First print alignment monitoring system based on YOLOv8 algorithm

Abstract

The present invention relates to the technical field of printing production systems, and discloses a first print alignment monitoring system based on the YOLOv8 algorithm, including: a visual detection module installed on a printing line, the visual detection module is used to obtain image data of printed objects on the printing line; a printing original manuscript scanning module is used to obtain original printing manuscript image data as first print alignment comparison data; a printing feature extraction module is used to identify and extract feature point parameters of the image data of printed objects on the printing line and feature point parameters of the original printing manuscript image data, the feature point parameters are shape parameters, which include center coordinates, length, width and tilt angle. In view of the high production volume and high-speed printing characteristics of the printing production line, the present invention improves the visual anomaly perception AI model to meet the efficient alignment requirements of printing production, and is applied to the first print alignment and subsequent printing detection work of the printing production line.

Description

Technical Field

[0001] The present invention relates to the technical field of printing production systems, and more specifically, to a first-print registration monitoring system based on the YOLOv8 algorithm. Background Art

[0002] First print registration refers to the process of comparing the printed manuscript with the finished part of the product during the printing process to ensure the accurate position and color consistency of the printed product. High-precision first print registration uses precise positioning equipment and high-resolution image processing technology to easily detect any minor deviations or errors and correct them in time.

[0003] In order to ensure recognition accuracy, the existing AI recognition takes a long time. For example, the SSD algorithm uses basic convolutional models (such as VGG, ResNet, etc.) to extract features and performs multi-scale target detection through multiple convolutional layers and feature layers. The depth and complexity of these networks may result in a relatively large amount of calculation, so SSD may be slightly slower than YOLOv8 in detection speed, which makes it unsuitable for target detection and recognition tasks in the high-speed production process of printing production lines. Summary of the invention

[0004] In view of the problems existing in the prior art, the object of the present invention is to provide a first-print registration monitoring system based on the YOLOv8 algorithm, comprising: a visual detection module installed on a printing line, the visual detection module being used to obtain image data of a printed object on the printing line;

[0005] The printing original manuscript scanning module is used to obtain the original printing manuscript image data as the first printing plate comparison data;

[0006] A printing feature extraction module is used to identify and extract feature point parameters of the printed object image data on the printing line and the feature point parameters of the original printed manuscript image data, wherein the feature point parameters are shape parameters, which include center coordinates, length, width and tilt angle;

[0007] The YOLOv8 target recognition model is used to identify the feature points of the image data of the printed object collected by the visual inspection module on the printing line and the feature points of the original printed manuscript image data collected by the printing manuscript scanning module, and clearly identify the registration differences;

[0008] Adaptive recognition optimization module, used to optimize the anchor-based rotation box detection step in the YOLOv8 object recognition model;

[0009] A model deep learning training database, wherein the model deep learning training database is used to store image data of printed objects on a printing line and image data of original printed manuscripts;

[0010] The optimization method of the adaptive recognition optimization module is to learn a bounding box, and the oriented bounding box of the object replaces the rotation box of the anchor used in the YOLOv8 target recognition model. The specific steps are as follows:

[0011] The original printed manuscript image data stored in the model deep learning training database is used as the training set. The adaptive recognition optimization module uses the training set to train the YOLOv8 target recognition model and extracts the feature point parameter information of the image data in the training set, that is, obtains the center coordinates, length, width and tilt angle as the real frame parameters;

[0012] The obtained center coordinates, length, width and tilt angle parameters are imported into the anchor of the original YOLOv8 in the YOLOv8 target recognition model to obtain the predicted box parameters, and then the predicted box parameters are compared with the real box parameters to calculate the loss function value. The original YOLOv8 of the YOLOv8 target recognition model is guided to learn according to the loss function value until the loss function value is the same as the preset threshold. The learning is stopped, and the combination of the predicted boxes of all feature points in the original printed manuscript image data is used as the bounding box to obtain the optimized YOLOv8.

[0013] The data interaction module is used to output the detected defect targets and adjust the printing content through the printing equipment.

