A machine vision-based method and system for intelligent overprinting and positioning of textile packaging bags, including deviation detection and correction control.

By eliminating texture interference through dual-layer adaptive filtering and edge enhancement algorithms, a three-dimensional digital standard template is constructed. Combined with texture recognition and color feature analysis, the accuracy problem of overprinting deviation detection for textile packaging bags is solved, achieving sub-pixel-level detection and closed-loop correction control, thereby improving production stability and yield.

CN122453717BActive Publication Date: 2026-11-10HANGZHOU SONGYANG PRINTING IND CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202610442745.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-04-07
Publication Date
2026-11-10
Estimated Expiration
2046-04-07

AI Technical Summary

Technical Problem

Existing technologies cannot effectively distinguish between the base texture of textile packaging bags and the overprinted pattern, resulting in decreased matching accuracy. Deformation of the textile substrate during high-speed printing causes distortion and pseudo-deviation of the Mark points, making it impossible to achieve sub-pixel level detection. Furthermore, the lagging correction control cannot solve the problem of cumulative deviation in multi-color overprinting.

Method used

A dual-layer adaptive filtering and edge enhancement algorithm is used to eliminate texture interference, a three-dimensional digital standard template is constructed, and texture recognition and color feature analysis are combined to perform accurate deviation detection. Deviation is decoupled through the elastic deformation model of textile substrate, a closed-loop correction control system is constructed, and the response delay of the mechanism is predicted and compensated by the time-series prediction model.

Benefits of technology

It achieves sub-pixel-level precise detection of printing deviations in textile packaging bags, eliminates false deviations, improves the production stability and yield of textile packaging bags, and avoids the generation of batches of defective products.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122453717B_ABST
    Figure CN122453717B_ABST
Patent Text Reader

Abstract

This invention belongs to the field of machine vision technology and discloses a machine vision deviation detection and correction control method and system for intelligent overprinting positioning of textile packaging bags. By designing a dual-layer adaptive filtering and edge enhancement algorithm, the method first uses frequency domain Fourier transform combined with adaptive notch filtering to suppress the frequency domain components of the substrate texture, then uses Gaussian difference filtering to remove high-frequency noise from the fuzz. It also optimizes the adaptive threshold Canny operator to achieve edge locking of the overprinting features, and combines image segmentation technology to isolate interference between the substrate and the overprinting area. Simultaneously, it constructs a three-dimensional digital standard template that integrates overprinting features, texture features, and position parameters to prevent the substrate texture from obscuring the overprinting features or misjudging them as noise. By constructing a triple-constraint judgment mechanism of texture recognition, pattern determination, and color feature analysis, it can accurately identify and eliminate false deviations caused by substrate stretching, wrinkles, uneven warp and weft density, and color halo, preventing the correction mechanism from malfunctioning due to misjudgment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of machine vision technology, specifically a machine vision deviation detection and correction control method and system for intelligent overprinting and positioning of textile packaging bags. Background Technology

[0002] With the environmental upgrade of the flexible packaging industry, textile packaging bags made of materials such as non-woven fabrics, cotton and linen, and Oxford cloth are widely used in various fields due to their advantages such as reusability, tensile strength, and biodegradability. Multi-color overprinting is the core production process. Currently, flexible packaging overprinting deviation detection and correction technologies are mostly applicable to rigid / semi-rigid substrates such as plastic films and paper packaging. Related solutions often employ a fixed printing mark point template matching method, using machine vision to identify mark point position deviations and drive correction. The core image recognition process relies on feature extraction and template matching technologies. However, for textile packaging bag scenarios, existing technologies have the following technical problems: Textile substrates have periodic warp and weft textures, surface fuzz, and strong elastic deformation characteristics. Existing preprocessing and feature extraction methods cannot effectively distinguish between substrate textures and overprinted patterns. They are prone to misjudging textures as noise or obscuring overprinted features, resulting in decreased matching accuracy and the inability to achieve sub-pixel level detection. Existing technologies rely heavily on registration with a single fixed Mark point. However, textile substrates are prone to stretching, wrinkling, and local deformation during high-speed traction and printing processes, which leads to Mark point distortion. Conventional Harris corner detection and SIFT / ORB template matching methods have a significantly reduced recognition rate for deformed Mark points, which can easily result in mismatches and missed detections, directly causing distortion in deviation detection results. Local stretching, wrinkling, and uneven warp and weft density of textile substrates can cause visual displacement of the substrate itself. Existing technology cannot distinguish between visual errors caused by the deformation of the substrate itself and actual overprinting deviations. A large number of false deviations trigger the correction mechanism to malfunction, which in turn aggravates overprinting errors and severely limits the improvement of yield. Existing solutions are mostly open-loop control with lag in deviation detection and correction execution. They do not take into account the elastic deformation characteristics of textile substrates and the mechanical response delay of high-speed production lines. The correction action always lags behind the occurrence of deviation, which cannot solve the problem of cumulative deviation in multi-color printing. In high-speed scenarios, batch defects are very likely to occur. Summary of the Invention

[0003] The purpose of this invention is to provide a machine vision deviation detection and correction control method and system for intelligent overprinting and positioning of textile packaging bags, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a machine vision deviation detection and correction control method for intelligent overprinting positioning of textile packaging bags, comprising the following specific steps: Preferably, in the acquisition and preprocessing stage, a linear array camera and a diffuse ring light source synchronous acquisition system are built in conjunction with the encoder of the traction roller of the printing production line to achieve equidistant trigger acquisition, obtain continuous printing images of textile packaging bags after multi-color overprinting, and eliminate image ghosting and distortion caused by high-speed movement. To address the texture interference characteristics of textile substrates, a dual-layer adaptive filtering and edge enhancement algorithm is designed. First, the characteristic frequency bands of the periodic warp and weft texture of the textile substrate are decomposed by frequency domain Fourier transform. Adaptive notch filtering is used to suppress the frequency domain components of the substrate texture. Combined with texture analysis methods, the periodic feature parameters of the texture are extracted to achieve the localization and suppression of texture frequency bands. Then, Gaussian difference filtering is used to remove high-frequency noise from surface fuzz. An edge detection optimization algorithm is used to optimize the adaptive threshold Canny operator. The threshold parameters are dynamically adjusted by combining the grayscale features of the overprinted pattern and the edge gradient characteristics to achieve edge feature locking of the overprinted mark point and the pattern, while preserving edge detail information. Image segmentation technology is used to segment the overprinted area and the substrate area of ​​the preprocessed image, and finally outputs a preprocessed image with texture suppression, noise removal, clear edges and complete region segmentation.

[0005] Preferably, the template construction stage is based on the preprocessed image output from the acquisition and preprocessing stage, selects standard unbiased textile packaging bag overprint samples, optimizes the Harris corner detection operator using a corner detection algorithm and sets a deformation adaptive weight factor, extracts sub-pixel level corner features of the reference color group Mark points, and performs feature encoding on the edge curvature and grayscale distribution of the Mark points to generate Mark point-specific feature vectors. Based on contour extraction technology, the closed contour feature points of the overprinted pattern are extracted. The contour moment algorithm is used to calculate the geometric feature parameters of the contour, including area, perimeter, and eccentricity. The contour feature descriptor is generated by combining the texture detail features of the pattern to realize the unique identification of the overprinted pattern. The Mark point feature vector and the contour feature descriptor are fused and the deformation invariant encoding is performed by the pattern recognition algorithm to generate a composite feature vector with robustness to rotation, scaling, and local deformation. Textile substrate texture features of each region in the standard sample are extracted and texture feature library is constructed using texture classification method. This library is then integrated with composite overprint features and standard overprint position parameters to form a three-dimensional digital standard template containing overprint features, texture features and position parameters, which is then stored in the template library.

