Intelligent digital positioning printing method and system for high-low cut tufted carpet

Abstract

The application relates to a high-low cut tufted carpet intelligent digital positioning printing method and system, and relates to the technical fields of digital printing and computer vision. The scheme comprises the following steps: acquiring a carpet elevation map and a texture image through a 3D structured light sensor and a two-dimensional camera, constructing a dynamic physical geometric model containing pile lodging information by using a deep convolutional neural network; extracting a demarcation contour line based on the model, and generating a digital compensation pattern for offsetting light and shadow distortion by using a reverse optical rendering algorithm in combination with virtual light source parameters; driving a nozzle to perform positioning printing and realizing closed-loop quality inspection through online monitoring. The application can effectively solve the problems of visual discontinuity and light and shadow distortion caused by pile height difference and random lodging, and can create a three-dimensional light and shadow effect with physical reality.

Description

Technical Field

[0001] This invention relates to the field of digital printing and textile fabric pattern processing technology, specifically to an intelligent digital positioning printing method and system for tufted carpets with a double-layer three-dimensional pile structure of high-cut pile and low-cut pile. Background Technology

[0002] In the digital printing process of tufted carpets, the pile structure of the raw carpet directly affects the positioning accuracy of the pattern and the final artistic expression. For computer jacquard tufted carpets with a double-layer three-dimensional structure of high-cut pile and low-cut pile, the carpet surface is composed of piles of different heights. The junction of the two pile surfaces forms a three-dimensional outline. During digital printing, it is necessary to accurately position and print in different areas according to the difference in pile height in order to present the layered relationship between the foreground and background patterns.

[0003] Currently, the industry's digital positioning printing technology for high-low tufted carpets mainly uses dual cameras to create two-dimensional images of the carpet surface from different angles. By analyzing the grayscale or color difference characteristics of the two-dimensional images, the height variation of the pile is indirectly inferred, and boundary lines are generated accordingly for zone printing. However, this solution has revealed the following problems in actual production: Recognition accuracy is limited by two-dimensional information: For complex areas with smooth transitions between high and low pile, localized pile collapse, or uneven lighting, the difference in reflected light in two-dimensional images is extremely small, making the system prone to misidentification or omission. This causes the extracted boundary lines to deviate from the true contours, ultimately resulting in problems such as printed floral patterns and misprints.

[0004] Poor adaptability and long production changeover cycle: When changing to a new pattern or when there are differences in the characteristics of carpet pile between different batches, it is necessary to manually perform long-term image parameter calibration and sample-specific training, which is cumbersome and seriously restricts production efficiency.

[0005] The 3D effect is stiff and rigid: the existing projection scheme that simply adds 1-3 fixed offset pixels along the dividing line does not take into account factors such as the actual pile height difference, pile density and falling direction. The resulting projection lacks physical realism and limits the development of design techniques and artistic expression.

[0006] Lack of online quality closed-loop control: The entire process focuses on identification and positioning before printing, but lacks real-time, online monitoring and compensation methods for defects that may occur during printing, such as nozzle clogging, ink splatter, and color difference. This makes it difficult to meet the stringent requirements of mid-to-high-end products for low defect rates. Summary of the Invention

[0007] This application provides a method and system for intelligent digital positioning printing of high and low cut tufted carpets, which can solve the technical problems of mispositioning, visual discontinuity and light and shadow distortion caused by the inability to obtain accurate physical height and pile collapse information when facing high and low cut transition areas.

[0008] To achieve the above objectives, this application provides the following technical solution: The first aspect of this application provides a method for intelligent digital positioning printing of high and low cut tufted carpets, comprising the following steps: S1. Use a 3D structured light contour sensor to obtain an elevation map of the carpet surface during movement, and use a 2D camera to simultaneously acquire texture images. S2. Input the fused elevation map and texture image into a pre-trained deep convolutional neural network to automatically identify and label the high-cutting area, low-cutting area and transition area. At the same time, construct a dynamic physical geometry model of the pile, including the direction of pile collapse, the degree of collapse and the probability of collapse, for the pixels in the transition area. S3. Based on the dynamic physical geometry model of the pile head, extract the sub-pixel level closed-loop boundary contour line between the high-cut pile area and the low-cut pile area; S4. Receive the virtual light source azimuth and elevation angle set by the user. Based on the dynamic physical geometry model of the pile, the boundary outline and the actual pile height difference on both sides of the outline, use the reverse optical rendering algorithm to dynamically generate a digital compensation pattern to offset the light and shadow distortion caused by the pile height difference and random falling, and to create a three-dimensional light and shadow effect under the virtual light source. Then, superimpose it onto the boundary transition area between the foreground and background patterns. S5 drives the digital printhead to perform positioning and printing based on the partitioned pattern and the complete pattern after superimposed compensation; while printing, the online monitoring camera simultaneously acquires images for quality inspection and performs automatic compensation when defects are detected.

[0009] In an optional embodiment, in step S2, the dynamic physical geometry model of the pile is a probabilistic model, which is constructed by outputting the pile's falling direction, falling angle, and probability entropy value representing the randomness of falling for each pixel in the transition area through a deep convolutional neural network.

[0010] In an optional embodiment, in step S4, the reverse optical rendering algorithm integrates a calculation model of light occlusion, multiple scattering, and penumbra among the multi-pile heads, and renders a digital compensation pattern with spatial gradient transparency and adaptive blurred edges based on the probability distribution in the dynamic physical geometry model of the multi-pile heads.

[0011] In an optional embodiment, in step S2, while identifying the partitions, the deep convolutional neural network assigns a material label with high reflectivity to the high-cut velvet area and a material label with diffuse reflectivity to the low-cut velvet area, and generates differentiated ink volume output instructions to enhance the visual contrast between high and low cuts.

[0012] In an optional embodiment, in step S5, the information on repetitive defects with fixed locations detected by the online monitoring camera is fed back to step S2 to dynamically update the local collapse probability parameter in the dynamic physical geometry model of the pile.

[0013] The second aspect of this application provides an intelligent digital positioning and printing system for high and low cut tufted carpets, which, along the carpet's travel direction, sequentially includes a three-dimensional shape recognition module, a deep learning main control module, a digital printing module, and an online closed-loop quality inspection module, wherein: The deep learning main control module has a built-in fuzzy physical model building unit and a reverse optical rendering unit; The pile physical model construction unit is used to perform the pile dynamic physical geometric model construction described in step S2 of technical solution 1 of this application; The reverse optical rendering unit is used to perform the digital compensation pattern generation described in step S4 of technical solution 1 of this application; The deep learning master control module integrates and outputs complete printing control signals containing digital compensation patterns.

[0014] In an optional embodiment, the reverse optical rendering unit is also used to execute the rendering algorithm described in technical solution 3 of this application, which integrates the ray occlusion, multiple scattering and penumbra calculation models.

[0015] In an optional embodiment, the deep learning master control module is also used to perform the operation described in technical solution 4 of this application, which assigns different material labels to different velvet areas and generates differentiated ink volume instructions.

[0016] In one optional embodiment, a data feedback channel is provided between the online closed-loop quality inspection module and the deep learning main control module to transmit the detected repetitive defect information back to the pile head physical model building unit for dynamic updating of model parameters.

[0017] In one alternative embodiment, the deep convolutional neural network is pre-trained using a dataset of fused 3D and 2D carpet images with light and shadow labels generated by physically based simulation rendering before application.

[0018] This application provides a method and system for intelligent digital positioning printing of high-low cut pile carpets. This scheme employs a 3D structured light contour sensor and a 2D camera to simultaneously acquire elevation maps and texture images of the carpet. The fused data is then input into a pre-trained deep convolutional neural network, enabling automatic identification of high-cut pile areas, low-cut pile areas, and transition zones. More importantly, it constructs a dynamic physical geometry model of the pile, including the direction, degree, and probability of pile collapse. Based on this dynamic physical geometry model, sub-pixel-level closed-loop boundary contours are extracted. Combined with user-defined virtual light source parameters, an inverse optical rendering algorithm integrating light occlusion, multiple scattering, and penumbra calculations is used to dynamically generate a digital compensation pattern that can offset the light and shadow distortion caused by pile height differences and random collapse. This process allows the system to proactively calculate the compensation strategy needed to offset future light and shadow distortions, rather than passively outlining boundaries. This drives the digital printing head to accurately position and print based on the complete pattern containing the compensation pattern. Simultaneously, an online monitoring and feedback mechanism corrects defects and updates model parameters in real time. This effectively solves the problem that traditional two-dimensional imaging schemes cannot perceive the true three-dimensional shape and the state of the pile, resulting in visual breaks and messy shadows at the junction of high and low points. Therefore, it avoids the situation where light and shadow distortion caused by physical structure does not match the design expectations, significantly improving the three-dimensionality, realism, and visual consistency of carpet products in complex lighting environments. It achieves a technological leap from simple image recognition to physical shape modeling and light and shadow reconstruction, ensuring that the finished carpet presents a clean and natural three-dimensional projection effect from different viewing angles.

