual values ​​in the time domain into a frequency domain amplitude spectrum. For example, performing a Fourier transform on -70, -65, -60, -55, and -50 yields a frequency domain amplitude spectrum, with the amplitudes of the frequency components in the spectrum being 0.5, 0.2, 0.1, etc. Find the frequency component with the largest amplitude in the amplitude spectrum as the dominant frequency component, and record its corresponding amplitude. For example, if the dominant frequency component is 0.1 Hz, the corresponding amplitude is 0.5. Combine the linear regression slope of 5, the dominant frequency component of 0.1 Hz, and the corresponding amplitude of 0.5 into a vector, and use this vector as the temporal feature parameter of the pixel coordinate position. Repeat the above process for each pixel coordinate position in the residual image to obtain the temporal feature parameter corresponding to each pixel position.

[0034] Defect identification module: Determines ink breakage defects and ink splatter defects based on time-domain feature parameters.

[0035] In another preferred embodiment of the present invention, the process of determining ink breakage defects is as follows: The temporal feature parameters of each pixel location are read from memory, including the linear regression slope. A slope threshold is pre-set, for example, to 3. All pixel locations are traversed, and the linear regression slope of each pixel location is extracted. The linear regression slope is compared with the slope threshold of 3. When the linear regression slope of a pixel location is greater than 3, that pixel location is marked as a candidate pixel for ink breakage. For example, the linear regression slope of a pixel at coordinates (200, 200) is 5, which is greater than 3, so it is marked as a candidate pixel for ink breakage. The linear regression slope of another pixel at coordinates (200, 201) is 2, which is less than 3, so it is not marked. After traversal, a binary labeled image is obtained, where the value of labeled pixels is 1, and the value of unlabeled pixels is 0.

[0036] Connectivity analysis is performed on the binary labeled image using the eight-neighborhood connectivity rule, meaning that adjacent pixels in eight directions (up, down, left, right, and diagonal) belong to the same connected region. The binary labeled image is scanned, and all pixels with a value of 1 that are interconnected are grouped into the same connected region. The set of pixel coordinates contained in each connected region is recorded. For example, if one connected region contains 3 pixels (200, 200), (200, 201), and (201, 200), and another connected region contains 4 pixels (300, 300), (300, 301), (301, 300), and (301, 301), these connected regions are considered candidate connected regions for ink breakage.

[0037] Extract the linear regression slopes of all pixels within each candidate connected region of ink breakage and calculate the average of these slopes. For example, the linear regression slopes of the three pixels in the first connected region are 5, 4, and 6, with an average of 5. The linear regression slopes of the four pixels in the second connected region are 4, 3, 4, and 5, with an average of 4. A connected region slope threshold is pre-set, for example, 3.5. The average slope of each connected region is compared with the connected region slope threshold of 3.5. When the average is greater than 3.5, the connected region is identified as an ink breakage defect. The first connected region has an average slope of 5 greater than 3.5, so it is identified as an ink breakage defect. The second connected region also has an average slope of 4 greater than 3.5, so it is identified as an ink breakage defect. Output the location and contour information of all ink breakage defects for subsequent alarm or shutdown processing.

[0038] The essence of ink breakage defects is that the volume of ejected ink droplets continuously decreases due to nozzle blockage or insufficient ink supply. During continuous printing, the grayscale value of the actual ink droplets gradually falls below the theoretical grayscale value, and this deviation accumulates and expands with the increase of printing frames. Therefore, the grayscale residual value at the corresponding position in the residual image shows a monotonically increasing trend, and the slope of linear regression can quantify this increasing rate. When the slope exceeds a preset threshold, it indicates that there is a continuously deteriorating ink loss at that position. Connected component analysis aggregates spatially adjacent abnormal pixels into regions and calculates the average slope of the regions, which can eliminate random noise interference from individual pixels. Only when the entire region shows a consistent increasing trend can it be confirmed as an ink breakage defect.