[0006] Preferably, the deviation detection stage is based on a three-dimensional digital standard template containing overprint features, texture features and position parameters generated in the template construction stage. The composite features of the reference color group are used as the registration reference. The cosine similarity of the composite feature vector is calculated by the ORB feature matching algorithm, and feature point pairs with a matching degree higher than a preset threshold are selected. For the remaining color sets, local non-rigid registration is performed in sequence. Taking into account the elastic deformation characteristics of the textile substrate, the feature point matching weight is dynamically adjusted using an elastic registration algorithm. The lateral deviation, longitudinal deviation, rotation angle deviation, and local flexible deformation deviation data of each color set relative to the standard template are initially obtained to complete the initial inspection of the registration deviation. At the same time, the feature matching data and deviation distribution during the initial inspection process are recorded.

[0007] Preferably, the precise detection stage is based on the initial detection data of overprinting deviation obtained in the initial deviation detection stage and the recorded feature matching data and deviation distribution. For areas where deviation is determined to exist in the initial detection, a texture recognition algorithm is used to extract the texture features of the textile substrate in that area. Cosine similarity is calculated between the texture features of the corresponding area in the standard template. At the same time, a texture consistency judgment index is constructed by combining the periodicity and uniformity parameters of the texture. A pattern judgment algorithm is used to analyze the validity of the overprinting features of the area. The outline features and color features of the overprinted pattern are combined to determine whether the deviation exceeds the allowable range of the overprinting process and whether it is a real deviation generated during the overprinting process. Based on color feature analysis technology, the color transition characteristics of the deviation area and the adjacent overprinting area are compared to eliminate false deviations caused by color halo and color deviation. Based on the comprehensive triple constraint judgment results, if the texture consistency is lower than the preset threshold, the deviation validity is not up to standard, or the color correlation is abnormal, it is judged as a false deviation caused by substrate stretching, wrinkles, uneven warp and weft density, or color halo and is removed. If all three constraints are met, it is judged as a real overprinting deviation. The overprinting features are fitted by a subpixel level edge fitting algorithm, and multi-dimensional deviation values ​​are output to complete the detection of overprinting deviation.

[0008] Preferably, the control quantity budgeting stage is based on the multi-dimensional deviation data obtained in the precise detection stage and the pre-constructed mathematical model of elastic deformation of textile substrate. It decouples and calculates four types of deviations: transverse, longitudinal, rotational, and local deformation, and separates the rigid deviation component that can be independently adjusted by the correction mechanism and the flexible deformation component that needs to be compensated by fine adjustment of the printing roller pressure. By combining time-series feature analysis technology to extract deviation time-series data from multiple consecutive frames of images, and combining the linear velocity collected in real time by the encoder of the production line and the response delay parameters of the correction mechanism, a time-series feature fitting algorithm is used to pre-calculate the decoupled rigid deviation components to generate an initial correction control quantity without correlation influence, thus compensating for the response delay of the mechanism in advance. At the same time, anomaly identification of deviation data is performed through pattern recognition technology to eliminate deviation data anomalies caused by sudden interference.

[0009] Preferably, the closed-loop correction stage is based on the initial correction control quantity without associated influence generated in the control quantity budget stage and the deviation data after anomaly identification. It collects deviation data of multiple consecutive frames of images, and constructs a printing deviation time-series prediction model based on long short-term memory network by combining time-series prediction technology. It inputs historical deviation sequence, production line speed, printing unit spacing parameters and integrates real-time feature matching data to predict the cumulative trend of printing deviation, generates a pre-adjusted correction control quantity, integrates the pre-adjusted control quantity with the feedforward pre-calculated initial correction control quantity and combines the pattern recognition algorithm to adaptively optimize the control quantity, generate the final correction control command and send it to the servo correction mechanism and printing roller pressure adjustment mechanism of the corresponding printing unit. Feature matching technology is used to collect the overprinting features and deviation data of subsequent images in real time, and the correction effect is fed back in a closed loop. Whether the feedback deviation meets the qualified standard is quantified according to the overprinting process level of textile packaging bags. The general grade overprinting requires lateral / longitudinal deviation ≤0.3mm, rotation angle deviation ≤0.3°, and local deformation ≤0.2mm. The precision grade overprinting requires lateral / longitudinal deviation ≤0.1mm, rotation angle deviation ≤0.1°, and local deformation ≤0.08mm. The system judges according to the preset process level. If all indicators meet the standard, it is qualified. If the standard is not met, the overprinting control strategy formulation process is restarted. The weights and correction control amplitude of the long short-term memory network prediction model are dynamically adjusted. At the same time, an anomaly detection algorithm is combined with a feature matching acceleration algorithm to improve the response speed of real-time detection and feedback, taking into account the dynamic characteristics of high-speed production lines.

[0010] Preferably, the model calibration stage is based on the real-time acquisition of overprint features, deviation data, and correction effect feedback information from the closed-loop correction stage. It performs real-time correlation analysis on deviation data, environmental parameters, and equipment operating parameters continuously acquired during production. Data correlation is performed using a synchronization method based on production timestamps, with data collected every 10ms and uniquely timestamped. A correlation dataset is constructed according to the dimensions of "deviation data - environmental parameters (temperature / humidity) - equipment operating parameters (traction speed / printing pressure / line speed)". The dataset is stored by production batch, with each batch's data retained for 30 days after production completion. During correlation analysis, only continuous data from the same batch and under the same operating conditions are selected, eliminating invalid data with mismatched timestamps to ensure the accuracy of the analysis results. An adaptive learning algorithm is used to identify the drift patterns in detection accuracy caused by environmental changes and slight equipment wear. The parameters of the feature extraction algorithm and the deviation judgment threshold are adaptively calibrated. Simultaneously, texture calibration technology is used to dynamically update the texture feature library in the standard template to compensate for errors caused by environmental and equipment factors. An abnormal deviation data early warning mechanism is established using an anomaly detection algorithm. When the deviation data shows a sudden change or the accuracy drift exceeds the preset range, the model calibration process is automatically triggered. The feature matching verification algorithm is used periodically to verify the standard template and the actual overprinted sample. The feature matching verification algorithm adopts the ORB feature matching algorithm consistent with the initial deviation detection. During verification, three indicators are calculated: the overall feature matching degree, the Mark point feature matching degree, and the contour feature matching degree between the actual sample and the standard template. The weights of the three indicators are set to 0.4, 0.3, and 0.3, respectively. The weighted average matching degree is the final verification result. At the same time, the feature points that fail to match are statistically analyzed. When the failure rate is ≤5%, the verification is considered qualified and no template correction is required. When the failure rate is 5%-10%, only the corresponding failed feature areas are corrected to ensure the targeted and efficient template verification.