[0014] As a further description of the above technical solution:

[0015] The YOLOv8 target recognition model uses the improved YOLOv8 to input the image data of the printed object on the printing line for recognition and detection, and uses a bounding box to replace the rotation box combination of multiple anchors to directly perform feature recognition on the image data of the printed object on the printing line. When the image data of the printed object on the printing line is recognized by the bounding box, it is divided into the following two recognition situations:

[0016] If the bounding box completely covers the recognition target, the image data recognition of the printed object can be completed directly, and the error is within the preset threshold, so it is passed;

[0017] If the bounding box cannot completely cover the recognition target, it means that the image data of the printed object does not meet the recognition requirements, the error exceeds the preset threshold, and there is a defect.

[0018] As a further description of the above technical solution:

[0019] The optimization method of the adaptive recognition optimization module is to learn a bounding box, and the oriented bounding box of the object replaces the rotation box of the anchor used in the YOLOv8 target recognition model. The specific steps are as follows:

[0020] The original printed manuscript image data stored in the model deep learning training database is used as a training set. The adaptive recognition optimization module 5 uses the training set to train the YOLOv8 target recognition model 4, and extracts the feature point parameter information of the image data in the training set, that is, obtains the center coordinates, length, width and tilt angle as the real frame parameters;

[0021] The obtained center coordinates, length, width and tilt angle parameters are imported into the anchor of the original YOLOv8 in the YOLOv8 target recognition model to obtain the predicted box parameters, and then the predicted box parameters are compared with the true box parameters to calculate the loss function value. The original YOLOv8 of the YOLOv8 target recognition model is guided to learn according to the loss function value until the loss function value is the same as the preset threshold. The learning is stopped, and the combination of the predicted boxes of all feature points in the original printed manuscript image data is used as the bounding box to obtain the optimized YOLOv8.

[0022] As a further description of the above technical solution:

[0023] The original YOLOv8 guided learning method of the YOLOv8 target recognition model is as follows:

[0024] The original YOLOv8 model executes the back-propagation algorithm to calculate the gradient of the loss function with respect to each parameter in the model, starting from the loss function and back-propagating the gradient along each layer of the original YOLOv8 network;

[0025] After obtaining the gradient, the model uses the gradient descent algorithm to update the parameters. For each parameter, the update rule is as follows:

[0026] parameter 新 =Parameter 旧 - learning rate × gradient;

[0027] Among them, the learning rate is a pre-set hyperparameter that determines the step size of the parameter update, and the gradient is the derivative of the loss function with respect to the parameter, which indicates the steepest descent direction of the loss function near the parameter value;

[0028] Then set the parameter 新 Replace the parameters in the original YOLOv8 旧 , again import the obtained center coordinates, length, width and tilt angle parameters into the original YOLOv8 anchor in the YOLOv8 target recognition model to obtain the predicted box parameters, and then compare the predicted box parameters with the true box parameters to calculate the loss function value;

[0029] Repeat the above steps until the loss function value is the same as the preset threshold and stop learning.

[0030] As a further description of the above technical solution:

[0031] The visual inspection module adopts a CCA chip linear array camera and a fill light source which is also set up on the printing line.

[0032] Compared with the prior art, the advantages of the present invention are:

[0033] This solution improves the visual anomaly perception AI model in view of the high production volume and high-speed printing characteristics of the printing production line. The bounding box that meets the printing error range replaces the original anchor's rotation box for direct recognition, improves the recognition efficiency of printed objects, and maximizes the recognition accuracy under high-speed printing to meet the efficient plate alignment requirements of printing production. It is applied to the first print plate alignment and subsequent printing inspection work of the printing production line. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram of the principle of the present invention;

[0035] Figure 2 It is a schematic diagram of the optimization process of the adaptive recognition optimization module of the present invention;

[0036] Figure 3 It is a schematic diagram of the principle of the adaptive recognition optimization module of the present invention.