[0039] In another preferred embodiment of the present invention, the process of determining ink splatter defects is as follows: The temporal feature parameters of each pixel location are read from memory. These parameters include the dominant frequency component and its corresponding amplitude. The nozzle ejection waveform data is read from the nozzle control system to obtain the ejection frequency for each nozzle orifice; for example, the ejection frequency of a certain type of nozzle is set to 10 kHz. All pixel locations are traversed, and the dominant frequency component of each pixel location is extracted. The frequency ratio of the dominant frequency component to the nozzle ejection frequency is calculated. For example, the dominant frequency component of a pixel at coordinates (200, 200) is 10 kHz, so the frequency ratio is 10 kHz divided by 10 kHz, which equals 1. The dominant frequency component of another pixel at coordinates (200, 201) is 20 kHz, so the frequency ratio is 2. A preset ratio range is set, for example, between 0.9 and 1.1. An amplitude threshold is also set, for example, to 0.8. It is then determined whether the frequency ratio of each pixel location is within the range of 0.9 to 1.1, and whether the amplitude of the dominant frequency component at that pixel location is greater than 0.8. When both conditions are met, the pixel location is marked as a candidate pixel for the ink-spraying algorithm. For example, if the frequency ratio of pixel coordinates (200, 200) is 1 and within the preset ratio range, and its dominant frequency component has an amplitude of 0.9, which is greater than 0.8, it is marked as a candidate pixel for the ink-spraying algorithm. If the frequency ratio of pixel coordinates (200, 201) is 2 and not within the preset ratio range, it is not marked even if its amplitude is greater than 0.8. If the frequency ratio of pixel coordinates (201, 200) is 1 and within the preset ratio range, but its amplitude is 0.6, which is less than 0.8, it is also not marked.

[0040] After traversal, a binary labeled image of candidate ink particles is obtained. Connectivity analysis is performed on this image, using four-neighbor or eight-neighbor connectivity rules to group all interconnected candidate ink particles into the same connected region, resulting in candidate connected regions. For example, a connected region containing three pixels (200, 200), (200, 201), and (201, 200) is found. These candidate connected regions are directly identified as ink defects, and the location coordinates, contour range, and corresponding nozzle number of each defect are output. The ink defect information is displayed on the human-machine interface and recorded in the production log file.

[0041] The essence of ink splatter defects is that ink droplets deviate from their intended landing point during flight due to airflow interference or electrostatic effects. These deviated droplets randomly land in an area around the expected location, and the ink droplets generated by splatter are usually synchronized with the printhead ejection cycle, as splatter can occur with each ejection. Therefore, the grayscale residual in the residual image corresponding to splatter will oscillate with the printhead ejection frequency as the dominant frequency. The ratio of the dominant frequency component to the nozzle ejection frequency in the time-domain feature parameters is close to 1. At the same time, the grayscale fluctuation amplitude generated by splatter is large, and the amplitude corresponding to the dominant frequency component will exceed the normal fluctuation range. By filtering based on both frequency ratio and amplitude, splatter defects can be distinguished from random noise or other types of defects. Connected component analysis ensures the spatial continuity of the splatter region, avoiding misclassification of isolated pixels as splatter. The above-mentioned judgment method directly utilizes the different temporal characteristics of ink breakage and ink splatter in terms of physical mechanism, so that defect identification no longer relies on the static analysis of ink dot morphology in the image, but distinguishes design elements from real defects from the dynamic change law. Because design elements such as ink splatter and breakage are part of the printed content, they remain stable in the continuous printing process and will not show monotonous increase of grayscale residual or oscillation related to the jetting frequency. Therefore, this method can accurately identify defects and avoid misjudgment.

[0042] In another preferred embodiment of the present invention, after obtaining a preset number of residual images, the following steps are performed to obtain a theoretical printing image: After accumulating 50 frames of residual images in memory, the correction process is initiated. The printhead configuration information is read to obtain the projection coordinates of each nozzle on the substrate surface; for example, the projection position of nozzle 1 is (0, 0), and the projection position of nozzle 2 is (0.5, 0). Based on the industrial camera's pixel resolution of 0.02, the projection position of each nozzle is converted into image coordinates; for example, the image coordinates corresponding to nozzle 1 are (0, 0), and the image coordinates corresponding to nozzle 2 are (25, 0). A square region with a side length of 20 pixels (or other values) is defined as the corresponding area for each nozzle, centered on its image coordinates.

[0043] The algorithm iterates through the corresponding region of each nozzle, extracting the grayscale residual values ​​of all pixels within that region from 50 frames of residual images. For nozzle 1, the algorithm extracts the grayscale residual values ​​of all pixels within the corresponding region from the first frame of the residual image and calculates the average value to -70. The average value is then extracted from the second frame to obtain -65, and so on until the 50th frame, where the average value is -20. These 50 average values ​​are arranged chronologically to form a grayscale residual time series. The average value of all values ​​in this time series is calculated; for example, the average of the 50 values ​​is -45. The ejected droplet volume of nozzle 1 is read as 6, and the theoretical grayscale value is read as 150. The average grayscale residual value -45 is multiplied by the ejected droplet volume 6 and then divided by the theoretical grayscale value 150 to obtain an effective ejected droplet volume correction of -1.8, meaning the actual ejected droplet volume is 1.8 smaller than the theoretical value.