[0011] This invention also provides a machine vision deviation detection and correction control system for intelligent overprinting and positioning of textile packaging bags, based on the above method, including: The acquisition and preprocessing module, targeting the texture interference characteristics of textile substrates, adopts a dual-layer adaptive filtering and edge enhancement algorithm. It suppresses the frequency domain components of the substrate texture through frequency domain Fourier transform, removes fuzz noise through Gaussian difference filtering, optimizes the adaptive threshold Canny operator to lock the overprinted mark points and the sub-pixel level edges of the pattern, and combines image segmentation technology to complete the separation of the overprinted area and the substrate area, outputting a preprocessed image with texture suppression and clear edges. The standard template construction module selects a standard unbiased overprint sample, extracts and encodes the subpixel-level corner features of the reference color group Mark points through the optimized Harris corner detection operator, and combines contour extraction and pattern recognition technology to extract the overprint pattern contour features and substrate texture features, and integrates them to form a three-dimensional digital standard template containing overprint features, texture features and position parameters, which is then stored in the template library. The initial deviation detection module uses the composite features of the reference color group in the standard template as the registration reference. It adopts the ORB feature matching algorithm, locks the global coordinate system through global rigid registration, and performs local non-rigid registration in combination with the elastic deformation characteristics of the textile substrate. It initially obtains the horizontal, vertical, rotation and local deformation deviation data of each printing color group, and records the feature matching and deviation distribution information. The precision detection module targets areas with deviations in the initial inspection. It compares the consistency of the substrate texture using a texture recognition algorithm, and combines pattern judgment and color feature analysis technology to construct a triple constraint judgment mechanism. This mechanism eliminates false deviations caused by substrate stretching, wrinkles, uneven warp and weft density, and color halo. It also uses a sub-pixel-level edge fitting algorithm to calculate the true deviation and outputs multi-dimensional deviation values. The control quantity budget module combines deviation data with the mathematical model of elastic deformation of textile substrate to decouple the calculation of four types of deviations: lateral, longitudinal, rotational, and local deformation. It separates rigid and flexible deviation components, extracts continuous multi-frame deviation time series data, and combines the production line speed and the response delay of the correction mechanism to pre-calculate and generate an initial correction control quantity without any related influence, thus eliminating sudden interference. The closed-loop correction module, based on the initial correction control quantity, predicts the cumulative trend of deviation through a long short-term memory network, generates a pre-adjusted control quantity, and adaptively fuses and optimizes it with the initial control quantity through a pattern recognition algorithm. It then sends control commands to the servo correction mechanism and the printing roller pressure adjustment mechanism, collects images in real time to provide feedback on the correction effect, and dynamically adjusts parameters. The model self-calibration module performs real-time correlation analysis of production deviation data, environmental parameters, and equipment operating parameters. It identifies the drift patterns in detection accuracy caused by environmental changes and minor equipment wear, updates the textile substrate texture feature library in the template library, and periodically verifies the accuracy of standard templates.

[0012] The beneficial effects of this invention are as follows: 1. This invention designs a dual-layer adaptive filtering and edge enhancement algorithm. First, it uses frequency domain Fourier transform combined with adaptive notch filtering to suppress the frequency domain components of the substrate texture. Then, it uses Gaussian difference filtering to remove high-frequency noise from the fuzz. It also optimizes the adaptive threshold Canny operator to achieve edge locking of overprint features. Combined with image segmentation technology, it isolates the substrate from interference between the substrate and the overprint area. At the same time, it constructs a three-dimensional digital standard template that integrates overprint features, texture features, and position parameters to avoid the substrate texture from overshadowing the overprint features or being misjudged as noise. This achieves accurate extraction of overprint features and improves the overall accuracy of overprint deviation detection for textile packaging bags.

[0013] 2. This invention constructs a triple-constraint judgment mechanism of texture recognition, pattern determination, and color feature analysis, which can accurately identify and eliminate false deviations caused by substrate stretching, wrinkles, uneven warp and weft density, and color halo, avoiding misjudgment by the correction mechanism and thus exacerbating the overprinting error. At the same time, it combines the mathematical model of elastic deformation of textile substrate to decouple the multi-dimensional deviations, separate the rigid and flexible deviation components, realize targeted correction control quantity budgeting, and eliminate the influence of sudden interference on deviation data, so that the detection results are consistent with the actual overprinting conditions, improve the reliability of deviation detection, and increase the product yield.

[0014] 3. The deviation time-series prediction model constructed by this invention based on time-series prediction technology can accurately predict the cumulative trend of printing deviation. Combined with feedforward pre-calculation to compensate for the response delay of the correction mechanism, the final correction command generated by the fusion is sent to the dedicated adjustment mechanism to form a complete closed-loop correction system. At the same time, the model parameters and correction amplitude can be dynamically adjusted through real-time feedback. The model self-calibration module can also adaptively calibrate algorithm parameters and update the texture feature library. The anomaly warning mechanism can automatically trigger the calibration process, realizing effective control of cumulative printing deviation in high-speed production scenarios, avoiding the generation of batch defective products, and improving the production stability of high-speed printing of textile packaging bags. Attached Figure Description

[0015] Figure 1 This is a flowchart of the intelligent overprinting positioning deviation detection and correction control for textile packaging bags according to the present invention; Figure 2 This is a flowchart of the overprinted marker feature extraction process of the present invention; Figure 3 The flowchart illustrates the overprinting correction control strategy of this invention. Detailed Implementation

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

[0017] like Figures 1 to 3 As shown, this embodiment of the invention provides a machine vision deviation detection and correction control method for intelligent overprinting and positioning of textile packaging bags, including the following specific steps: In the acquisition and preprocessing stage, a synchronous acquisition system of a line array camera and a diffuse ring light source linked with the encoder of the traction roller of the printing production line is built to achieve equal-interval trigger acquisition, obtain continuous printing images of textile packaging bags after multi-color overprinting, and eliminate image ghosting and distortion caused by high-speed motion. The line scan camera uses an industrial line scan camera with a resolution of 4096 pixels and a line frequency of 20000Hz. The diffused ring light source uses a white LED light source with an illumination intensity of 8000-12000 lux and a vertical distance of 15-20cm between the light source and the camera lens. The traction roller encoder uses an incremental photoelectric encoder with a resolution of 1024 lines / revolution. The interval for triggering acquisition at equal intervals is set according to the actual interval of the overprinted Mark points on the textile packaging bag, with a trigger interval of 5-10cm to ensure that the overprinted features are fully presented in the acquired image.

[0018] A dual-layer adaptive filtering and edge enhancement algorithm is designed to address the texture interference characteristics of textile substrates. First, the characteristic frequency bands of the periodic warp and weft texture of the textile substrate are decomposed by frequency domain Fourier transform. Adaptive notch filtering is used to suppress the frequency domain components of the substrate texture. Combined with texture analysis methods, the periodic feature parameters of the texture are extracted to achieve the localization and suppression of the texture frequency bands, avoiding the loss of over-filtering features. Frequency domain Fourier transform uses two-dimensional fast Fourier transform to process the acquired grayscale image, converting the image from the spatial domain to the frequency domain. By detecting the texture period, the frequency domain peak region corresponding to the warp and weft texture is located. Adaptive notch filtering sets a filtering window for this peak region. The window radius is dynamically adjusted according to the texture period. When the warp and weft texture period of the textile substrate is 5-20 pixels, the filtering window radius is set to 3-8 frequency domain pixels. The attenuation coefficient of the notch filter is set to 0.8, which only suppresses the texture feature frequency band and retains the frequency domain feature information of the overprinted pattern.