[0019] In summary, this application constructs a complete closed-loop system from 3D perception and physical modeling to optical reverse compensation, incorporating the mechanical morphology of the pile into the printing compensation model. This achieves logical self-consistency of the technical solution and completeness of system functions, providing a highly adaptive and robust systematic solution for the intelligent production of high and low cut tufted carpets. Attached Figure Description

[0020] Figure 1 A flowchart of a method for intelligent digital positioning printing of high and low cut tufted carpet provided in this application; Figure 2 This is a comparison of the light and shadow effects at the junction of the velour surface before and after using this method (i.e., a comparison before and after printing). Figure 3 A schematic diagram of the intelligent digital positioning and printing system for high and low cut tufted carpets provided in this application. Detailed Implementation

[0021] The technical solutions in the embodiments of this application will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0022] Example 1: In the production of tufted carpets with a double-layered three-dimensional structure of high-cut and low-cut pile areas, enhancing their three-dimensional expression through digital printing while simultaneously concealing the visual discontinuity at the boundary between the two pile heights is a major technical challenge for the industry. Traditional dual-camera two-dimensional imaging recognition solutions can only infer pile height changes through color differences, failing to obtain precise physical height and pile collapse information. Therefore, when facing high-low cut transition areas, misalignment is easily produced, creating harsh color or structural discontinuities at the boundary. While existing technologies have attempted to simulate projection by adding fixed-width color bars, this method completely deviates from the actual pile height difference and collapse state, resulting in a stiff and artificial three-dimensional effect that exacerbates visual disharmony. More importantly, the pile heads of high-cut and low-cut piles will randomly collapse under different external forces, causing chaotic and unrealistic shadows on the carpet surface under natural light or indoor lighting that do not conform to the design expectations. This distortion of light and shadow caused by physical structure is a deep-seated technical problem that traditional pattern compensation methods based on two-dimensional design cannot foresee or solve.

[0023] Based on the above issues, please refer to Figure 1 A flowchart of a method for intelligent digital positioning printing of high and low cut tufted carpets provided in this application embodiment is shown. The method includes the following steps: Step 1: Use a 3D structured light contour sensor to acquire an elevation map of the carpet surface during movement, and use a 2D camera to simultaneously acquire texture images; The 3D structured light profile sensor refers to a device that uses the principle of structured light triangulation to project specifically coded grating stripes onto the carpet surface and reconstructs the three-dimensional shape of the object by analyzing stripe deformation. This sensor specifically includes a light source projection unit and an imaging receiving unit, whose function is to acquire real-time elevation data of the carpet surface with micron-level precision, generating an elevation map describing the absolute height of the carpet pile. The 2D camera refers to an industrial area array camera used to capture information about the carpet surface's color, texture, and macroscopic patterns. The elevation map is obtained by the sensor scanning the carpet line by line while it is moving, while the texture image is obtained by the 2D camera simultaneously capturing images at the same field of view. Both are strictly aligned in timestamps and spatial coordinates for subsequent multimodal fusion processing. For example, when a carpet moves at a speed of 20 meters per minute, the 3D structured light profile sensor acquires point cloud data at a frequency of 10 kHz, generating an elevation matrix with a resolution of up to 0.5 mm, while the 2D camera captures RGB texture images at a matching frame rate. This synchronous acquisition method ensures that each pixel contains both accurate height information and rich color and texture information, providing a complete data foundation for subsequent accurate identification of velvet structures.

[0024] Step 2: Input the fused elevation map and texture image into a pre-trained deep convolutional neural network to automatically identify and label high-cutting areas, low-cutting areas, and transition areas. At the same time, construct a dynamic physical geometry model of the pile, including the direction of pile collapse, degree of collapse, and probability of collapse, for the pixels in the transition area. The deep convolutional neural network is a deep learning model pre-trained on a fused 3D and 2D carpet image dataset with physically simulated rendering labels. Its input is the fused elevation map and texture image, and its output is a partition label and pile physical parameters. High-cut pile areas, low-cut pile areas, and transition areas are different functional regions automatically labeled based on pile height thresholds and gradient changes. The pile dynamic physical geometry model is a probabilistic model used to quantitatively describe the microscopic morphology of the pile in the transition area and its behavioral characteristics under illumination. This model specifically includes three core dimensions: pile lodging direction, lodging degree, and lodging probability. The pile lodging direction can refer to the azimuth angle of the pile tilt (0-360 degrees); the lodging degree can refer to the angle between the pile axis and the vertical normal; and the lodging probability characterizes the uncertainty or random entropy value of the pile morphology at that location. These parameters are obtained through multidimensional regression prediction of each pixel using a neural network. For example, for a row of pixels in the transition zone, the network might output a 45-degree folding direction (to the right and front), a 30-degree folding degree, and a folding probability of 0.8, meaning that there is an 80% probability that the pile at that location will exhibit this state. The folding direction and folding degree are used together to determine the reflection vector of light on the pile surface; the folding probability is used to introduce softened edges in subsequent rendering, avoiding the generation of harsh computational boundaries. By constructing this dynamic physical geometry model, the system can move beyond simple image recognition to understanding the physical morphology of soft fabrics, providing a precise mathematical description for eliminating lighting distortion.

[0025] Step 3: Based on the dynamic physical geometry model of the pile head, extract the sub-pixel level closed-loop boundary contour line between the high-cut pile area and the low-cut pile area; The subpixel-level closed-loop boundary contour line refers to a boundary curve with a precision higher than that of a single pixel, capable of continuously and closedly dividing the high-cut pile area and the low-cut pile area. This contour line is extracted based on the yield gradient and elevation change points in the aforementioned constructed dynamic physical geometry model of the pile. Its function is to provide precise spatial constraints for subsequent reverse optical rendering, ensuring that the compensation pattern only acts on the necessary transition areas without contaminating the main pattern. The extraction process typically employs a gradient-based edge detection algorithm combined with morphological optimization to transform discrete pixel classification results into smooth vector contours. For example, when the elevation difference between adjacent pixels exceeds 1.5mm and the yield direction changes drastically, the system uses an interpolation algorithm to calculate the precise boundary position within the pixel, thereby obtaining a subpixel-level boundary line. This result provides a high-precision geometric benchmark for subsequent digital compensation pattern generation, effectively avoiding printing misalignment or visual discontinuities caused by boundary jaggedness.

[0026] Step 4: Receive the azimuth and elevation angles of the virtual light source set by the user. Based on the dynamic physical geometry model of the pile, the boundary outline, and the actual pile height difference on both sides of the outline, use the reverse optical rendering algorithm to dynamically generate a digital compensation pattern to offset the light and shadow distortion caused by the pile height difference and random falling, and to create a three-dimensional light and shadow effect under the virtual light source. Then, overlay it onto the transition area between the foreground and background patterns. The virtual light source azimuth and elevation angles are ideal lighting parameters set by the user in the software interface according to design requirements, used to define the desired stereoscopic projection direction. The reverse optical rendering algorithm is a physically based lighting transmission simulation algorithm. Its core logic is reverse derivation: given the target visual effect (i.e., the ideal brightness distribution under the virtual light source) and scene geometric properties (i.e., the dynamic physical geometry model of the pile and the actual pile height difference), the algorithm calculates the ink distribution to be printed on the carpet surface. This algorithm integrates calculation models of light occlusion, multiple scattering, and penumbra among multiple piles, enabling the simulation of the interaction between light and complex pile structures. The digital compensation pattern refers to image data with spatially gradient transparency and adaptive blurred edges generated by the above algorithm. Its function is to actively counteract the messy shadows caused by the random falling of piles in the real environment and superimpose a stereoscopic projection that conforms to the virtual light source settings. For example, if the user sets the virtual light source to come from the upper left at 45 degrees, but the actual pile falls to the right, which may cause unexpected dark areas on the right side, the algorithm will calculate that a gradient light-colored compensation pattern needs to be printed in that area to neutralize the real shadows, while generating a natural dark projection on the left side. like Figure 2 As shown in the image, this is a comparison of the lighting and shadow effects at the junction of the textured surfaces before and after using this method. The image reveals that the area on the left, where this method was not used, exhibits obvious cluttered shadows and visual breaks. In contrast, the area on the right, after applying this method, displays a clean, natural, and 3D projection at the junction that aligns with the direction of the virtual light source, completely eliminating lighting distortion. This reverse calculation and compensation mechanism achieves a shift from passively adapting to ambient light to actively shaping lighting and shadow effects.