[0037] Description of the numbers in the figure:

[0038] 1. Visual inspection module; 2. Printing original scanning module; 3. Printing feature extraction module; 4. YOLOv8 target recognition model; 5. Adaptive recognition optimization module; 6. Model deep learning training database; 7. Data interaction module. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention;

[0040] See also Figure 1 and Figure 2 , the present invention provides the following embodiments:

[0041] The first printing registration monitoring system based on the YOLOv8 algorithm includes: a visual detection module 1 installed on a printing line, the visual detection module 1 is used to obtain image data of a printed object on the printing line;

[0042] The printing original manuscript scanning module 2 is used to obtain the original printing manuscript image data as the first printing plate comparison data;

[0043] The printing feature extraction module 3 is used to identify and extract the feature point parameters of the printed object image data on the printing line and the feature point parameters of the original printed manuscript image data. The feature point parameters are shape parameters, which include center coordinates, length, width and tilt angle;

[0044] YOLOv8 target recognition model 4, used to identify the feature points of the image data of the printed object on the printing line collected by the visual inspection module 1 on the printing line and the feature points of the original printing manuscript image data of the printing original manuscript scanning module 2, and clearly identify the registration difference;

[0045] The adaptive recognition optimization module 5 is used to optimize the anchor-based rotation box detection step in the YOLOv8 target recognition model 4;

[0046] A model deep learning training database 6, which is used to store image data of printed objects on the printing line and image data of original printed manuscripts;

[0047] The data interaction module 7 is used to output the detected defect targets and adjust the printed content through the printing equipment;

[0048] Among them, the visual inspection module 1 adopts a CCA chip linear array camera and a fill light source which is also installed on the printing line.

[0049] The CCA chip linear array camera of the visual inspection module 1 is used to photograph the initial draft to obtain the image data of the printed part, and the initial draft is scanned by the printed original draft scanning module 2 to obtain the image data of the original draft, and the photographed image data of the printed part and the image data of the original draft are uploaded to the printing feature extraction module 3 and the model deep learning training database 6, and the image data is recognized by the YOLOv8 target recognition model 4, and the defect targets of the printed manuscript are recognized based on the original manuscript, and then the defect targets of the first printing are uploaded to the printing equipment through the defect analysis module 8 for adjustment, so as to quickly obtain the defect data of the printed matter on the printing line, and according to the detection data, the first printing is replaced, changed and adjusted in time, so as to facilitate the proofreading of the printing style;

[0050] Among them, the YOLOv8 target detection algorithm can simultaneously predict the positions and categories of multiple targets in an image through a forward propagation process. Compared with traditional target detection algorithms, YOLOv8 has faster speed and higher real-time performance. It is suitable for target detection and recognition tasks in the high-speed production process of printing production lines, and ensures the accuracy of detection and recognition.

[0051] See also Figure 2 The optimization method of the adaptive recognition optimization module 5 is to replace the rotation box of the anchor used in the YOLOv8 target recognition model 4 with the oriented bounding box of the object by learning a bounding box. The specific steps are as follows:

[0052] The original printed manuscript image data stored in the model deep learning training database 6 is used as a training set, and the training set is used for training and optimization of the adaptive recognition optimization module 5, and the feature point parameter information of the image data in the training set is extracted, that is, the center coordinates, length, width and tilt angle are obtained as the real frame parameters;

[0053] The obtained center coordinates, length, width and tilt angle parameters are imported into the anchor of the original YOLOv8 in the YOLOv8 target recognition model 4 to obtain the predicted box parameters, and then the predicted box parameters are compared with the real box parameters to calculate the loss function value. The original YOLOv8 of the YOLOv8 target recognition model 4 is guided to learn according to the loss function value until the loss function value is the same as the preset threshold value. The preset threshold value is determined according to the printing requirements of the printed matter. The combination of the predicted boxes of all feature points in the original printed manuscript image data is used as the bounding box to obtain the optimized YOLOv8;

[0054] YOLOv8 target recognition model 4 uses the improved YOLOv8 to input the image data of the printed object on the printing line for recognition and detection, and uses a bounding box to replace the rotation box combination of multiple anchors to directly perform feature recognition on the image data of the printed object on the printing line. When the image data of the printed object on the printing line is recognized by the bounding box, it is divided into the following two recognition situations:

[0055] If the bounding box completely covers the recognition target, the image data recognition of the printed object can be completed directly, and the error is within the preset threshold, so it is passed;

[0056] If the bounding box cannot completely cover the recognition target, it means that the image data of the printed object does not meet the recognition requirements, the error exceeds the preset threshold, and there is a defect;

[0057] The data interaction module 7 is used to output the detected defect targets and adjust the printed content through the printing equipment.