[0019] High-frequency noise from the surface fuzz is then removed using Gaussian difference filtering. This filtering employs two Gaussian kernels for the difference operation: the first kernel has a standard deviation of 1.0 and a size of 3×3, while the second kernel has a standard deviation of 1.5 and a size of 5×5. Pixel-level difference calculations are performed on the filtered image to remove only high-frequency noise from the fuzz with grayscale value abrupt changes ≤15, preserving the grayscale abrupt change features of the overprinting edges and avoiding over-filtering that results in the loss of overprinting details. An edge detection optimization algorithm is used to optimize the adaptive threshold Canny operator, dynamically adjusting the threshold parameters based on the grayscale features of the overprinted pattern and the edge gradient characteristics. This achieves sub-pixel-level edge feature locking between the overprinted mark points and the pattern, while preserving edge detail information. The adaptive threshold Canny operator dynamically adjusts the high and low thresholds based on the grayscale difference between the overprinted area and the substrate area. When the grayscale difference is between 20 and 50, the low threshold is set to 20 and the high threshold to 60; when the grayscale difference is between 50 and 80, the low threshold is set to 40 and the high threshold to 90; when the grayscale difference is greater than 80, the low threshold is set to 60 and the high threshold to 120. The gradient calculation uses the 8-neighbor gradient operator, and the edge connection threshold is set to 5 pixels to achieve sub-pixel-level edge locking between the overprinted mark point and the pattern.

[0020] Image segmentation technology is used to segment the overprinted area and the substrate area of ​​the preprocessed image, further isolating substrate texture interference, and finally outputting a preprocessed image with texture suppression, noise removal, clear edges and complete region segmentation.

[0021] The image segmentation technology employs a segmentation algorithm based on a combination of grayscale thresholding and region growing. First, an initial segmentation threshold is set according to the mean grayscale difference between the overprinted pattern and the substrate. Regions with grayscale values ​​higher than this threshold are marked as candidate overprinted regions. Then, region growing is performed using the overprinted Mark point as the seed point. The growing neighborhood is set to 4 neighborhoods, and the grayscale similarity threshold is set to 0.9. The growth stops when the grayscale change amplitude at the region boundary is greater than 30. Finally, accurate segmentation of the overprinted region and the substrate region is achieved, with segmentation accuracy reaching the pixel level.

[0022] The template construction stage is based on the preprocessed image output from the acquisition and preprocessing stage. It selects standard unbiased textile packaging bag overprint samples, optimizes the Harris corner detection operator using a corner detection algorithm and sets a deformation adaptive weight factor, extracts sub-pixel level corner features of the reference color group Mark points, and performs feature encoding on the edge curvature and grayscale distribution of the Mark points to generate Mark point-specific feature vectors, thereby improving the recognition adaptability of deformed Mark points. The window size of the optimized Harris corner detection operator is set to 7×7, the threshold of the corner response function is set to 0.04, and the neighborhood for non-maximum suppression is set to 3×3. The deformation adaptive weight factor is dynamically allocated according to the deformation degree of the Mark point. When the pixel deformation rate of the Mark point is ≤3%, the corner position weight is set to 1.0 and the grayscale feature weight is set to 1.0. When the deformation rate is 3%-8%, the corner position weight is linearly reduced to 0.7 and the grayscale feature weight is increased to 1.3. When the deformation rate is >8%, the corner position weight is set to 0.7 and the grayscale feature weight is set to 1.3, thereby improving the feature extraction accuracy of deformed Mark points.

[0023] The closed contour feature points of the overprinted pattern are extracted using contour extraction technology. The geometric feature parameters of the contour, including area, perimeter, and eccentricity, are calculated using a contour moment algorithm. The contour moment calculation adopts the central moment calculation method of a 7×7 pixel neighborhood. The area is counted based on the effective pixels enclosed by the contour. The perimeter is calculated by accumulating the connection lengths of the eight neighbors of the contour edge pixels. The eccentricity is adapted to the contour type of the overprinted pattern within a certain range: 0-1 for circular / elliptical overprinted patterns and 0-0.5 for square / rectangular overprinted patterns. All geometric feature parameters are extracted with a precision down to the single pixel level. After extraction, normalization is performed to generate a contour feature descriptor, which is then combined with the texture details of the pattern to generate a contour feature description. To achieve a unique identifier for the overprinted pattern, the Mark point feature vector and contour feature descriptor are fused and a pattern recognition algorithm is used for deformation-invariant encoding to generate a composite feature vector that is robust to rotation, scaling, and local deformation. The deformation-invariant encoding adopts a concatenated encoding method based on scale-invariant features. First, the Mark point feature vector (64-dimensional) and contour feature descriptor (128-dimensional) are fused to obtain a 192-dimensional basic feature. Then, the final 256-dimensional composite feature vector is generated through rotation, scaling, and translation-invariant mapping processing. Each dimension value in the vector is normalized to 0-1. After encoding, the feature vector is optimized for dimensionality reduction, retaining feature dimensions with a contribution rate of over 95%, thereby improving the efficiency of subsequent matching.

[0024] Textile substrate texture features of each region in the standard sample are extracted and a texture feature library is constructed using a texture classification method. This library is then integrated with composite overprinting features and standard overprinting position parameters to form a three-dimensional digital standard template containing overprinting features, texture features, and position parameters, which is stored in the template library. The three dimensions are defined as overprinting feature dimension, substrate texture dimension, and spatial position dimension. Each dimension is stored hierarchically according to feature type. The overprinting feature dimension stores Mark point feature vectors and contour feature descriptors, the texture dimension stores texture feature parameters extracted from the gray-level co-occurrence matrix, and the position dimension stores subpixel-level coordinates and relative offset parameters. The templates are stored in binary format. The template library is categorized and managed into folders according to the substrate type and overprinting specifications of textile packaging bags. It supports quick retrieval and calling by production work order. The storage size of a single template is controlled within 100KB to ensure matching speed.

[0025] Texture feature extraction employs the gray-level co-occurrence matrix method, extracting four core feature parameters: contrast, energy, entropy, and correlation. These features are stored in separate databases based on the substrate type of textile packaging bags (non-woven fabric / cotton linen / Oxford cloth). Each substrate type is further divided into sub-databases based on the printing area. Texture features in each sub-database are extracted using a 5×5 pixel grid, with a set of feature parameters extracted from each block. The constructed texture feature database supports rapid retrieval and similarity comparison by region.

[0026] The deviation detection stage is based on a three-dimensional digital standard template containing overprint features, texture features and position parameters generated in the template construction stage. The ORB feature matching algorithm is adopted with the composite features of the reference color group as the registration reference and a feature point screening mechanism is set. First, the reference color group and the standard template are globally rigidly registered. The cosine similarity of the composite feature vector is calculated by the feature point matching algorithm. Feature point pairs with matching degree higher than the preset threshold are selected to lock the global coordinate system. When using the ORB feature matching algorithm, the cosine similarity preset threshold is set to 0.85, and feature points below this threshold are directly discarded. Feature point detection uses the FAST corner detection operator, with a threshold of 20, a pyramid layer of 4, and a scaling factor of 1.2 for each layer. In local non-rigid registration, the feature point matching weight is dynamically adjusted according to the elastic deformation coefficient of the textile substrate. The deformation coefficient is calculated based on the substrate elongation rate. When the elongation rate is ≤5%, the weight coefficient is set to 1.0; when the elongation rate is 5%-10%, the weight coefficient linearly decreases to 0.7; and when the elongation rate is >10%, the weight coefficient is set to 0.7. Furthermore, the feature points in the deformation region are weighted by their neighborhood, with a neighborhood radius of 5 pixels, to achieve accurate feature matching under elastic deformation.