[0044] The geometric center coordinates of the grayscale residual values ​​within the region corresponding to nozzle 1 are extracted from each frame of the residual image. The geometric center coordinates are calculated by weighting the pixel coordinates using the grayscale residual values ​​of all pixels within the region. For example, the geometric center of the grayscale residual values ​​in this region is (0.2, 0.1) in frame 1, (0.3, 0.1) in frame 2, and so on up to (0.2, 0.2) in frame 50. The average geometric center coordinates across 50 frames are calculated to be (0.23, 0.13). The difference between the average geometric center coordinates (0.23, 0.13) and the image coordinates (0, 0) of the nozzle projection position is calculated to obtain the ink droplet landing point offset as (0.23, 0.13).

[0045] Extract the grayscale residual gradient sequence from the center point to the edge within the region corresponding to nozzle 1 in each frame of residual image. Starting from the geometric center, take a sampling point every 1 coordinate unit along the radial direction. Calculate the difference between the grayscale residual value of this sampling point and the previous sampling point, divide by the interval to obtain the gradient value. Average all gradient values ​​in the radial direction to obtain the average gradient sequence of the frame. For example, the average gradient sequence of frame 1 is [-5, -4, -3, -2, -11], frame 2 is [-6, -4, -3, -2, -1], and so on up to frame 50, which is [-5, -4, -3, -2, -1]. Calculate the average gradient sequence of 50 frames to obtain the average value at each gradient position, for example, the average gradient sequence is [-5.1, -4.0, -3.0, -2.0, -1.0]. The edge diffusion gradient of the nozzle is read as 1.86. The theoretical gradient sequence corresponding to the edge diffusion gradient is generated by the theoretical grayscale value 150 and the edge diffusion gradient 1.86 using a grayscale value allocation formula. The theoretical gradient sequence is [-7.5, -5.0, -3.0, -1.5, 0.1]. The average gradient sequence is subtracted from the theoretical gradient sequence at corresponding positions to obtain the difference at each gradient position. The average of these differences is then taken as the ink droplet edge spreading correction amount. For example, the difference is [2.4, 1.0, 0, -0.5, -1.0], and the average value is 0.38.

[0046] The effective droplet volume correction factor of -1.8, the droplet landing position offset of (0.23, 0.13), and the droplet edge spreading correction factor of 0.38 are combined to form the ejection state correction parameters. Based on these parameters, the printhead ejection waveform data is corrected: the droplet volume of this nozzle in subsequent time steps is adjusted from 6 to 4.2; the edge diffusion gradient used in generating the theoretical printed image in subsequent time steps is adjusted from 1.86 to 1.86 plus the droplet edge spreading correction factor of 0.38, i.e., 2.24; the coordinates of the center point of the ink dot corresponding to this nozzle in the theoretical printed image in subsequent time steps are adjusted from the original nozzle projection position (0, 0) to (0.23, 0.13) after adding the landing position offset. Based on the corrected droplet volume of 4.2, the corrected edge diffusion gradient of 2.24, and the corrected center point coordinates of the ink dot (0.23, 0.13), the theoretical printed image generation process is re-executed to obtain the theoretical printed image. The above process is repeated for each nozzle to achieve dynamic updating of the theoretical printed image according to the actual ejection state of the nozzle.

[0047] It should be noted that after obtaining multiple consecutive frames of residual images, the effective jet droplet volume correction is inferred by using the average grayscale residual, the droplet landing position offset is inferred by using the offset of the grayscale residual geometric center, and the droplet edge spreading correction is inferred by using the difference between the grayscale residual radial gradient and the theoretical gradient sequence. Essentially, this decomposes the deviation between the actual and theoretical jetting states of the nozzle contained in the residual image into three independent components: volume deviation, landing position deviation, and spreading deviation. The grayscale residual value in the residual image is proportional to the difference between the actual and theoretical ink droplet grayscale values. Since the ink droplet grayscale and the jet droplet volume have an approximately linear relationship in the central region of the ink droplet, multiplying the average grayscale residual by the jet droplet volume and then dividing by the theoretical grayscale value can restore the absolute value of the volume deviation. The geometric center coordinates of the grayscale residual value reflect the offset direction and distance of the actual ink droplet center relative to the nozzle projection position. Because after the ink droplet landing point shifts, the area with the largest negative residual in the residual image—that is, the region where the actual grayscale is most severely lower than the theoretical grayscale—will shift synchronously, and its weighting center corresponds to the landing point position. The radial gradient sequence of the grayscale residual characterizes the difference between the actual ink droplet edge grayscale decay rate and the theoretical decay rate. This difference directly originates from the deviation of the maximum spreading radius from the theoretical value during ink droplet spreading. Therefore, the average gradient difference can be used as a correction amount for the edge diffusion gradient. The volume correction is directly superimposed on the jetting ink droplet volume, the landing point offset is superimposed on the ink droplet center coordinates, and the spreading correction is superimposed on the edge diffusion gradient. Then, the theoretical printing image is regenerated, ensuring that the ink volume, position, and edge blurring degree of each ink droplet in the newly generated theoretical printing image are consistent with the actual jetting behavior of the current nozzle. Only when the theoretical printed image remains synchronized with the current actual spraying state of the printhead can the residual between the subsequently acquired actual printed image and the theoretical printed image eliminate systematic deviations caused by non-defect factors such as nozzle wear, ink viscosity changes, and temperature drift. This ensures that the grayscale residuals retained in the residual image truly correspond to abnormal spraying events such as ink skipping or ink splatter. This closed-loop dynamic update mechanism ensures that the theoretical benchmark will not become invalid due to parameter drift caused by long-term printhead operation, thus continuously providing a reliable comparison benchmark in personalized printing scenarios without fixed templates.

