Digital printing enhancement system

Abstract

The invention discloses a digital printing enhancement system. Printing images are captured in real time in the printing process. The preprocessed image is analyzed through the trained defect recognition deep learning model, and parameters of the printing machine are automatically adjusted so that deviation can be corrected in real time; by analyzing the cut image, the cutting effect is judged, and the parameters of the cutting machine are adjusted in real time, so that the cut paper is smooth and errorless. Multi-type and multi-roll production task management is carried out, so that the production efficiency is improved; the setting of the printing machine is automatically adjusted according to the production requirements of different styles. According to the method, real-time monitoring and adjustment can be achieved in the printing process, and high-level printing quality is guaranteed. Manual intervention is reduced, production efficiency is improved, and the risk of human errors is reduced. Automatic deviation correction and smooth slitting in the printing process are achieved, and the quality of printed matter is improved. And the accuracy and efficiency of deviation rectification and slitting are improved. The device is suitable for synchronous production of multiple types and multiple rolls, and the production efficiency is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of digital printing, and in particular to a digital printing enhancement system. Background Art

[0002] In the modern digital printing industry, as the market demand for personalized and efficient printed products continues to grow, the accuracy and reliability of printing technology have become particularly critical. Digital printing technology has become an important branch of the printing industry due to its flexibility and rapid response capabilities. However, there are some inherent challenges in the digital printing process, such as printing offset, uneven slitting, etc. These problems not only affect the aesthetics and quality of printed products, but may also lead to reduced production efficiency and increased costs.

[0003] Traditional correction methods mainly rely on the operator's experience and manual adjustment, which is not only time-consuming and labor-intensive, but also difficult to ensure consistency and accuracy. With the increase in printing speed and the expansion of production scale, traditional methods can no longer meet the needs of modern printing production. In addition, the problem of uneven slitting has always been a difficult problem in the printing industry, especially in the environment of simultaneous production of multiple styles and multiple rolls, the control of slitting quality becomes more complicated. Summary of the invention

[0004] The object of the present invention is to provide a digital printing enhancement system to solve the above-mentioned problems existing in the prior art.

[0005] An embodiment of the present invention provides a digital printing enhancement system, the system comprising:

[0006] Image capture module: uses a camera to capture the printed image in real time during the printing process;

[0007] Data preprocessing module: denoising, enhancing and normalizing the printed image to obtain a preprocessed image;

[0008] Deep learning analysis module: Use the trained defect recognition deep learning model to analyze the pre-processed image and identify the type of printing defects;

[0009] Among them, the defect recognition deep learning model includes: an input layer, a basic convolution layer, a dynamic attention module, a multi-scale feature extraction module, a convolution module and a fully connected layer; the basic convolution layer includes a first convolution layer with 64 3x3 convolution kernels and a second convolution layer with 128 3x3 convolution kernels; the first convolution layer is used to extract the first feature of the preprocessed image, and the second convolution layer is used to extract the first feature to obtain the second feature; the dynamic attention module is used to generate a first feature map based on the second feature, obtain the attention weight based on the first feature map, and obtain the third feature based on the attention weight and the first feature map; the multi-scale feature extraction module is used to perform convolution operations on the third feature through scale spaces of multiple different scales to obtain multiple second feature maps of the same scale; multiple second feature maps are spliced ​​to obtain a third feature map; the convolution module is used to extract the third feature map through multiple convolution blocks of different scales to obtain a fourth feature; the fully connected layer is used to flatten the fourth feature and obtain the printing defect type based on the fourth feature.

[0010] Optionally, the system further comprises an automatic real-time deviation correction module and a slitting and flattening guarantee module, wherein the automatic real-time deviation correction module obtains an adjustment strategy according to the deep learning analysis result, and automatically adjusts the parameters of the printing press according to the strategy to correct the deviation in real time;

[0011] The slitting smoothness guarantee module includes:

[0012] Slitting monitoring submodule: uses sensors and cameras to monitor the slitting position and paper edges;

[0013] Feedback adjustment submodule: By analyzing the image after slitting, the cutting effect is judged, and the parameters of the cutting machine are adjusted in real time to ensure that the paper is flat and correct after slitting.