[0027] For the remaining color sets, local non-rigid registration was performed sequentially. Taking into account the elastic deformation characteristics of the textile substrate, an elastic registration algorithm was used to dynamically adjust the matching weight of feature points. The elastic registration algorithm adopted the thin plate spline interpolation registration method, and the registration grid was divided according to the deformation area of ​​the substrate stretching and wrinkling. The grid size was set to 10×10 pixels. The feature points in each grid were independently registered and calculated. The registration error threshold of the deformation grid was set to 0.05mm. During the registration process, the feature points of the grids that exceeded the error threshold were re-matched to ensure that the registration accuracy of the elastic deformation area was consistent with that of the non-deformation area of ​​the substrate. The lateral deviation, longitudinal deviation, rotation angle deviation and local flexible deformation deviation data of each color set relative to the standard template were initially obtained to complete the initial inspection of the registration deviation. At the same time, the feature matching data and deviation distribution during the initial inspection were recorded.

[0028] The precise detection stage is based on the initial detection data of overprinting deviation and the recorded feature matching data and deviation distribution obtained in the initial deviation detection stage. For areas where deviation is determined in the initial detection, a texture recognition algorithm is used to extract the texture features of the textile substrate in that area. The cosine similarity is calculated between this area and the texture features of the corresponding area in the standard template. At the same time, a texture consistency judgment index is constructed by combining the periodicity and uniformity parameters of the texture. This index is calculated by weighting texture similarity, periodic deviation rate and uniformity deviation rate, with the weights of the three set to 0.5, 0.3 and 0.2, respectively. Texture similarity is the cosine similarity between the measured texture and the standard texture, with a threshold of 0.8. Periodic deviation rate is the percentage difference between the period of the measured texture and the period of the standard texture, with a threshold of ≤10%. Uniformity deviation rate is the percentage difference between the pixel grayscale variance of the measured texture and the standard texture, with a threshold of ≤15%. When all three indicators meet the threshold requirements, the texture consistency is judged to be up to standard. If any indicator does not meet the requirements, it is judged to be a pseudo-deviation related area caused by substrate deformation. A pattern determination algorithm is used to analyze the validity of the overprinting features of the region. The algorithm combines the contour features and color features of the overprinted pattern to determine whether the deviation exceeds the allowable range of the overprinting process and whether it is a real deviation generated during the overprinting process. Pattern determination employs a template matching-based classification pattern recognition algorithm. The allowable range for overprinting is quantified as follows: horizontal / vertical deviation ≤ 0.5mm, rotation angle deviation ≤ 0.5°, and local deformation ≤ 0.3mm. Deviations exceeding this range are considered valid, while those within this range are considered invalid. In color transition characteristic analysis, the average RGB colors of the deviation area and adjacent overprinting areas are extracted, and the color difference is calculated. When the RGB single-channel color difference is ≤ 20, it is determined to be a pseudo-deviation caused by color halo / color shift and is discarded. When the single-channel color difference is > 20, it is determined to be a color feature related to the true overprinting deviation.

[0029] Based on color feature analysis technology, the color transition characteristics of the deviation area and the adjacent overprinted area are compared. First, the image is converted from the RGB color space to the HSV color space. The mean and variance of hue (H), saturation (S), and lightness (V) of the deviation area and the adjacent overprinted area are extracted respectively. When the difference in H channel is ≤10°, the difference in S channel is ≤0.1, and the difference in V channel is ≤0.1, it is determined that the color transition is abnormal due to color halo / color cast. Otherwise, it is determined that the color feature change is caused by overprinting deviation. At the same time, a 5×5 pixel neighborhood mean filter is applied to the edge transition area. To avoid single-pixel color abrupt changes affecting the judgment results, and to eliminate false deviations caused by color halo and color shift, the judgment results are based on the triple constraints. If the texture consistency is lower than the preset threshold, the deviation validity is not up to standard, or the color correlation is abnormal, it is judged as a false deviation caused by substrate stretching, wrinkles, uneven warp and weft density, or color halo and is removed. If all three constraints are met, it is judged as a true overprinting deviation. A subpixel-level edge fitting algorithm is used to fit the overprinting features and output multi-dimensional deviation values ​​with an accuracy of ≤±0.08mm, thus achieving high reliability and subpixel-level accurate detection of overprinting deviation.

[0030] The subpixel-level edge fitting algorithm uses quadratic polynomial interpolation to fit the pixel-level edges extracted by the Canny operator. The fitting neighborhood is set to 3×3 pixels. The gray value of each edge pixel is interpolated to determine the precise coordinates of the subpixel-level edge. During the fitting process, outliers with gray value deviations >10 are removed. The final output horizontal and vertical deviation values ​​are accurate to 0.01mm, the rotation angle deviation is accurate to 0.01°, and the local deformation is accurate to 0.01mm.

[0031] The control quantity budgeting stage is based on the multi-dimensional deviation data obtained in the precise detection stage and the pre-constructed mathematical model of elastic deformation of textile substrate. It decouples and calculates four types of deviations: transverse, longitudinal, rotational, and local deformation, and separates the rigid deviation component that can be independently adjusted by the correction mechanism and the flexible deformation component that needs to be compensated by the pressure of the printing roller. Deviation decoupling determines rigid and flexible deviation components based on deformation recovery characteristics. When the deviation recovery rate is ≤10% after the printing / traction force is removed, it is determined to be a rigid deviation component, which is independently adjusted by the lateral, longitudinal, and rotation adjustment modules of the servo correction mechanism. When the deviation recovery rate is >10%, it is determined to be a flexible deformation component, and the printing roller pressure compensation coefficient is matched according to the degree of deformation. When the flexible deformation is ≤0.2mm, the compensation coefficient is set to 0.8; when it is 0.2-0.5mm, the compensation coefficient linearly increases to 1.2; and when it is >0.5mm, the compensation coefficient is set to 1.2. Flexible deformation deviation compensation is achieved through fine adjustment of printing roller pressure.

[0032] By combining time-series feature analysis technology to extract deviation time-series data from multiple consecutive frames of images, and using a time-series feature fitting algorithm to combine the linear velocity collected in real time by the production line encoder and the response delay parameters of the correction mechanism, feedforward pre-calculation is performed on the decoupled rigid deviation components to generate an initial correction control quantity without correlation influence. The feedforward pre-calculation is time-compensated according to the production line linear velocity and the mechanical response time of the correction mechanism. The mechanical response time is calibrated to 20ms. When the linear velocity is ≤80m / min, the compensation quantity is executed at 90% of the calculated value; when it is 80-120m / min, the compensation quantity is executed at 100% of the calculated value; and when it is >120m / min, the compensation quantity is executed at 110% of the calculated value. During the pre-calculation process, the three types of rigid deviation components—lateral, longitudinal, and rotational—are calculated independently to avoid mutual interference between components. The calculation results are retained to three decimal places to adapt to the adjustment accuracy of the servo mechanism and compensate for the mechanism response delay in advance. At the same time, pattern recognition technology is used to identify anomalies in the deviation data and eliminate deviation data anomalies caused by sudden interference, ensuring the accuracy of the control quantity budget.