[0048] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A real-time identification system for ink skipping and ink splatter defects in digital printing based on image recognition, characterized in that, include: The acquisition module acquires a sequence of printed images from the surface of the substrate during the online production process of the digital printing press. The frame rate of the acquired printed image sequence is synchronized with the running speed of the digital printing press. Extract the nozzle ejection waveform data corresponding to each frame of the printed image. The nozzle ejection waveform data includes the theoretical ejection state of each nozzle at the moment of print image acquisition. Generation module: Generates a theoretical printing image sequence synchronized with the printing image sequence based on the nozzle jet waveform data; Feature extraction module: Performs pixel-level grayscale value subtraction between the printed image and the theoretical printed image at the same time to obtain the residual image, and the residual images at all times form a residual image sequence; Extract the temporal feature parameters of the grayscale residual value of each pixel position in the residual image sequence as a function of time; Defect identification module: Determines ink breakage defects and ink splatter defects based on time-domain feature parameters.

2. The real-time identification system for ink skipping and ink splatter defects in digital printing based on image recognition according to claim 1, characterized in that, The process of extracting the nozzle spray waveform data is as follows: A timestamp mapping table is established between the printing image acquisition time and the printhead ejection waveform data output time in the printhead control system. The timestamp mapping table records the storage address of the printhead ejection waveform data output at the corresponding time, using the printing image acquisition time as an index. A one-to-one correspondence is established between each frame of printed image and the printhead ejection waveform data through the timestamp mapping table. The printhead ejection waveform data is read from the storage address. The printhead ejection waveform data is stored in the form of a binary bitmap. Each binary bit in the binary bitmap corresponds to a nozzle of the printhead. A binary bit value of 1 indicates that the corresponding nozzle is in the ejection state at the printing image acquisition time, and a binary bit value of 0 indicates that the corresponding nozzle is in the non-ejection state at the printing image acquisition time. For nozzles with a binary bit value of 1, the ejection voltage amplitude, ejection duration, and ejection droplet volume of the corresponding nozzle at the printing image acquisition time are read from the printhead control system.

3. The real-time identification system for ink skipping and ink splatter defects in digital printing based on image recognition according to claim 1, characterized in that, The process of obtaining the theoretical printing image sequence is as follows: Based on the formula for the volume of the jet ink droplet and the volume of the spherical cap, the radius of the bottom surface of the spherical cap is calculated, and twice the radius of the bottom surface is taken as the diameter of the ink droplet formed on the surface of the substrate. The ejection kinetic energy of the ink droplets is obtained based on the ejection voltage amplitude and ejection duration. The maximum spreading radius during the ink droplet spreading process is calculated based on the ejection kinetic energy, the viscosity parameters, and the surface tension parameters of the ink droplets. The difference between the maximum spreading radius and the ink droplet diameter is divided by the ink droplet diameter to obtain the edge diffusion gradient. Taking the projection position of the nozzle with a binary bit value of 1 on the substrate surface as the center point, a circular area is delineated according to the ink droplet diameter. Within the circular area, pixel gray values ​​are assigned according to the edge diffusion gradient. The gray value of the pixel at the center point of the circular area is equal to the theoretical gray value. The theoretical gray value is the gray value of the ink droplet center point formed on the substrate surface when the nozzle ejects ink droplets according to the ejection volume. The pixel gray values ​​within the circular area decrease pixel by pixel along the edge diffusion gradient from the center point to the edge. The gray values ​​of all pixels within the circular area are filled into a blank image according to the pixel coordinate position to obtain the theoretical printed image. The above process is repeated for each printed image acquisition time to obtain the theoretical printed image sequence.