[0014] Optionally, the system further includes a multi-model and multi-roll synchronous production module, and the multi-model and multi-roll synchronous production module includes:

[0015] Production task scheduling submodule: manage production tasks for multiple models and multiple volumes, optimize scheduling through algorithms, and improve production efficiency;

[0016] Dynamic adjustment submodule: Automatically adjust the settings of the printing press according to the production requirements of different styles.

[0017] Optionally, the system further includes a data monitoring and feedback module, which includes:

[0018] Real-time data recording submodule: real-time recording of various data in the production process, including printing speed, equipment status, and environmental conditions;

[0019] Abnormal alarm submodule: when abnormal conditions are detected, alarm information is issued and corresponding solutions are generated;

[0020] Visualization module: visualize real-time data and historical records.

[0021] Optionally, the training method of the defect recognition deep learning model includes:

[0022] A training set is obtained, where the training set includes a plurality of training printed images, and defect types are annotated in the training printed images;

[0023] The defect recognition deep learning model is trained using the training set, and the loss function of the defect recognition deep learning model is calculated for each iteration;

[0024] When the loss function converges, it is determined that the training of the defect recognition deep learning model is completed.

[0025] Optionally, the loss function of the defect recognition deep learning model is:

[0026] L total =L cls +λ1L bbox +λ2L feat +λ3L att

[0027] Among them, L total is the loss function of the deep learning model for defect recognition, L cls is the cross entropy loss function, L bbpx is the bounding box regression loss function, L feat is the feature reconstruction loss function, L att is the dynamic attention loss function, λ1, λ2 and λ3 are hyperparameters, and the values ​​of λ1, λ2 and λ3 are 1.0, 1.0 and 0.5 respectively.

[0028] Optionally, the cross entropy loss function is calculated as:

[0029]

[0030] Where N is the number of training print images in the training set, C is the number of sample categories, and y i,c is the true label of sample i in category c, The predicted probability of sample i in category c, for The logarithmic function of .

[0031] Optionally, the bounding box regression loss function is calculated as:

[0032]

[0033] Among them, t i,j is the true bounding box parameter of the printing defect; is the predicted bounding box parameter of printing defects; SmoothL1() is the smoothing loss function,

[0034]

[0035] express The absolute value of .

[0036] Optionally, the feature reconstruction loss function is calculated as:

[0037]

[0038] Among them, F i is the true feature of the training printed images in the training set, is the feature reconstructed by the network, express The norm of .

[0039] 10. The digital printing enhancement system according to claim 9, wherein the dynamic attention loss function is calculated as follows:

[0040]

[0041] in, is the attention weight, A target is the preset target attention weight, A i is the importance score; β is the control term, β = 0.1.

[0042] Compared with the prior art, the embodiments of the present invention achieve the following beneficial effects:

[0043] The embodiment of the present invention also provides a digital printing enhancement system, which monitors the printing quality in real time through the application of deep learning technology. If there is a quality problem, the parameters can be adjusted in time, which can significantly reduce the printing defect rate and improve the yield rate. The automated deviation correction and production scheduling methods make the production line run more efficiently, reduce the errors caused by human operation, and improve production efficiency. The optimized production process and efficient quality control help reduce the cost of raw materials and labor and achieve higher economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 The present invention provides a flowchart of a digital printing enhancement method. DETAILED DESCRIPTION

[0045] The present invention will be described in detail below in conjunction with the accompanying drawings.

[0046] Embodiments Embodiments of the present invention provide a digital printing enhancement system. When the system implements its functions, the system performs the following steps: Figure 1 The steps in the digital printing enhancement method shown in the figure include:

[0047] S101: Using a camera to capture a printing image in real time during the printing process.

[0048] The camera used in the embodiment of the present invention may be a high-resolution camera, which refers to a camera device capable of capturing high-quality images. Its image resolution is usually high, and common resolutions include 1080p (1920x1080), 4K (3840x2160), and even 8K (7680x4320) and above. A high-resolution camera is used to capture images during the printing process to capture information such as color, position, and speed.