[0033] A pattern recognition anomaly detection algorithm based on the 3σ principle is adopted to perform statistical analysis on continuous multi-frame deviation data, calculate the mean μ and standard deviation σ of each type of deviation data. When the deviation data of a single frame exceeds the range of [μ-3σ, μ+3σ], it is determined to be a deviation anomaly caused by sudden interference, and the deviation data of this frame will not participate in the calculation of the initial correction control quantity. At the same time, a continuous anomaly judgment rule is set. When three consecutive frames of data exceed the threshold, a sudden interference warning is triggered, the calculation of the batch of control quantities is suspended, and feedback is sent to the industrial control system.

[0034] The closed-loop correction stage is based on the initial correction control quantity without related influence generated in the control quantity budget stage and the deviation data after anomaly identification. It collects deviation data of multiple consecutive frames of images, and constructs a printing deviation time-series prediction model based on long short-term memory network by combining time-series prediction technology. It inputs historical deviation sequence, production line speed, printing unit spacing parameters and integrates real-time feature matching data to predict the cumulative trend of printing deviation in the next 3-5 printing units and generate a pre-adjusted correction control quantity. This Long Short-Term Memory (LSTM) bias time-series prediction model consists of an input layer, two hidden layers, and an output layer. The input layer is a fully connected layer with 16 neurons to accommodate multidimensional input data. The first hidden layer has 64 LSTM neurons, and the second hidden layer has 32 LSTM neurons. A Dropout layer is used to suppress overfitting with a Dropout coefficient of 0.2. The output layer is a fully connected layer with 4 neurons, corresponding to the predicted values ​​of four types of bias: lateral, longitudinal, rotational, and local deformation. The model is trained using the Adam optimizer with an initial learning rate of 0.001, an iteration batch size of 32, and 100 training epochs. The loss function is the mean squared error. The training dataset consists of 10,000+ sets of deviation time-series data collected in the overprinting production of textile packaging bags. It is divided into training set, validation set, and test set in an 8:1:1 ratio. During model training, training is stopped when the loss function of the validation set does not decrease for 10 consecutive rounds, and the optimal model parameters are saved. The model input data is normalized to a range of [0, 1]. The input historical deviation sequence is taken as deviation data of 50 consecutive frames. The production line speed and printing unit spacing parameters are scalar quantized and incorporated into the input features. The output is the four types of deviation prediction values ​​for the next 3-5 printing units, which are directly used as the basis data for pre-adjustment and correction control.

[0035] The system integrates the pre-adjusted control quantity with the feedforward pre-calculated initial correction control quantity and combines it with a pattern recognition algorithm to adaptively optimize the control quantity, generating the final correction control command and issuing it to the servo correction mechanism and printing roller pressure adjustment mechanism of the corresponding printing unit. The two types of mechanisms adopt a timing sequence of servo correction first and pressure compensation later. The servo correction mechanism takes priority in execution response, with an action completion time ≤30ms, a horizontal / vertical adjustment step size of 0.01mm, and a rotation adjustment step size of 0.01°. The printing roller pressure adjustment mechanism starts within 5ms after the servo correction action is completed, with a pressure adjustment step size of 0.05MPa and an adjustment range of ±20% of the preset process pressure. When the flexible deformation is ≤0.1mm, only the servo correction mechanism is activated, and the pressure adjustment is not triggered to reduce the invalid actions of the mechanism.

[0036] Feature matching technology is used to collect the overprinting features and deviation data of subsequent images in real time, and closed-loop feedback is provided for the correction effect. The weights and correction control amplitude of the long short-term memory network prediction model are dynamically adjusted. At the same time, anomaly detection algorithms are combined to avoid over-correction and under-correction. Feature matching acceleration algorithms are used to improve the response speed of real-time detection and feedback for the dynamic characteristics of high-speed production lines. It is ensured that the correction response time is ≤50ms under production conditions with online speed ≥120m / min, so as to effectively control the cumulative deviation of multi-color overprinting.

[0037] The feature matching acceleration algorithm adopts an ORB feature matching optimization algorithm based on feature point dimensionality reduction. Principal component analysis (PCA) is used to reduce the dimensionality of the extracted ORB feature points from 256 dimensions to 64 dimensions. At the same time, a feature point screening rule is set for high-speed production scenarios. When the line speed is ≥120m / min, the number of feature points is screened at 70% of the basic number to retain high-response feature points. The algorithm is accelerated by GPU parallel computing on the embedded industrial control board with 16 parallel threads to ensure that the feature matching response time is ≤20ms when the line speed is ≥120m / min, which meets the real-time requirements of high-speed production.

[0038] The correction control amplitude is dynamically adjusted according to the magnitude and trend of the deviation. When the deviation value is ≤0.2mm and the trend is stable, the correction amplitude is 80% of the calculated value to avoid overcorrection. When the deviation value is 0.2-0.5mm and the trend is slightly upward, the correction amplitude is 100% of the calculated value. When the deviation value is >0.5mm or the trend is rapidly upward, the correction amplitude is 110% of the calculated value. At the same time, the LSTM model weights are adjusted according to the feedback of the correction effect. When the correction error is ≤0.05mm, the corresponding feature weight is increased by 0.1, and when the correction error is >0.1mm, the corresponding feature weight is decreased by 0.1. The weight adjustment range is 0.1-1.5.

[0039] The model calibration stage, based on the real-time acquisition of overprinting features, deviation data, and correction effect feedback information from the closed-loop correction stage, performs real-time correlation analysis on deviation data continuously collected during production, environmental parameters (including temperature and humidity), and equipment operating parameters (including traction speed and printing pressure). An adaptive learning algorithm is used to identify the drift patterns in detection accuracy caused by environmental changes and minor equipment wear. The parameters and deviation judgment thresholds of the feature extraction algorithm are adaptively calibrated. Simultaneously, texture calibration technology is used to dynamically update the texture feature library in the standard template. The technique extracts substrate texture samples in batches according to changes in the production environment. Each batch extracts 200 5×5 pixel texture blocks. After removing abnormal texture blocks with abrupt grayscale changes, the contrast, energy, entropy, and correlation feature parameters of the grayscale co-occurrence matrix are recalculated. The new parameters are then weighted and fused with the parameters in the original feature library. The fusion weight is set according to the magnitude of environmental changes: when the temperature / humidity change is ≤5%, the weight of the new parameter is set to 0.2; when the change is 5%-10%, the weight is set to 0.5; and when the change is >10%, the weight is set to 0.8. The updated texture feature library is synchronized to the template library in real time to ensure the accuracy of texture matching. An adaptive learning algorithm combining linear regression optimized by gradient descent and incremental reinforcement learning is used. The algorithm is trained using continuously collected deviation data, environmental parameters (temperature: 0-50℃, humidity: 30%-90%), and equipment operating parameters (traction speed: 0-150m / min, printing pressure: 0-10MPa) as training samples, and the output is the calibration of feature extraction parameters and deviation judgment threshold. The algorithm iteration steps are as follows: 1. Real-time acquisition of single batch production data (data corresponding to 500 frames of images), and labeling of accuracy drift values; 2. Inputting sample data into the algorithm model, fitting the correlation between parameters and drift values ​​through linear regression, and calculating the initial calibration amount; 3. Using the improvement rate of detection accuracy after calibration as the reward function, the reinforcement learning module incrementally corrects the initial calibration amount, with the reward function threshold set to 95%; 4. Applying the corrected calibration amount to actual detection, collecting feedback data for verification, and repeating the iteration if the accuracy does not meet the standard, with an iteration step size of 0.01. The key parameters of the algorithm are: learning rate 0.005, iteration convergence error ≤0.001, and sample update cycle once every 1000 meters of textile packaging bags produced, so as to realize dynamic adaptive calibration of feature extraction parameters and deviation judgment threshold.