[0027] Step 5: Drive the digital print head to position and print according to the partition pattern and the complete pattern after superimposed compensation; at the same time, the online monitoring camera simultaneously collects images for quality inspection and performs automatic compensation when defects are found.

[0028] The digital printhead, controlled by printing control signals, is the actuator that sprays the partitioned pattern and the complete pattern superimposed with the digitally compensated pattern onto the carpet surface. Positioning printing refers to the printhead precisely controlling the ink droplet placement based on the carpet's real-time position and the previously extracted boundary contour lines, ensuring perfect alignment of the compensated pattern with the physical boundary. The online monitoring camera, located behind the printing station, synchronously captures images of the printed carpet. Quality inspection compares the captured images with a standard pattern to identify defects such as ink splatter, ink breaks, or positional deviations. Automatic compensation means that upon detecting a defect, the system immediately generates a correction command, driving the printhead to perform reprinting in the next cycle or immediately. For example, when the online monitoring camera detects a missing compensation pattern due to nozzle blockage, the system records the defect coordinates and controls adjacent printheads or performs a second reprint at that location during the return stroke. This significantly improves the consistency and yield of the finished carpet, ensuring that complex three-dimensional light and shadow effects are completely and accurately reproduced on the physical carrier.

[0029] This application constructs a complete closed loop from 3D perception to physical modeling and then to optical inverse compensation through the synergistic effect of the aforementioned technical features. By synchronously acquiring data from a 3D structured light sensor and a 2D camera, precise 3D morphology and texture information of the carpet is obtained, laying the data foundation for subsequent processing. The dynamic physical geometry model of the pile, constructed using a deep convolutional neural network, not only achieves intelligent region division but, more importantly, quantifies the direction and probability of pile collapse, enabling the system to understand the microscopic morphology of soft fabrics. The sub-pixel-level boundary contour lines extracted based on this model provide geometric constraints for high-precision printing. Furthermore, by combining the user-defined virtual light source with the actual pile height difference, a digital compensation pattern generated using an inverse optical rendering algorithm can actively counteract the light and shadow distortion caused by pile height differences and random collapse, creating a three-dimensional light and shadow effect that meets design expectations. Finally, online monitoring and automatic compensation mechanisms ensure the robustness of the printing results. These steps work together closely to fundamentally solve the problem of visual discontinuity and light and shadow distortion at the junction of high and low sections, which traditional technologies cannot handle, and realize a revolutionary leap in carpet printing from two-dimensional planar decoration to three-dimensional light and shadow shaping.

[0030] Example 2: In one embodiment, the method further refines the process of constructing the dynamic physical geometry model of the pile in step S2.

[0031] Step 1: Input the fused elevation map and texture image into a pre-trained deep convolutional neural network to automatically identify and label high-cutting areas, low-cutting areas, and transition areas. At the same time, construct a dynamic physical geometry model of the pile, including the direction of pile collapse, degree of collapse, and probability of collapse, for the pixels in the transition area. The dynamic physical geometry model of the pile is a probabilistic model, constructed by using a deep convolutional neural network to output the pile's falling direction, falling angle, and probability entropy value representing the randomness of falling for each pixel in the transition region. Specifically, this probabilistic model means that the shape of the pile is no longer considered a single deterministic geometric object, but rather described as a set with statistically distributed characteristics. The deep convolutional neural network serves as the model's generation source, and its output layer is configured to output a multi-dimensional feature vector in parallel for each pixel in the transition region. This vector includes at least three core components: the pile falling direction, the falling angle, and the probability entropy value. The pile falling direction characterizes the tilt of the pile in the horizontal plane, typically represented by an azimuth angle from 0 to 360 degrees. The pile falling angle characterizes the degree of tilt of the pile relative to the vertical normal, reflecting the severity of the falling. The probability entropy value is a quantitative indicator measuring the randomness or uncertainty of the falling state; its magnitude directly corresponds to the degree of disorder or prediction confidence of the pile arrangement at that pixel. For example, when a neural network processes pixels in a transition zone, if the output shows a 45-degree tilt direction, a 30-degree tilt angle, and a probability entropy value of 0.1, it indicates that the pile at that location is consistently tilted to the upper right, indicating a strong degree of certainty in its state. Conversely, if the probability entropy value at the same location is as high as 0.8, it indicates that the pile at that location may be subject to complex external forces, resulting in an irregular and chaotic tilting state, making single-direction prediction unreliable. Through this modeling method that incorporates probability entropy values, the system can accurately quantify the uncertainty of pile tilting in transition zones. This allows subsequent lighting and shadow compensation algorithms to adaptively adjust the edge blurring and transparency gradient of the compensation pattern based on the entropy value, thereby generating a soft and natural transition effect in areas with chaotic pile arrangement. This effectively avoids the visual harshness caused by forcibly fitting deterministic boundaries.

[0032] This application achieves a leap from deterministic geometric description to probabilistic statistical description by introducing probability entropy values ​​to construct a dynamic physical geometric model of the pile. The folding direction and folding angle output by the deep convolutional neural network provide the basic geometric information of the pile's micro-morphology, while the probability entropy values ​​endow the system with the ability to perceive uncertainty. The combined use of these three elements enables the inverse optical rendering algorithm to not only simulate the reflection of light on regularly tilted piles, but also to simulate the scattering and penumbra effect of light on randomly scattered piles based on the probability entropy values. This collaborative mechanism ensures that the generated digital compensation pattern has both accurate spatial positioning and a smooth gradation characteristic that conforms to real physical laws, fundamentally solving the problem of light and shadow distortion caused by random pile folding, and significantly improving the three-dimensionality and naturalness of the finished high-low tufted carpet print.

[0033] Example 3: In one embodiment, the method further optimizes the reverse optical rendering algorithm in step S4 to enhance the physical realism of the digitally compensated pattern.

[0034] Step 1: The reverse optical rendering algorithm integrates models for calculating light occlusion, multiple scattering, and penumbra among multiple light sources; The calculation model for light occlusion among multiple fibers simulates the physical process of higher fibers blocking light from being projected onto lower areas or the back of adjacent fibers. Its function is to accurately calculate the core region of the shadow (umbra) based on the height difference and falling direction in the dynamic physical geometry model of the fibers. The multiple scattering calculation model simulates the complex interactive behavior of light reflecting, refracting, and diffusely reflecting within a dense array of fibers. It is based on the two-way scattering distribution function (BSDF) of the fiber material and the spatial distribution density of the fibers, used to correct for overly dark or unnatural shadows caused by single ray tracing, ensuring that the shadow area retains appropriate ambient light brightness. The penumbra calculation model simulates the blurring of shadow boundaries caused by a light source with a certain area or irregular edges of the fibers. Its setting method dynamically determines the width of the penumbra based on the size parameters of the virtual light source and the probability entropy value of the fallen fibers.

[0035] For example, when the virtual light source is set to an upper-left 45-degree azimuth angle, for a tuft of fluff pointing downwards to the right, the occlusion model calculates that a dark umbra is formed on its right side; while the multiple scattering model calculates the supplementary lighting effect of light reflected from surrounding upright fluff on this umbra, avoiding stark blackness; simultaneously, if the probability entropy value of fluff falling over in this area is high (i.e., the falling state is chaotic), the penumbra model automatically expands the blur range of the shadow edge, simulating a soft transition zone. Through the coordinated operation of the three sub-models of occlusion, multiple scattering, and penumbra, the algorithm can reproduce the complex interaction between light and three-dimensional fluff structures in the real physical world, avoiding the defects of harsh shadows and lack of depth in traditional methods.

[0036] Step 2: Based on the probability distribution in the dynamic physical geometry model of the plush head, render and generate a digital compensation pattern with spatial gradient transparency and adaptive blurred edges; Spatial gradient transparency refers to the fact that the opacity (Alpha channel) of pixels in the compensation pattern is not a fixed value, but a property that continuously changes according to the light intensity attenuation curve calculated in the dynamic physical geometry model of the pile. Its function is to ensure that the printed projection pattern presents a smooth grayscale or color transition from the core shadow area to the non-shadow area, eliminating step artifacts. Adaptive blurring edge refers to the technical feature that the outline sharpness of the compensation pattern can be automatically adjusted according to the consistency of local pile fall. The determination method is: when the pile fall direction in a certain area is highly consistent (low probability entropy), the edge remains relatively sharp to simulate a sharp projection; when the pile fall direction in a certain area is disordered (high probability entropy), the edge is Gaussian blurred to simulate diffuse shadows.