[0058] The original YOLOv8 model uses anchors to assist in object detection and guides the model to learn correct object detection by calculating the loss function;

[0059] The loss function generally includes measurements of position error, size error, etc. to ensure that the predicted box output by the model is as close as possible to the true box.

[0060] Redesigning the anchor and loss function of YOLOv8 according to the above method can be adjusted according to the specific task and data set to obtain better target detection performance;

[0061] There are some disadvantages of anchor-based rotation box detection in target detection: the design of anchor is very complicated. It is necessary to design different proportions and sizes. The positive and negative samples will be very unbalanced, which will lead to slow training and failure to achieve the optimal effect. However, for any rotated object, the width and height actually depend on different rotation coordinate systems, so direct regression will be somewhat difficult. The anchor box of the YOLOv8 target recognition model 4 is optimized by the adaptive recognition optimization module 5 and applied to the detection of oriented objects. However, it does not directly regress w, h and θ, but learns a bounding box and then obtains the oriented bounding box of the object. This can avoid the difficulty of regressing w, h and θ in the anchor-based rotation box detection process and improve the efficiency of anchor-based rotation box detection.

[0062] Among them, the original YOLOv8 guided learning method of the YOLOv8 target recognition model 4 is as follows:

[0063] The original YOLOv8 model executes the back-propagation algorithm to calculate the gradient of the loss function with respect to each parameter in the model, starting from the loss function and back-propagating the gradient along each layer of the original YOLOv8 network;

[0064] After obtaining the gradient, the model uses the gradient descent algorithm to update the parameters. For each parameter, the update rule is as follows:

[0065] parameter 新 =Parameter 旧 - learning rate × gradient;

[0066] Among them, the learning rate is a pre-set hyperparameter that determines the step size of the parameter update, and the gradient is the derivative of the loss function with respect to the parameter, which indicates the steepest descent direction of the loss function near the parameter value;

[0067] Then set the parameter 新 Replace the parameters in the original YOLOv8 旧 , again import the obtained center coordinates, length, width and tilt angle parameters into the original YOLOv8 anchor in the YOLOv8 target recognition model 4 to obtain the predicted box parameters, and then compare the predicted box parameters with the true box parameters to calculate the loss function value;

[0068] Repeat the above steps until the loss function value is the same as the preset threshold and stop learning;

[0069] Gradient descent is used to optimize the YOLOv8 target recognition model 4, and the standard of the original printed manuscript image data obtained by the printed manuscript scanning module 2 and other qualified printing standards corresponding to the error range are determined in advance as the range of the prediction box, and all prediction box ranges are combined to form a bounding box to meet the image data recognition work of the printed objects on the printing line, thereby improving the monitoring and recognition efficiency and accuracy of the image data of the printed objects on the printing line.

[0070] In the target detection task, for the detection of irregular features in a single image, the number of anchor boxes used is usually multiple. This is because different targets may have different sizes, aspect ratios and rotation angles. In order to cover as many potential target shapes as possible, a set of anchor boxes with different sizes, aspect ratios and rotation angles are usually preset at different positions of the image. This solution combines the prediction boxes of multiple anchors that meet the preset threshold as a whole, and directly recognizes the image data of the printed object on the printing line to improve the detection efficiency. The bounding box formed by combining the prediction boxes of multiple anchors that meet the preset threshold improves the detection accuracy.