[0040] An abnormal deviation data early warning mechanism is established using an anomaly detection algorithm. When deviation data shows sudden changes or accuracy drifts beyond the preset range, the model calibration process is automatically triggered without manual intervention. A feature matching verification algorithm is used periodically to verify the standard template and the actual overprinted samples. The verification cycle is set according to both production time and production batch. Verification is automatically triggered after 4 hours of continuous production or after 5 production batches. During verification, 10 overprinted samples without deviations from actual production are taken and their features are matched with the standard template. The average matching degree is calculated. When the average matching degree is ≥98%, the template is considered to be qualified and no update is required. When the average matching degree is 95%-98%, local feature corrections are made to the template. When the average matching degree is <95%, the overprinting features and texture features are re-extracted, and a new standard template is generated to replace the original template. The verification process is fully automated, and the verification results are uploaded to the production control system in real time for storage.

[0041] The threshold for judging sudden changes in deviation data is: the change in deviation data in a single frame is greater than 50% compared to the previous frame, or the cumulative change in deviation data in 5 consecutive frames is greater than 80%; the threshold for judging drift in detection accuracy is: the difference between the actual detection deviation and the true deviation is greater than 0.1mm, or the pass rate of detection accuracy in 100 consecutive frames is less than 95%; after the calibration process is triggered, the system first pauses the issuance of new correction instructions, calls the latest standard sample in the template library to re-extract features, and then performs a one-time calibration on the feature extraction parameters and deviation judgment threshold. The calibration process takes ≤2s. After completion, the detection and correction work is automatically resumed without manual intervention.

[0042] This invention also provides a machine vision deviation detection and correction control system for intelligent overprinting and positioning of textile packaging bags, based on the above method, including: The acquisition and preprocessing module, targeting the texture interference characteristics of textile substrates, adopts a dual-layer adaptive filtering and edge enhancement algorithm. It suppresses the frequency domain components of the substrate texture through frequency domain Fourier transform, removes fuzz noise through Gaussian difference filtering, optimizes the adaptive threshold Canny operator to lock the overprinted mark points and the sub-pixel level edges of the pattern, and combines image segmentation technology to complete the separation of the overprinted area and the substrate area, outputting a preprocessed image with texture suppression and clear edges. The standard template construction module selects a standard unbiased overprint sample, extracts and encodes the subpixel-level corner features of the reference color group Mark points through the optimized Harris corner detection operator, and combines contour extraction and pattern recognition technology to extract the overprint pattern contour features and substrate texture features, and integrates them to form a three-dimensional digital standard template containing overprint features, texture features and position parameters, which is then stored in the template library. The initial deviation detection module uses the composite features of the reference color group in the standard template as the registration reference. It adopts the ORB feature matching algorithm, locks the global coordinate system through global rigid registration, and performs local non-rigid registration in combination with the elastic deformation characteristics of the textile substrate. It initially obtains the horizontal, vertical, rotation and local deformation deviation data of each printing color group, and records the feature matching and deviation distribution information. The precision detection module targets areas with deviations in the initial inspection. It compares the consistency of the substrate texture using a texture recognition algorithm, and combines pattern judgment and color feature analysis technology to construct a triple constraint judgment mechanism. This mechanism eliminates false deviations caused by substrate stretching, wrinkles, uneven warp and weft density, and color halo. It also uses a sub-pixel-level edge fitting algorithm to calculate the true deviation and outputs multi-dimensional deviation values. The control quantity budget module combines deviation data with the mathematical model of elastic deformation of textile substrate to decouple the calculation of four types of deviations: lateral, longitudinal, rotational, and local deformation. It separates rigid and flexible deviation components, extracts continuous multi-frame deviation time series data, and combines the production line speed and the response delay of the correction mechanism to pre-calculate and generate an initial correction control quantity without any related influence, thus eliminating sudden interference. The closed-loop correction module, based on the initial correction control quantity, predicts the cumulative deviation trend through a long short-term memory network, generates a pre-adjusted control quantity, and adaptively fuses and optimizes it with the initial control quantity using a pattern recognition algorithm. Control commands are then sent to the servo correction mechanism and the printing roller pressure adjustment mechanism. The servo correction mechanism adopts a three-axis linkage adjustment mode; horizontal and vertical adjustments are driven by stepper motors, and rotational adjustments are driven by servo motors. During adjustment, a maximum adjustment range limit is set: the maximum horizontal / vertical adjustment range in a single operation is ≤0.5mm, and the maximum rotational adjustment range in a single operation is ≤0.5°, to avoid new registration deviations due to excessive single adjustment ranges. The printing roller pressure adjustment mechanism adjusts by printing unit zone, with each printing unit independently set with a pressure compensation value. The pressure difference between adjacent units is ≤0.1MPa to prevent new wrinkles and deformations in the substrate due to uneven pressure. Real-time image acquisition provides feedback on the correction effect, and parameters are dynamically adjusted. The model self-calibration module performs real-time correlation analysis of production deviation data, environmental parameters, and equipment operating parameters. It identifies the drift patterns in detection accuracy caused by environmental changes and minor equipment wear, updates the textile substrate texture feature library in the template library, and periodically verifies the accuracy of standard templates.

[0043] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0044] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine vision deviation detection and correction control method for intelligent overprinting and positioning of textile packaging bags, characterized in that, The specific steps include the following: In the acquisition and preprocessing stage, a synchronous acquisition system linked with the traction roller encoder of the printing production line is built to acquire multi-color overprinted continuous printing images; in response to the texture interference of textile substrates, an appropriate dual-layer adaptive filtering and edge enhancement processing is adopted to complete texture suppression, noise removal, edge enhancement and region segmentation, and output preprocessed images. In the template construction stage, standard unbiased overprint samples are selected. Overprint features are extracted and encoded through optimized corner detection, contour extraction and pattern recognition technologies. A textile substrate texture feature library is constructed. Overprint features, texture features and position parameters are integrated to form a three-dimensional digital standard template and store it. In the initial deviation detection stage, the composite features of the reference color group are used as the registration benchmark. Through feature matching and registration technology, the deviation data of each printing color group is initially obtained and relevant information is recorded. In the accurate detection stage, for areas with deviations, a triple constraint judgment mechanism is constructed through texture recognition, pattern determination and color feature analysis to eliminate false deviations and fit the real overprinting deviations, and output multi-dimensional deviation values. In the control quantity budgeting stage, deviation decoupling calculation is performed by combining the mathematical model of elastic deformation of textile substrate to generate initial correction control quantities without related influences and eliminate abnormal interference. In the closed-loop correction stage, a time-series prediction model is constructed to predict the deviation trend and generate pre-adjustment control quantity. After fusion and optimization, the final correction command is issued. Through real-time feedback, the parameters are dynamically adjusted to ensure the control of cumulative overprint deviation in high-speed scenarios. During the model calibration phase, the deviation data, environmental parameters, and equipment operating parameters are correlated and analyzed to identify the drift pattern of detection accuracy and adaptively calibrate the feature extraction parameters and deviation judgment threshold. An anomaly early warning mechanism is established, and the calibration process is automatically triggered when the deviation changes abruptly or the accuracy drift exceeds the limit. The matching degree between the standard template and the actual overprinted sample is checked regularly.