[0037] For example, at the boundary between high-cut and low-cut pile, if the detected pile angles in the transition zone are mainly concentrated between 30 and 35 degrees with a concentrated probability distribution, the rendering engine will generate a compensation pattern with high edge sharpness and a steep transparency gradient to present a strong sense of three-dimensional cutting. Conversely, if the pile in this area is randomly distributed between 0 and 90 degrees due to external pressure, the generated compensation pattern will have a significant blurring effect at the edges and a gradual change in transparency to match this chaotic light effect. This probability-based adaptive rendering mechanism ensures that the final generated digital compensation pattern not only conforms to the illumination logic of the virtual light source on a macroscopic level, but also perfectly matches the actual physical uncertainty of the carpet pile surface on a microscopic level, thus achieving a realistic three-dimensional visual effect after printing.

[0038] This application achieves a leap from two-dimensional image synthesis to three-dimensional physical light field reconstruction through the synergistic effect of the aforementioned technical features. By deeply integrating light occlusion, multiple scattering, and penumbra calculation models into the inverse optical rendering algorithm, the system no longer relies on preset fixed filters but instead calculates the propagation path of light in a specific pile geometry in real time. Based on this, the probability distribution data provided by the pile dynamic physical geometry model drives the generation logic of spatial gradient transparency and adaptive blurred edges, enabling the digital compensation pattern to dynamically adapt to the randomness of pile fall. This synergy ensures that the printed light and shadow effects possess both the rigor of physical optics (such as correct shadow depth and reflection) and the natural texture of the fabric material (such as soft edge transitions), effectively solving the technical problems of false three-dimensional effects and obvious visual discontinuities caused by neglecting the complex interaction of light in existing technologies. This significantly enhances the artistic expression and added value of high-low tufted carpets.

[0039] Example 4: In an optional implementation, the method further includes introducing a material perception and differentiated ink volume control mechanism in step S2 to further enhance the visual hierarchy contrast of high and low cut areas.

[0040] Step 1: Use a 3D structured light contour sensor to acquire an elevation map of the carpet surface during movement, and use a 2D camera to simultaneously acquire texture images; The elevation map reflects the three-dimensional topography of the carpet surface, while the texture image records the color and micro-texture information of the carpet surface. These two types of data serve as the input foundation for subsequent deep convolutional neural networks, providing clues to the height distribution characteristics of the pile and the surface optical properties, respectively. Simultaneous acquisition ensured strict spatial correspondence, laying the data foundation for the subsequent accurate division of high-pile and low-pile areas and the analysis of their material properties.

[0041] Step 2: Input the fused elevation map and texture image into a pre-trained deep convolutional neural network to automatically identify and label high-cutting areas, low-cutting areas, and transition areas. At the same time, construct a dynamic physical geometry model of the pile, including the direction of pile collapse, degree of collapse, and probability of collapse, for the pixels in the transition area. In this process, the deep convolutional neural network performs the task of partition recognition in parallel with the task of classifying material properties. Specifically, based on the physical laws learned from the training data, the network determines that the high-cut pile area, due to its longer and relatively upright pile head, has a smoother and flatter top surface, tending to produce specular reflection, and therefore assigns it the material label of high reflectivity. Conversely, it determines that the low-cut pile area, due to its shorter and denser pile head, has a higher surface roughness, and light is scattered in multiple directions on its surface, thus assigning it the material label of diffuse reflectivity.

[0042] For example, when the network detects that the average texture height of a certain area is greater than 8mm and the surface normal vector variance is small (i.e., relatively flat), it marks it as a specular reflection area; while when it detects that the average texture height of a certain area is less than 4mm and the surface normal vector distribution is messy, it marks it as a diffuse reflection area. The generation of this material label is not a simple binary classification, but a continuous probability output that combines elevation gradient and texture frequency features.

[0043] Based on the aforementioned material labels, the system further generates differentiated ink volume output instructions. For high-cut velour areas marked as having high gloss reflectivity, the instructions tend to reduce the ink volume per unit area or use high-transparency ink to preserve the reflectivity of the velour itself, resulting in a bright visual effect under illumination. For low-cut velour areas marked as having diffuse reflectivity, the instructions tend to increase the ink volume or use high-saturation ink to enhance their light absorption capacity, resulting in a deep shadow effect. Through this ink volume control based on material properties, a significant contrast between light and dark areas is created between the high-cut and low-cut velour areas, thereby visually enhancing the three-dimensional layering of the high- and low-cut structure.

[0044] Step 3: Based on the dynamic physical geometry model of the pile head, extract the sub-pixel level closed-loop boundary contour line between the high-cut pile area and the low-cut pile area; This step utilizes the partitioning results containing material attribute information constructed above to help optimize the extraction accuracy of the boundary contour lines. The abrupt change points of the material tags often coincide with the boundary height of the pile height change. By jointly judging the geometric height jump and the material attribute jump, the true position of the high-low intersection can be more accurately located, generating a smooth and continuous sub-pixel level closed-loop boundary contour line, providing precise spatial constraints for the subsequent generation of compensation patterns.

[0045] Step 4: Receive the azimuth and elevation angles of the virtual light source set by the user. Based on the dynamic physical geometry model of the pile, the boundary outline, and the actual pile height difference on both sides of the outline, use the reverse optical rendering algorithm to dynamically generate a digital compensation pattern to offset the light and shadow distortion caused by the pile height difference and random falling, and to create a three-dimensional light and shadow effect under the virtual light source. Then, overlay it onto the transition area between the foreground and background patterns. In generating the 3D compensation pattern, the aforementioned differentiated ink volume output instructions are used as the base background layer in the compositing process. When calculating the grayscale and color of the compensation pattern, the reverse optical rendering algorithm comprehensively considers the specular reflection characteristics of the high-cut velvet area and the diffuse reflection characteristics of the low-cut velvet area. Specifically, the algorithm simulates the reflection behavior of light shining on velvet surfaces with different material labels, ensuring that the generated compensation pattern, after being superimposed with the base ink volume, presents a natural light and shadow transition in the human eye that conforms to the virtual light source settings. For example, in the specular reflection area, the compensation pattern reserves more specular channels to avoid over-coverage that causes reflections to disappear; in the diffuse reflection area, the compensation pattern deepens shadow details and utilizes the light absorption characteristics of the low-cut velvet to enhance the sense of depth.

[0046] Step 5: Drive the digital print head to position and print according to the partition pattern and the complete pattern after superimposed compensation; at the same time, the online monitoring camera simultaneously collects images for quality inspection and performs automatic compensation when defects are found.

[0047] The digital printhead strictly executes complete pattern data including differentiated ink volume instructions. In high-cut areas, the printhead ejects ink with a lower droplet density or smaller droplet volume to maintain the smoothness and reflectivity of the textured surface; in low-cut areas, the printhead fills with a higher droplet density to ensure color richness and light absorption. An online monitoring camera captures the printed image in real time to verify whether the visual contrast of the high- and low-cut areas meets expectations. If blurring due to ink volume control deviations is detected, the system will immediately adjust subsequent printing parameters or perform local reprinting to ensure that the final product presents a clear and vivid high- and low-cut stereoscopic visual effect.

[0048] This application utilizes a deep convolutional neural network to intelligently identify the material properties of high-cut and low-cut pile areas, and assigns differentiated ink volume output instructions to different material areas, achieving a leap from geometric shape compensation to material optical property optimization. The high-reflectivity label guidance system in the high-cut pile area reduces ink coverage, preserving the natural specular reflection ability of the pile, making it visually brighter and more prominent; the diffuse reflection label guidance system in the low-cut pile area increases ink deposition, enhancing light absorption and color saturation in this area, making it visually deeper and more restrained. This differentiated inkjet strategy based on physical material perception forms a deep synergy with the reverse optical rendering compensation pattern in step S4: the former constructs the basic light and shadow framework, while the latter refines the light and shadow details. The combined use of both not only eliminates the visual discontinuity at the high-low cut boundary but also actively amplifies the optical differences between the high and low pile surfaces themselves, significantly improving the three-dimensional layering, realism, and artistic expression of the finished carpet, effectively solving the problem of dull three-dimensional effects and unclear layers caused by neglecting material differences in traditional printing methods.