4. The real-time identification system for ink skipping and ink splatter defects in digital printing based on image recognition according to claim 1, characterized in that, The process of obtaining the residual image is as follows: The mapping relationship between the pixel coordinate system of the industrial camera and the physical coordinate system of the printhead is obtained. Based on the mapping relationship, the correspondence between the pixel coordinates of each pixel in the printed image and the corresponding pixel coordinates of the same physical position in the theoretical printed image is established. All pixel coordinates in the actual printed image are traversed, and the corresponding pixel coordinates of the same physical position in the theoretical printed image are found according to the correspondence. The gray value of the pixel in the printed image is subtracted from the gray value of the corresponding pixel in the theoretical printed image to obtain the gray value residual. The gray value residuals of all pixel positions are arranged according to the pixel coordinates of the actual printed image to form a residual image.

5. The real-time identification system for ink skipping and ink splatter defects in digital printing based on image recognition according to claim 1, characterized in that, The process of extracting time-domain feature parameters is as follows: The grayscale residual values ​​of the same pixel coordinate position in multiple consecutive residual images are extracted from the residual image sequence to form a time series. The time series is fitted with linear regression to obtain the linear regression slope. The time series is then subjected to Fourier transform to obtain the frequency domain amplitude spectrum. The dominant frequency component and the amplitude corresponding to the dominant frequency component are extracted from the frequency domain amplitude spectrum. The linear regression slope, the dominant frequency component, and the amplitude corresponding to the dominant frequency component constitute the time domain feature parameters.

6. The real-time identification system for ink skipping and ink splatter defects in digital printing based on image recognition according to claim 1, characterized in that, The process for determining ink breakage defects is as follows: The linear regression slope is extracted from the temporal feature parameters. The linear regression slope is compared with a preset slope threshold. When the linear regression slope is greater than the preset slope threshold, the pixel position corresponding to the linear regression slope is marked as a candidate pixel for ink breakage. Connectivity analysis is performed on the candidate pixels to obtain the candidate connected regions for ink breakage. The average linear regression slope of all pixels in the candidate connected regions for ink breakage is extracted. When the average value is greater than the preset connected region slope threshold, the candidate connected regions for ink breakage are judged as ink breakage defects.

7. The real-time identification system for ink skipping and ink splatter defects in digital printing based on image recognition according to claim 1, characterized in that, The process for determining ink splatter defects is as follows: Extract the dominant frequency component from the time-domain feature parameters, obtain the jet frequency of the corresponding nozzle in the nozzle jet waveform data, calculate the frequency ratio of the dominant frequency component to the jet frequency, and mark the pixel position corresponding to the dominant frequency component as a candidate pixel for ink spraying when the frequency ratio is within the preset ratio range and the amplitude corresponding to the dominant frequency component is greater than the preset amplitude threshold. Perform connected component analysis on the candidate pixels to obtain the candidate connected regions for ink spraying, and determine the candidate connected regions for ink spraying as ink spraying defects.

8. The real-time identification system for ink skipping and ink splatter defects in digital printing based on image recognition according to claim 3, characterized in that, After obtaining a preset number of residual images, perform the following steps to obtain the theoretical printing image: Based on the projection position of the nozzle on the substrate surface, the grayscale residual value of the region corresponding to the nozzle in each frame of the residual image sequence is extracted from the residual image sequence. The grayscale residual values ​​of the region corresponding to the nozzle in multiple consecutive frames of residual images are combined to form the grayscale residual time series corresponding to the nozzle. The average value of all grayscale residual values ​​in the grayscale residual time series is calculated. The average value is multiplied by the volume of the ejected ink droplets of the nozzle and then divided by the theoretical grayscale value to obtain the effective ejected ink droplet volume correction amount. Extract the geometric center coordinates of the grayscale residual values ​​in the region corresponding to the nozzle from each frame of the residual image. Calculate the difference between the average value of the geometric center coordinates in multiple consecutive frames of residual images and the projection position of the nozzle on the substrate surface to obtain the offset of the ink droplet landing point. Extract the grayscale residual gradient sequence from the center point to the edge within the region corresponding to the nozzle from each frame of residual image. Perform a difference operation between the average value of the grayscale residual gradient sequence in multiple consecutive frames of residual images and the edge diffusion gradient to obtain the ink droplet edge spreading correction amount. Correct the printhead ejection waveform data and obtain the theoretical printing image based on the effective ejected ink droplet volume correction amount, the ink droplet landing point offset amount, and the ink droplet edge spreading correction amount.

Images