2. The machine vision deviation detection and correction control method for intelligent overprinting positioning of textile packaging bags according to claim 1, characterized in that, The acquisition and preprocessing stage employs a synchronous acquisition system consisting of a linear array camera and a diffuse ring light source, linked with a traction roller encoder to achieve equidistant trigger acquisition, eliminating image trailing and distortion caused by high-speed motion. In the dual-layer adaptive filtering and edge enhancement processing, the frequency domain components of the periodic latitude and longitude texture of the substrate are first suppressed by frequency domain Fourier transform combined with adaptive notch filtering, and then surface fuzz noise is removed by Gaussian difference filtering. The adaptive threshold Canny operator is optimized to lock the overprinted mark points and pattern edge features, and the overprinted area and substrate area are separated by image segmentation technology.

3. The machine vision deviation detection and correction control method for intelligent overprinting positioning of textile packaging bags according to claim 2, characterized in that, The template construction stage selects standard samples based on the preprocessed image, optimizes the Harris corner detection operator and sets a deformation adaptive weight factor, extracts and encodes sub-pixel level corner features of the reference color group Mark points; Extract the closed contour feature points of the overprinted pattern, calculate the geometric feature parameters including the area, perimeter, and eccentricity of the contour, and generate the contour feature descriptor. Integrate the sub-pixel corner features of Mark points with the contour feature descriptor to encode deformation invariance. Integrate the overprinted composite features, the substrate texture feature library, and standard position parameters to form a three-dimensional digital standard template and store it in the template library.

4. The machine vision deviation detection and correction control method for intelligent overprinting positioning of textile packaging bags according to claim 3, characterized in that, The initial deviation detection stage uses the ORB feature matching algorithm to calculate the cosine similarity between each set of printing color groups and the composite feature vector of the standard template, and selects high-matching feature point pairs with a matching degree higher than a preset threshold; combined with the elastic deformation characteristics of the textile substrate, local non-rigid registration is performed, the feature point matching weight is dynamically adjusted, and the lateral, longitudinal, rotation angle and local flexible deformation deviation data are initially obtained, and the feature matching and deviation distribution information are recorded.

5. The machine vision deviation detection and correction control method for intelligent overprinting positioning of textile packaging bags according to claim 4, characterized in that, The precise detection stage targets the initial detection deviation area, extracts the substrate texture features and compares them with the corresponding area of ​​the standard template using cosine similarity, and constructs a consistency judgment index by combining texture periodicity and uniformity; analyzes the effectiveness of overprint feature deviation by combining pattern judgment with the allowable range of overprinting process, and eliminates false deviations caused by color halo and color deviation by combining color transition characteristics; calculates the real deviation using a sub-pixel level edge fitting algorithm and outputs multi-dimensional deviation values.

6. The machine vision deviation detection and correction control method for intelligent overprinting positioning of textile packaging bags according to claim 5, characterized in that, The control quantity budgeting stage decouples the calculation of four types of deviations: lateral, longitudinal, rotational, and local deformation. It separates the rigid deviation component that can be independently adjusted by the correction mechanism and the flexible deformation component that needs to be compensated by fine-tuning the printing roller pressure. It extracts the deviation time sequence data of multiple consecutive frames, and performs feedforward pre-calculation in combination with the production line speed and the response delay of the correction mechanism to generate the initial correction control quantity. It also eliminates deviation anomalies caused by sudden interference through pattern recognition.

7. The machine vision deviation detection and correction control method for intelligent overprinting positioning of textile packaging bags according to claim 6, characterized in that, The closed-loop correction stage employs a long short-term memory network to construct a deviation time-series prediction model. Input parameters, including historical deviation sequences and production line speed, are used to predict the cumulative deviation trend and generate a pre-adjustment control quantity. The pre-adjustment control quantity is fused with the initial control quantity and adaptively optimized using a pattern recognition algorithm. Commands are then sent to the servo correction mechanism and the printing roller pressure adjustment mechanism. Real-time image acquisition provides feedback on the correction effect, dynamically adjusting the model weights and correction amplitude. A feature matching acceleration algorithm is used to improve the response speed.

8. The machine vision deviation detection and correction control method for intelligent overprinting positioning of textile packaging bags according to claim 7, characterized in that, The model calibration phase involves correlation analysis of deviation data, environmental parameters, and equipment operating parameters. An adaptive learning algorithm is used to identify the accuracy drift pattern, calibrate feature extraction parameters and deviation judgment thresholds, dynamically update the texture feature library, establish an anomaly early warning mechanism, trigger the calibration process when deviation changes abruptly or accuracy drift exceeds the limit, and periodically verify the matching degree between the standard template and the actual sample.

9. A machine vision deviation detection and correction control system for intelligent overprinting and positioning of textile packaging bags, based on the method described in any one of claims 1-8, characterized in that, include: The acquisition and preprocessing module uses a dual-layer adaptive filtering and edge enhancement algorithm to suppress substrate texture and remove fuzz noise. It optimizes the adaptive threshold Canny operator to lock overprinted edge features and combines image segmentation to complete region separation, outputting a preprocessed image. The standard template construction module selects standard unbiased samples, and through optimized corner detection, contour extraction and pattern recognition technologies, extracts overprint features and substrate texture features, integrates them to form a three-dimensional digital standard template and stores it. The initial deviation detection module uses the composite features of the reference color group as the registration benchmark. It first performs global rigid registration between the reference color group and the standard template using the ORB feature matching algorithm to lock the global coordinate system. Then, it combines the elastic deformation characteristics to perform local non-rigid registration, initially acquiring various deviation data and recording relevant information. The precision detection module constructs a triple constraint judgment mechanism for the initial detection deviation area, eliminates false deviations, uses a sub-pixel level edge fitting algorithm to calculate the true deviation, and outputs multi-dimensional deviation values. The control quantity budget module decouples the deviation by combining the elastic deformation mathematical model, separates the rigid and flexible deviation components, pre-calculates and generates the initial correction control quantity, and eliminates sudden interference. The closed-loop correction module predicts the deviation trend through a long short-term memory network, integrates the pre-adjusted control quantity and the initial control quantity and optimizes its distribution, and provides real-time feedback on adjustment parameters. The model self-calibration module correlates and analyzes various operating parameters, identifies accuracy drift patterns and calibrates parameters, updates the texture feature library, and periodically verifies the accuracy of standard templates.

Citation Information

Patent Citations

  • Printing and overprinting monitoring system and monitoring method thereof

    CN110875996A

  • Automatic pattern overprinting and printing method for textile fabric

    CN119814935A