[0049] Example 5: In another optional embodiment, the method further includes introducing a closed-loop feedback mechanism during the printing process to dynamically correct internal model parameters.

[0050] Step 1: Receive the azimuth and elevation angles of the virtual light source set by the user. Based on the dynamic physical geometry model of the pile, the boundary outline, and the actual pile height difference on both sides of the outline, the reverse optical rendering algorithm is used to dynamically generate a digital compensation pattern to offset the light and shadow distortion caused by the pile height difference and random falling, and to create a three-dimensional light and shadow effect under the virtual light source. This pattern is then superimposed on the transition area between the foreground and background patterns. The inverse optical rendering algorithm is based on the dynamic physical geometry model of the textured surface constructed in the previous steps. This model includes the direction, degree, and probability distribution of the textured surface's fall over pixels within the transition zone, serving as the input for ray tracing calculations. By setting the azimuth and elevation angles of the virtual light source, the algorithm simulates the propagation path of light on textured surfaces with different height differences and fall over states, calculating a compensation pattern that can counteract cluttered shadows in the real environment and generate a stereoscopic projection effect. This compensation pattern features spatially gradient transparency and adaptive blurred edges, aiming to visually eliminate the sense of discontinuity at the junction of high and low points.

[0051] Step 2: Drive the digital print head to position and print according to the partition pattern and the complete pattern after superimposed compensation; while printing, the online monitoring camera simultaneously collects images for quality inspection and performs automatic compensation when defects are found; The online monitoring camera captures real-time images of the carpet surface during printing, not only to trigger immediate re-spraying actions from the printhead but also to collect data to support model iteration. Specifically, the system performs spatiotemporal correlation analysis on the collected defect information, filtering out repetitive defects that appear at the same physical coordinates in multiple carpet travel cycles. These repetitive defects with fixed locations are usually not accidental ink splattering but rather indicate a continuous change in the physical state of the local pile surface. For example, long-term ink accumulation may cause the pile to stick together, the direction of the pile to shift irreversibly, or there may be a slight deformation of the carpet substrate at that location.

[0052] For example, when an online monitoring camera repeatedly detects color deviations caused by abnormal pile collapse at the same coordinate point 1.5 meters from the edge along the width of the carpet, and these deviations cannot be completely eliminated by a single touch-up spray, the system determines that the actual probability distribution of pile collapse at that location has deviated from the initial model prediction. In this case, the coordinates and defect features of that location are marked as repetitive defect information with a fixed location.

[0053] Based on the above screening results, the system feeds back the information on repetitive defects with fixed locations to the dynamic physical geometry model construction stage of the pile in step S2. Specifically, these defect data are used as new training samples or constraints to dynamically update the collapse probability parameters of the corresponding local areas in the model. For example, if an abnormal shadow pointing in a specific direction frequently appears in a certain area, the model will adjust the collapse probability entropy value of the pixels in that area, increasing the weight of that specific collapse direction, thereby correcting the perception of the pile morphology in that area. Through this feedback mechanism, the dynamic physical geometry model of the pile is no longer static, but can continuously absorb measured data from the physical world as the production process progresses, achieving self-calibration and evolution. This ensures that the subsequently generated digital compensation patterns are always based on model data that is closest to the real physical state, effectively overcoming the model mismatch problem caused by equipment aging, process drift, or material batch differences, and significantly improving the robustness and printing accuracy of the system under long-term operation.

[0054] This application establishes a tight closed-loop collaboration between online monitoring in step S5 and model building in step S2. By capturing repetitive defects with fixed positions through online monitoring, the system can identify subtle changes occurring in the physical world and transform this information into a driving force for model updates. Furthermore, the updated dynamic physical geometry model of the pile guides the inverse optical rendering algorithm to generate more accurate compensation patterns, which are ultimately executed by the digital printhead for high-quality positioning printing. This closed-loop logic of perception-decision-execution-feedback-optimization gives the system a biological-like adaptive capability, enabling it to proactively adapt to dynamic disturbances in the production environment and ensure that the three-dimensional light and shadow effect of the high-low tufted carpet remains consistent and of high quality throughout the entire production cycle.

[0055] Example 6: In one embodiment, such as Figure 3 As shown, Figure 3 This application provides a schematic diagram of the intelligent digital positioning and printing system for high and low cut tufted carpets. The system, along the carpet's travel direction, includes a three-dimensional shape recognition module, a deep learning main control module, a digital printing module, and an online closed-loop quality inspection module. The deep learning main control module incorporates a pile physical model construction unit and a reverse optical rendering unit. The pile physical model construction unit is used to execute step S2 of the above embodiment to construct a dynamic physical geometric model of the pile. The reverse optical rendering unit is used to execute step S4 of the above embodiment to generate a digital compensation pattern. The deep learning main control module integrates and outputs a complete printing control signal containing the digital compensation pattern.

[0056] The 3D shape recognition module can refer to a sensing component located upstream of the carpet production line. Its function is to acquire the raw 3D data and 2D texture data of the carpet surface during its movement. This module connects to the subsequent deep learning main control module, transmitting the acquired elevation map and texture image to the main control module in real time, serving as the data foundation for the system's intelligent decision-making. In practical applications, the 3D shape recognition module may include a structured light sensor and an industrial camera. The structured light sensor projects a grating and receives reflected light to calculate the pile height information, while the industrial camera simultaneously captures the carpet surface color and texture information. The two are synchronized via timestamps to ensure spatial consistency.

[0057] The deep learning master control module refers to the core computing and control hub of the system, which integrates a high-performance processor and storage unit to run complex deep convolutional neural network algorithms and optical rendering algorithms. This module communicates with the 3D shape recognition module, digital printing module, and online closed-loop quality inspection module, acting as a crucial link between them. Specifically, the deep learning master control module receives perception data from upstream, processes it through internal algorithms, and then sends precise motion control and inkjet commands to the downstream printing module. This module can be an embedded system based on an industrial control computer architecture or an edge computing node connected to cloud computing power; this application does not impose any special limitations on this.

[0058] The velvet physics model building unit can be a software functional unit integrated within the deep learning main control module or a dedicated hardware acceleration module. Its name derives from the fact that this unit is specifically responsible for building a mathematical model describing the physical state of the velvet. This unit is connected to the data stream output by the 3D shape recognition module, receiving elevation maps and texture images, and processing them through a pre-trained deep convolutional neural network to output a dynamic physical geometry model of the velvet, including the direction, degree, and probability of velvet collapse. In the system's interconnected relationships, the velvet physics model building unit generates refined physical parameters for the transition zone. These parameters directly serve as input variables for the inverse optical rendering unit, determining the accuracy of subsequent lighting compensation calculations. This unit can be implemented as a neural network inference engine loaded with specific weight parameters, or as parallel computing logic based on an FPGA.

[0059] The reverse optical rendering unit, also integrated within the deep learning main control module, is a computational unit that reverse-engineers the required printing compensation pattern based on the virtual light source settings and the textured surface physical model. This unit works closely with the textured surface physical model construction unit, receiving the collapse probability distribution and boundary contour data output by the latter, as well as the user-defined azimuth and elevation angle parameters of the virtual light source. By running the reverse optical rendering algorithm, this unit calculates a digital compensation pattern that can offset real-world lighting distortion and create a virtual 3D effect. In the system workflow, the compensation pattern generated by the reverse optical rendering unit is sent to the compositing interface of the main control module, where it is superimposed with the foreground and background patterns to form the final printing control signal.

[0060] A digital printing module refers to the actuator that performs the final pattern transfer, typically including a printhead array, ink path system, and motion drive mechanism. This module operates based on the complete printing control signals output by the deep learning main control module, precisely jetting the fused pattern, including the compensation pattern, onto the carpet surface. The digital printing module and the main control module are connected via a high-speed bus to ensure real-time response to high-frequency inkjet commands. Depending on actual production needs, the digital printing module can be a single-channel scanning printer or a multi-channel single-pass industrial printer. Its inkjet resolution, droplet size, and supported ink types can be set according to the carpet material and production speed requirements.

[0061] An online closed-loop quality inspection module refers to a detection and feedback component located downstream of the carpet's travel direction. Its function is to immediately acquire images of the finished product and perform quality assessment after printing. This module has a bidirectional data channel with the deep learning main control module. On one hand, it transmits the acquired image data to the main control module for defect identification; on the other hand, when a defect is detected, it triggers the main control module to execute automatic compensation commands or adjust subsequent printing parameters. The online closed-loop quality inspection module typically includes a high-resolution line scan camera and an illumination source, and its installation position must ensure effective capture of surface features both before and after ink drying.