[0049] S102: De-noising, enhancing and normalizing the printed image to obtain a pre-processed image. This step can improve the accuracy of subsequent analysis. The de-noising, enhancing and normalizing processes adopt conventional de-noising, enhancing and normalizing processes.

[0050] S103: Analyze the preprocessed image using the trained defect recognition deep learning model to identify the type of printing defects.

[0051] Optionally, step S103 may also identify printing defects, color deviations and position errors, and generate a real-time analysis report including information such as defect type, location and severity.

[0052] S104: Obtain an adjustment strategy based on the deep learning analysis results, and automatically adjust the parameters of the printing press according to the strategy to correct the deviation in real time.

[0053] According to the deep learning analysis results (printing defect type), the parameters of the printing machine (such as color, position, speed, etc.) are automatically adjusted to correct the deviation in real time to ensure the printing quality. In the embodiment of the present invention, a table of one-to-one correspondence between printing defect types and adjustment strategies is stored in the database in advance. When the printing defect type is identified, the adjustment strategy corresponding to the printing defect type is retrieved from the database. The adjustment strategy contains the parameters of the printing machine that need to be adjusted and the values ​​to be adjusted.

[0054] S105: Use sensors and cameras to monitor the slitting position and paper edges; judge the cutting effect by analyzing the image after slitting, and adjust the parameters of the cutting machine in real time to ensure that the paper is flat and correct after slitting.

[0055] S106: Manage production tasks for multiple styles and multiple volumes, optimize scheduling through algorithms, and improve production efficiency; automatically adjust the settings of the printing press according to the production requirements of different styles.

[0056] S107: Real-time recording of various data in the production process, including printing speed, equipment status, and environmental conditions; when abnormal conditions are detected, an alarm message is issued and a corresponding solution is generated; real-time data and historical records are visualized.

[0057] In the embodiment of the present invention, the parameters of the printing press that need to be adjusted include: color-related parameters, position-related parameters, speed-related parameters, pressure-related parameters, temperature-related parameters, material-related parameters and other parameters. Among them:

[0058] 1. Color related parameters

[0059] Ink Density: Adjust the density of ink to ensure color accuracy.

[0060] Color matching: Mix the proportions of different color inks as needed to achieve the desired color effect.

[0061] Drying time: Adjust the drying time of the ink to prevent blurring or sticking of the pattern after printing.

[0062] 2. Position related parameters

[0063] Alignment: Adjust the alignment of the printed pattern on the material to ensure accurate printing of the pattern.

[0064] Print Spacing: Adjust the spacing between different patterns to avoid overlap or gaps.

[0065] 3. Speed ​​related parameters

[0066] Printing speed: Adjust printing speed according to material and pattern complexity to ensure printing quality and production efficiency.

[0067] Drying speed: Adjust the speed of the drying system to ensure that the ink dries in the appropriate time.

[0068] 4. Pressure related parameters

[0069] Printing Pressure: Adjust the pressure of the print head or roller to ensure that the ink adheres evenly to the surface of the material.

[0070] Embossing Strength: Adjust the embossing strength according to the material type to avoid material damage or unclear patterns.

[0071] 5. Temperature related parameters

[0072] Ink temperature: Adjust the temperature of the ink to ensure its fluidity and adhesion.

[0073] Drying equipment temperature: Adjust the temperature of the drying equipment to speed up the ink drying process.

[0074] 6. Material related parameters

[0075] Material Type: Adjust printing settings based on the characteristics of different materials (such as paper, plastic, cloth, etc.).

[0076] Material Thickness: Adjust the height and pressure of the print head according to the thickness of the material.

[0077] 7. Other parameters

[0078] Cleaning cycle: Adjust the cleaning frequency of the printing press according to the type of ink and usage to avoid blockage.

[0079] Transport system speed: adjust the material transport speed to ensure synchronization with the printing speed.