[0071] The above are preferred specific implementations of the present invention; however, the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and improved concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. The first printing plate monitoring system based on the YOLOv8 algorithm is characterized by: include: A visual inspection module (1) installed on a printing line, the visual inspection module (1) being used to obtain image data of a printed object on the printing line; A printing manuscript scanning module (2) is used to obtain image data of the original printing manuscript as first printing comparison data; A printing feature extraction module (3) is used to identify and extract feature point parameters of the printed object image data on the printing line and feature point parameters of the original printed manuscript image data, wherein the feature point parameters are shape parameters, which include center coordinates, length, width and tilt angle; A YOLOv8 target recognition model (4) is used to identify feature points of image data of a printed object on the printing line collected by a visual inspection module (1) on the printing line and feature points of image data of an original printed manuscript collected by a printing manuscript scanning module (2), and to clearly identify registration differences; An adaptive recognition optimization module (5) is used to optimize the anchor-based rotation box detection step in the YOLOv8 object recognition model (4); A model deep learning training database (6), wherein the model deep learning training database (6) is used to store image data of printed objects on a printing line and image data of original printed manuscripts; The optimization method of the adaptive recognition optimization module (5) is to replace the rotation box of the anchor used in the YOLOv8 target recognition model (4) with a directional bounding box composed of all feature point prediction boxes in the original printed manuscript image data by learning a bounding box. The specific steps are as follows: The original printed manuscript image data stored in the model deep learning training database (6) is used as a training set, and the adaptive recognition optimization module (5) uses the training set to train the YOLOv8 target recognition model (4), and extracts the feature point parameter information of the image data in the training set, that is, obtains the center coordinates, length, width and tilt angle as the real frame parameters; The obtained center coordinates, length, width and tilt angle parameters are imported into the anchor of the original YOLOv8 in the YOLOv8 target recognition model (4) to obtain the predicted box parameters, and then the predicted box parameters are compared with the real box parameters to calculate the loss function value. The original YOLOv8 of the YOLOv8 target recognition model (4) is guided to learn according to the loss function value, and then the combination of the predicted boxes of all feature points in the original printed manuscript image data is used as the bounding box to obtain the optimized YOLOv8; The data interaction module (7) is used to output the detected defect targets and adjust the printed content through the printing equipment.

2. The first printing plate alignment monitoring system based on the YOLOv8 algorithm according to claim 1, characterized in that: The YOLOv8 target recognition model (4) uses the improved YOLOv8 to input the image data of the printed object on the printing line for recognition and detection, and uses a bounding box to replace the rotation box combination of multiple anchors to directly perform feature recognition on the image data of the printed object on the printing line. When the image data of the printed object on the printing line is recognized by the bounding box, it is divided into the following two recognition situations: If the bounding box completely covers the recognition target, the image data recognition of the printed object can be completed directly, and the error is within the preset threshold, so it is passed; If the bounding box cannot completely cover the recognition target, it means that the image data of the printed object does not meet the recognition requirements, the error exceeds the preset threshold, and there is a defect.

3. The first printing plate alignment monitoring system based on the YOLOv8 algorithm according to claim 1, characterized in that: The original YOLOv8 of the YOLOv8 target recognition model (4) is guided to learn according to the loss function value as follows: The original YOLOv8 model executes the back-propagation algorithm to calculate the gradient of the loss function with respect to each parameter in the model, starting from the loss function and back-propagating the gradient along each layer of the original YOLOv8 network; After obtaining the gradient, the model uses the gradient descent algorithm to update the parameters. For each parameter, the update rule is as follows: parameter 新 =Parameter 旧 - learning rate × gradient; Among them, the learning rate is a pre-set hyperparameter, and the gradient is the derivative of the loss function with respect to the parameter, which indicates the steepest descent direction of the loss function near the parameter value; Then set the parameter 新 Replace the parameters in the original YOLOv8 旧 , the obtained center coordinates, length, width and tilt angle parameters are imported into the original YOLOv8 anchor in the YOLOv8 target recognition model (4) to obtain the predicted box parameters, and then the predicted box parameters are compared with the true box parameters to calculate the loss function value. The above steps are repeated until the loss function value is the same as the preset threshold and the learning is stopped.

4. The first printing plate alignment monitoring system based on the YOLOv8 algorithm according to claim 1, characterized in that: The visual inspection module (1) adopts a CCA chip linear array camera and a fill light source which is also installed on the printing line.

Images