[0062] Specifically, the working process of this application system is as follows: The carpet travels along the guide belt and first passes through the 3D shape recognition module. This module simultaneously acquires the elevation data and texture image of the carpet surface and sends them to the deep learning main control module. The pile physical model construction unit in the main control module immediately processes the data, identifies the high-cut pile area, low-cut pile area, and transition area, and constructs a dynamic physical geometric model describing the pile collapse state in the transition area. Subsequently, the reverse optical rendering unit calls this model and, combined with the user-preset virtual light source parameters, generates a digital compensation pattern for eliminating visual tomography and simulating stereoscopic projection through reverse calculation. The deep learning main control module merges this compensation pattern with the designed base pattern to generate a complete printing control signal and sends it to the digital printing module. The digital printing module drives the printhead to complete the positioning and printing according to the signal. Finally, the online closed-loop quality inspection module scans the printed carpet. If defects are found, a feedback signal is sent to the main control module to trigger the compensation mechanism, thus forming a complete perception-cognition-execution-feedback closed loop.

[0063] As a preferred embodiment, the solution of this application is implemented as follows: In the carpet production line, the three-dimensional shape recognition module collects carpet surface point cloud data at a frequency of 10kHz. The deep learning main control module uses the built-in GPU acceleration card to run a deep convolutional neural network and completes the probabilistic modeling of the pile lying direction in the transition zone within milliseconds. Based on this probabilistic model, the reverse optical rendering unit simulates the occlusion and scattering effect of light between tilted piles to generate compensation stripes with gradient transparency. The digital printing module accurately prints compensation color blocks at the junction of high and low cuts according to the synthesized pattern signal. At the same time, the online closed-loop quality inspection module monitors the printing effect in real time. Once a local color difference caused by ink splatter is detected, the coordinate information is immediately sent back to the main control module. The main control module dynamically adjusts the inkjet volume of the next cycle or notifies the rework organization for processing.

[0064] Through the above technical solution, this application realizes the solidification of the intelligent digital positioning printing method for high and low cut tufted carpets into a specific engineering system. Because the system sequentially arranges sensing, calculation, execution, and detection modules along the direction of travel, and integrates pile physical model construction and reverse optical rendering functions in the core main control module, the system can understand the three-dimensional physical form of the carpet in real time and dynamically generate light and shadow compensation strategies. This solves the problem of light and shadow distortion caused by random pile collapse, which traditional equipment cannot handle. It achieves the transformation from theoretical algorithms to stable industrial production, ensuring the consistency and artistic expression of high-end carpet products.

[0065] Example 7: In one possible implementation, the method further includes a reverse optical rendering unit, which is also used to execute the rendering algorithm of the above embodiments that integrates ray occlusion, multiple scattering and penumbra calculation models.

[0066] The inverse optical rendering unit refers to a functional component deployed within the deep learning main control module. Its core responsibility is to perform high-fidelity inverse lighting and shadow calculations based on the dynamic physical geometry model of the carpet pile. When performing rendering tasks, this unit does not employ simplified planar mapping logic but instead embeds a complex ray transmission calculation mechanism. Specifically, this unit integrates a ray occlusion calculation model between multiple piles to simulate the shadow occlusion effect produced by high-cut piles on low-cut piles or adjacent piles when illuminated by a virtual light source. Simultaneously, this unit also integrates a multiple scattering calculation model to quantify the repeated reflection and diffusion behavior of light on the surface and gaps between dense pile fibers, thus restoring the unique soft luster of the carpet material. Furthermore, this unit integrates a penumbra calculation model to handle shadow transition areas caused by the non-point light source characteristics and the blurring of pile edges, thereby calculating shadow boundaries with a natural gradient effect.

[0067] There is a close data linkage between the reverse optical rendering unit and the pile physics model building unit in the system. The probability distribution data output by the pile physics model building unit, which includes the direction, degree, and probability of pile collapse, is transmitted to the reverse optical rendering unit in real time as input parameters. Based on these probability distributions and combined with the azimuth and elevation angles of the virtual light source set by the user, the reverse optical rendering unit calculates the ink compensation value required for each pixel through the aforementioned occlusion, multiple scattering, and penumbra calculation models. This collaboration ensures that the generated digital compensation pattern not only contains color information but also light and shadow structure information with spatial gradient transparency and adaptive blurred edges. This ensures that the final pattern printed on the carpet surface can accurately offset the light and shadow distortion caused by the actual pile height difference and random collapse, and create a stereoscopic visual effect that conforms to the virtual light source settings.

[0068] Specifically, the working process of this application is as follows: After the system receives the elevation map and texture image of the carpet to be printed, the pile physical model construction unit first generates a dynamic physical geometric model of the pile in the transition area; then, the reverse optical rendering unit calls the model and starts the integrated light occlusion, multiple scattering and penumbra calculation engine; the engine simulates the propagation path of light in a pile group with a specific collapse probability distribution according to the position of the virtual light source, and calculates the umbra area formed by the mutual occlusion of the pile, the reflective area formed by multiple scattering and the penumbra area formed by edge blurring; finally, based on the simulation results, the compensation pattern that needs to be pre-printed is derived in reverse. This pattern presents a continuous change in transparency in space and an adaptive blurring effect at the edges to match the optical response characteristics of the real pile surface.

[0069] As a preferred embodiment, the solution of this application is implemented as follows: When printing on a tufted carpet with a high-low cut structure, the reverse optical rendering unit reads data on the probability of pile collapse in a certain local area of ​​the transition zone. The data shows that the pile in this area has a 70% probability of tilting 30 degrees to the left. At this time, if the virtual light source is set to the upper right, the rendering unit uses a light occlusion model to calculate that the high-cut pile on the right will cast a shadow in the low-cut area on the left. It uses a multiple scattering model to calculate the weak brightness inside the shadow due to fiber reflection, and uses a penumbra model to determine the blurred width of the shadow edge. Based on this, the rendering unit generates a compensation pattern with a gradient from completely opaque to completely transparent. This pattern is superimposed on the main pattern, causing the printhead to reduce ink volume or print a darker compensation color in the preset shadow area during printing, while printing normally in the non-shadow area, thereby creating a realistic three-dimensional projection on the physical pile surface.

[0070] Through the above technical solution, this application achieves the embedding of advanced optical rendering algorithms at the system hardware level, ensuring that the reverse optical rendering unit can execute a complete rendering process including ray occlusion, multiple scattering, and penumbra calculation. Because the system fully considers the real physical behavior of light in complex textured structures when generating compensation patterns, the generated digital compensation patterns have extremely high physical realism, effectively eliminating visual breaks at the junctions of high and low points, and presenting a natural and soft three-dimensional light and shadow effect, avoiding the harsh boundaries and false shadows caused by traditional simplified algorithms.

[0071] Example 8: In another optional embodiment, the method further includes a deep learning master control module, which is also used to perform the operations described in the above embodiment of assigning different material labels to different velvet areas and generating differentiated ink volume instructions.

[0072] Assigning different material labels to different pile areas refers to the process where, during the deep learning main control module's partitioning and recognition of the carpet surface, it not only divides the carpet into high-cut pile areas, low-cut pile areas, and transition areas based on elevation maps and texture images, but also further infers the optical reflection characteristics of each area based on its surface micro-geometric features and assigns corresponding virtual material labels. These material labels are not physical labels attached to the carpet surface, but rather logical attribute markers stored in the main control module's internal data stream, used to characterize the reflection behavior pattern of the pile in that area when exposed to light. For example, high-cut pile areas with higher elevation and relatively flat surfaces can be assigned material labels characterizing high-gloss reflection or specular reflection characteristics; while low-cut pile areas with lower elevation and rougher surfaces due to their tufted structure can be assigned material labels characterizing diffuse reflection or Lambertian reflection characteristics. This process of assigning material labels is essentially the system mapping the physical pile morphology to optical parameters in the digital world. The specific classification thresholds or judgment logic can be set according to the actual type and density of carpet fibers and the requirements of the lighting environment; this embodiment does not impose any special limitations on this.

[0073] The material tags are closely linked to the currently defined deep learning master control module and other preceding technical features. Specifically, while executing the physical model construction and inverse optical rendering of the carpet pile in the above embodiments, the deep learning master control module uses the aforementioned material tags as additional input variables to participate in the subsequent generation of printing control signals. After receiving elevation and texture data from the 3D shape recognition module, the master control module outputs the partitioning results and corresponding material properties in parallel through the built-in deep convolutional neural network. Subsequently, when generating the digital compensation pattern and the final printing instructions, the module dynamically adjusts the ink droplet ejection strategy according to the material tags of different areas. For example, in areas marked as high-reflectivity, the system may tend to reduce the coverage of dark ink or adjust the ink droplet size to preserve the carpet pile gloss, while in areas marked as diffuse reflection, it may increase the ink penetration or use a specific color mixing ratio to enhance the light absorption effect. This cooperative relationship enables the master control module to achieve multi-dimensional collaborative control of the visual texture of the carpet surface, ensuring that the output printing control signal not only includes position and color information but also embeds fine-tuning instructions for material characteristics.

[0074] Generating differentiated ink volume instructions can refer to the deep learning main control module calculating a sequence of inkjet control parameters that are different for different textured areas based on the aforementioned material labels. Specifically, this instruction manifests as a combination of pulse width, jet frequency, droplet volume, or ink path opening time that drives the digital printhead. In practice, for high-cut textured areas labeled with high-reflectivity material, the generated ink volume instruction can control the printhead to jet less ink or use high-gloss spot color ink to avoid overwhelming the textured surface and maintain its reflective properties. Conversely, for low-cut textured areas labeled with diffuse reflection material, the generated ink volume instruction can control the printhead to jet more ink or use highly saturated color combinations to utilize the light-absorbing properties of the low-cut texture to create a deeper visual depth. The specific numerical range, ink color ratio curve, and jetting sequence of such differentiated ink volume instructions can be flexibly set according to the ink absorption performance of the carpet substrate, the physicochemical properties of the ink used, and the actual needs of the target visual effect. For example, it can be a linear correspondence or a non-linear lookup table mapping relationship. This application embodiment does not impose any special limitations on this.

[0075] Specifically, the working process of this application is as follows: After acquiring real-time 3D topographic data and 2D texture data of the carpet during its movement, the deep learning main control module first completes region segmentation and material attribute inference simultaneously through a neural network model, attaching a material tag representing its optical characteristics to each pixel or processing unit; then, the module introduces these material tags as key weight factors into the printing path planning and ink volume calculation algorithms, calculating differentiated ink volume distribution maps adapted to the reflective characteristics of high-cut pile areas and low-cut pile areas respectively; finally, the module sends a complete control signal containing position coordinates, color information, and differentiated ink volume parameters to the digital printing module, driving the printhead to perform precise zone printing operations. In this process, the material tags act as a bridge, directly converting the perceived physical topographic features into process parameters at the execution end, realizing closed-loop control from structure recognition to texture expression.

[0076] As a preferred embodiment, the solution of this application is implemented as follows: On a carpet production line, when a carpet with a high-low cut structure passes through a 3D shape recognition module, the system detects that the pile height in a certain area is significantly higher than the surrounding area and the surface point cloud distribution is uniform. The deep learning main control module then determines that this area is a high-cut pile area and assigns it a high-gloss material label; at the same time, the adjacent low-lying area is determined to be a low-cut pile area and assigned a diffuse reflection material label. Based on this, when generating printing instructions, the main control module issues a 15% reduction ink volume control instruction to the nozzles corresponding to the high-cut pile area and prioritizes the use of yellow and cyan ink channels to simulate the high-gloss color under sunlight; while for the low-cut pile area, it issues a 10% increase instruction and deepens the proportion of magenta and black ink channels to enhance the shadow depth. Finally, the pattern printed by the digital printing module according to this differentiated instruction presents a strong contrast in visual appearance, with the high-cut area being bright and transparent and the low-cut area being deep and heavy, effectively enhancing the three-dimensional layering of the carpet.

[0077] Through the above technical solution, this application achieves the ability of the deep learning main control module to assign material labels to different pile areas and generate differentiated ink volume instructions, enabling the system to adaptively adjust the inkjet strategy according to the physical reflection characteristics of the pile. This solves the technical problem that single ink volume control cannot take into account the visual performance differences between high and low cut areas. As a result, without changing the physical structure of the carpet, the technical effect of significantly enhancing the visual layer contrast of high and low cuts and improving the artistic expression of the finished product is achieved through the fine control of ink volume and color.

[0078] Example 9: In another optional embodiment, the method further includes: a data feedback channel is provided between the online closed-loop quality inspection module and the deep learning main control module to transmit back the detected repetitive defect information for the pile head physical model construction unit to dynamically update the model parameters.

[0079] The data feedback channel can refer to a bidirectional or unidirectional high-speed data transmission link established between the online closed-loop quality inspection module and the deep learning main control module. Its physical form can be a wired connection (such as industrial Ethernet or fiber optic) or a wireless connection (such as industrial Wi-Fi or 5G private network), and this application embodiment does not impose any special limitations on it. The functional positioning of this data feedback channel in the overall technical solution is to construct the reverse information flow path in the perception-decision-execution-learning closed loop. Its role is to map the real physical state captured by the back-end quality inspection link back to the front-end cognitive model. Through this channel, the image data and defect analysis results collected by the online closed-loop quality inspection module can be transmitted to the deep learning main control module in real time or near real time, enabling the carpet physical model construction unit to obtain a second layer of verification data in addition to the initial sensor input, thereby forming a quantitative basis for the deviation between the actual printing effect on the carpet surface and the expected model.

[0080] Repetitive defect information refers to a set of printing defects that repeatedly occur within the same spatial location or area with the same texture characteristics during a continuous production cycle, such as ink splatter, broken ink streaks, color shifts, or pattern misalignments. Unlike random, sporadic defects, repetitive defect information often indicates a systematic deviation between the local physical state of the carpet pile (such as the direction of collapse and density distribution) and the predicted values ​​of the current dynamic physical geometry model of the pile. In this scheme, repetitive defect information serves as the core data source driving model evolution. Its content includes not only the type and coordinates of the defects but also statistical values ​​of their morphological characteristics. The interaction between this information and the pile physical model building unit is as follows: after receiving this information, the model building unit uses it as part of the loss function or as a constraint to reverse-correct the collapse probability distribution parameters or material label weights of specific pixels in the transition zone, thereby making the model's description of the local pile state closer to physical reality.

[0081] Specifically, the working process and principle of this application are as follows: During the carpet printing production process, the online closed-loop quality inspection module first performs a full-width scan of the printed carpet surface to identify and mark various defects. The system's internal processing logic performs time-series correlation analysis on these defects. If it is determined that a certain type of defect appears multiple times in the same relative position, it is classified as repetitive defect information. Subsequently, this information is sent to the deep learning main control module via the data feedback channel. After receiving this information, the pile physical model construction unit initiates a parameter fine-tuning mechanism. It no longer relies solely on the initial elevation map and texture image obtained in step S1, but combines the actual printing results from the quality inspection feedback to recalculate the pile collapse probability entropy value of the affected area or adjust the local optical rendering parameters. This mechanism enables the system to have self-correction capabilities, eliminating model drift caused by mechanical wear, environmental changes, or batch differences in raw materials due to long-term operation, ensuring that the subsequently generated digital compensation patterns are always based on the latest and most accurate physical state of the pile surface.

[0082] As a preferred embodiment, the solution of this application is implemented as follows: Assuming that during the continuous printing of a batch of high-low cut tufted carpets, the online closed-loop quality inspection module discovers that a fine ink diffusion band always appears at the edge of the low-cut pile area, and this phenomenon repeats at a fixed position per meter of carpet. The system determines this to be a repetitive defect and sends the image features and coordinates of this area back to the deep learning main control module through the data feedback channel. The pile physical model construction unit analyzes and concludes that the original model underestimates the degree of pile collapse in this area, resulting in insufficient blurred edges in the compensation pattern generated by the reverse optical rendering algorithm. Therefore, the unit automatically increases the average collapse angle of the pixels in this local area and increases the collapse probability entropy value to simulate a more severe pile disorder state. The updated model parameters are immediately applied to the generation of the compensation pattern in the next cycle, making the edges of the printed compensation pattern softer and the coverage area larger, thereby physically offsetting the ink diffusion and eliminating the repetitive defect.

[0083] Through the above technical solution, this application realizes the establishment of a reverse information flow from quality inspection to modeling, and truly realizes a complete AI closed loop of perception-decision-execution-learning. Since the repetitive defect information is introduced as the basis for model updates, the system has the ability to continuously optimize and can continuously correct its own cognitive biases during operation. This solves the problem of the system losing its adaptive ability due to the lack of an effective data feedback mechanism, thereby improving the stability of long-term operation and the consistency of product quality.

[0084] Example 10: In one possible implementation, the method further includes: the deep convolutional neural network has been pre-trained using a dataset of fused 3D and 2D carpet images with light and shadow labels generated by physically based simulation rendering before application.

[0085] The dataset of fused 3D and 2D carpet images with light and shadow labels generated based on physical simulation rendering can refer to a synthetic dataset containing a large number of virtual carpet samples constructed using computer graphics technology. This dataset is not derived directly from simple photographs of real carpets, but is generated based on physical laws. Its construction process can be as follows: First, a 3D geometric model of a carpet with different pile height distributions, pile densities, and different collapse states is established in a virtual environment; then, various virtual light source conditions (including different azimuth angles, elevation angles, and light intensities) are set, and ray tracing or radiosity algorithms are used to calculate the propagation paths of light on the surfaces of these 3D models, simulating light occlusion, multiple scattering, and penumbra effects between piles; finally, corresponding 2D texture images and label images that accurately reflect the light and shadow distribution are rendered. These label images serve as ground truth values ​​to supervise the training process of a deep convolutional neural network, enabling it to learn the mapping relationship between 2D image features and the physical morphology of the 3D pile and the laws of light and shadow. This application does not impose special limitations on this; the specific size of the dataset, the complexity of the virtual scene, and the type of physical simulation engine can be set according to actual training needs.

[0086] A deep convolutional neural network (CNN) can refer to a deep learning model with multiple convolutional structures, pooling layers, and fully connected layers. Its architecture can be customized according to specific needs, such as the ResNet series, U-Net series, or a specially designed encoder-decoder structure. In this technical solution, the network functions as the core parsing unit of the system, responsible for receiving input data that integrates elevation and texture information, and outputting predictions about pile partitioning, pile flattening parameters, and material properties. The network works closely with the aforementioned dataset: the dataset provides physically realistic input-output sample pairs, and the network iteratively optimizes its internal weight parameters to minimize the error between the prediction results and the physical simulation labels, thereby achieving a generalized understanding of the complex light and shadow features of high-low tufted carpets. Through this collaboration, the network can maintain stable recognition accuracy even when facing the ever-changing pile flattening and lighting conditions in real production environments, avoiding model failure or misjudgment due to a lack of physical consistency in the training data.

[0087] Specifically, the working process and principle of this application are as follows: Before the system is officially put into operation, the aforementioned training dataset based on physical simulation is first constructed. This dataset is used to pre-train a deep convolutional neural network offline, establishing a deep correlation between the two-dimensional visual representation and the three-dimensional physical structure and the causes of light and shadow within the network. When the network is deployed in the high-low cut tufted carpet intelligent digital positioning and printing system, upon receiving real-time carpet image data collected by sensors, it can invoke the physical prior knowledge obtained through pre-training to accurately analyze the boundaries of the high-cut, low-cut, and transition zones of the current pile surface, and quantify the direction and probability distribution of the pile's collapse. Because the training labels themselves contain precise physical laws of light and shadow, the model parameters output by the network naturally possess optical consistency, thus providing a reliable data foundation for the subsequent inverse optical rendering unit to generate compensation patterns that eliminate visual distortions and conform to the virtual light source settings.

[0088] As a preferred embodiment, the solution of this application is implemented as follows: During the model training phase, technicians use a physics engine to generate 10,000 sets of virtual carpet samples. Each set of samples contains randomly varying pile height differences (0.5mm to 3mm), random lodging angles (0° to 90°), and random lodging directions. For each set of samples, four typical lighting environments—morning, noon, evening, and indoor artificial lighting—are simulated to generate corresponding two-dimensional images and light and shadow label maps containing precise shadow distributions. These 36,000 fused images (10,000 × 4 lighting conditions) are input into the deep convolutional neural network to be trained, and backpropagation training is performed using the light and shadow label maps as supervision signals until the network loss function converges. The trained network is loaded into the deep learning main control module of the system. On the actual printing production line, when the moving carpet passes through the three-dimensional shape recognition module, the real-time acquired elevation map and texture image are sent to the pre-trained network. The network immediately outputs the dynamic physical geometry model of the pile on the current carpet surface, guiding the subsequent digital printing module to perform high-precision stereoscopic projection compensation printing.

[0089] Through the above technical solution, this application achieves the following: by using light and shadow labels generated based on physical simulation rendering to pre-train the deep convolutional neural network, the model not only masters the surface features of the image during the learning process, but also deeply understands the physical mechanism of the interaction between light and the pile. Therefore, it can effectively overcome the subjective error and physical inconsistency problems of traditional manual annotation data, significantly improve the generalization ability and robustness of the neural network when facing complex and ever-changing real pile surface conditions, and ensure that the entire printing system can continuously and stably generate high-quality stereoscopic visual effects during long-term operation.

[0090] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for intelligent digital positioning printing of high and low cut tufted carpets, characterized in that, Includes the following steps: S1. Use a 3D structured light contour sensor to obtain an elevation map of the carpet surface during movement, and use a 2D camera to simultaneously acquire texture images. S2. Input the fused elevation map and texture image into a pre-trained deep convolutional neural network to automatically identify and label the high-cutting area, low-cutting area and transition area. At the same time, construct a dynamic physical geometry model of the pile, including the direction of pile collapse, the degree of collapse and the probability of collapse, for the pixels in the transition area. S3. Based on the dynamic physical geometry model of the pile head, extract the sub-pixel level closed-loop boundary contour line between the high-cut pile area and the low-cut pile area; S4. Receive the azimuth and elevation angles of the virtual light source set by the user. Based on the dynamic physical geometry model of the pile, the boundary contour line and the actual pile height difference on both sides of the contour, use the reverse optical rendering algorithm to dynamically generate a digital compensation pattern to offset the light and shadow distortion caused by the pile height difference and random falling, and to create a three-dimensional light and shadow effect under the virtual light source. Then, superimpose it onto the boundary transition area between the foreground and background patterns. S5 drives the digital printhead to perform positioning and printing based on the partitioned pattern and the complete pattern after superimposed compensation; while printing, the online monitoring camera simultaneously acquires images for quality inspection and performs automatic compensation when defects are detected.

2. The method according to claim 1, characterized in that, In step S2, the dynamic physical geometry model of the pile is a probabilistic model, which is constructed by outputting the pile's falling direction, falling angle, and probability entropy value representing the randomness of falling for each pixel in the transition area through a deep convolutional neural network.

3. The method according to claim 1, characterized in that, In step S4, the reverse optical rendering algorithm integrates a calculation model of light occlusion, multiple scattering, and penumbra between the multi-pile heads, and renders a digital compensation pattern with spatial gradient transparency and adaptive blurred edges based on the probability distribution in the dynamic physical geometry model of the multi-pile heads.

4. The method according to claim 1, characterized in that, In step S2, while identifying the partitions, the deep convolutional neural network assigns a material label with high reflectivity to the high-cut velvet area and a material label with diffuse reflectivity to the low-cut velvet area, and generates differentiated ink volume output instructions to enhance the visual contrast between high and low cuts.

5. The method according to claim 1, characterized in that, In step S5, the repetitive defect information with fixed location detected by the online monitoring camera will be fed back to step S2 to dynamically update the local collapse probability parameter in the dynamic physical geometry model of the pile.

6. A method for intelligent digital positioning printing of high and low cut tufted carpets, characterized in that: The system, arranged sequentially along the carpet travel direction, includes a 3D shape recognition module, a deep learning main control module, a digital printing module, and an online closed-loop quality inspection module. The deep learning main control module has a built-in physical model construction unit for the fuzzy head and a reverse optical rendering unit; The pile physical model construction unit is used to perform the pile dynamic physical geometry model construction described in step S2 of claim 1; The reverse optical rendering unit is used to perform the digital compensation pattern generation described in step S4 of claim 1; The deep learning master control module integrates and outputs a complete printing control signal containing the digital compensation pattern.

7. The system according to claim 6, characterized in that, The reverse optical rendering unit is also used to execute the rendering algorithm described in claim 3, which integrates the light occlusion, multiple scattering and penumbra calculation models.

8. The system according to claim 6, characterized in that, The deep learning main control module is also used to perform the operation described in claim 4, which assigns different material labels to different velvet areas and generates differentiated ink volume instructions.

9. The system according to claim 6, characterized in that, A data feedback channel is provided between the online closed-loop quality inspection module and the deep learning main control module to transmit the detected repetitive defect information back, so that the pile physical model construction unit can dynamically update the model parameters.

10. The system according to any one of claims 6 to 9, characterized in that, Before application, the deep convolutional neural network was pre-trained using a dataset of fused 3D and 2D carpet images with light and shadow labels generated based on physical simulation rendering.

Images