[0080] By collecting real-time printing images to monitor and adjust these parameters in real time, the printing quality can be effectively improved, the scrap rate can be reduced, and production efficiency can be improved. Therefore, based on deep learning analysis, automatically adjusting these parameters is the key to achieving high-quality printing.

[0081] In an embodiment of the present invention, the digital printing enhancement system includes an image capture module, a data preprocessing module, and a deep learning analysis module.

[0082] The image capturing module is used to execute step S101, using a camera to capture a printing image in real time during the printing process.

[0083] The data preprocessing module is used to execute step S102 to perform denoising, enhancement and normalization processing on the printed image to obtain a preprocessed image.

[0084] The deep learning analysis module is used to execute step S103, using the trained defect recognition deep learning model to analyze the preprocessed image and identify the type of printing defects.

[0085] The digital printing enhancement system also includes an automatic real-time deviation correction module, which is used to execute step S104, obtain an adjustment strategy based on the deep learning analysis results, and automatically adjust the parameters of the printing press according to the strategy to correct the deviation in real time.

[0086] In order to monitor and correct deviations in real time during the printing process, ensure printing quality, and reduce scrap rate, the digital printing enhancement system provided by the embodiment of the present invention also includes a slitting and flattening guarantee module, which includes:

[0087] Slitting monitoring submodule: Use sensors and cameras to monitor the slitting position and paper edges to ensure the accuracy of the cutting line.

[0088] Feedback adjustment submodule: By analyzing the image after slitting, the cutting effect is judged, and the parameters of the cutting machine are adjusted in real time to ensure that the paper is flat and correct after slitting.

[0089] In response to the market demand for personalized and small-batch production, the system also includes a multi-model and multi-roll synchronous production module, which includes:

[0090] Production task scheduling submodule: manage production tasks for multiple models and multiple volumes, optimize scheduling through algorithms, and improve production efficiency;

[0091] Dynamic adjustment submodule: Automatically adjust the settings of the printing press according to the production requirements of different styles.

[0092] By adopting the above solution, the simultaneous production of multiple models and multiple rolls is supported, which improves the flexibility of the production line and adapts to the rapidly changing market needs.

[0093] In order to ensure the stability and reliability of the system, the system also includes a data monitoring and feedback module, which includes:

[0094] Real-time data recording submodule: real-time recording of various data in the production process, including printing speed, equipment status, and environmental conditions;

[0095] Abnormal alarm submodule: when abnormal conditions are detected, alarm information is issued and corresponding solutions are generated;

[0096] Visualization module: visualize real-time data and historical records.

[0097] Optional, multi-model and multi-roll synchronous production module also includes:

[0098] Task scheduling module: dynamically adjust the production sequence and schedule of printing different models according to production needs to ensure that multiple products can be efficiently completed on the same production line.

[0099] Resource allocation: Through real-time data analysis, optimize the allocation of production resources to ensure that different products do not interfere with each other during the production process and maintain an efficient production rhythm.

[0100] Data integration and synchronization: Integrate the production data of different products to ensure that the printing quality and progress of each roll can be consistent when printing multiple rolls.

[0101] By adopting the above scheme, the following beneficial effects are achieved:

[0102] Intelligence: The application of deep learning technology enables the system to learn and optimize autonomously to adapt to different printing environments and needs, reduce manual intervention, improve production efficiency, and reduce the risk of human error.

[0103] Real-time: The system can monitor and correct deviations in real time during the printing process, ensuring printing quality and reducing scrap rates. This method relies on the fast reasoning ability of deep learning models, which can monitor and adjust in real time during the printing process to ensure high levels of printing quality.

[0104] Flexibility: Supports simultaneous production of multiple models and multiple rolls, improves the flexibility of the production line, adapts to the rapidly changing needs of the market, and meets the market's demand for personalized and small-batch production.

[0105] Data-driven: Through data monitoring and feedback modules, comprehensive production data analysis is provided to improve the scientific nature of management decisions.

[0106] Cost-effectiveness: By reducing manual intervention and improving production efficiency, operating costs can be reduced and corporate competitiveness can be enhanced.

[0107] In the embodiment of the present invention, the defect recognition deep learning model plays a vital role in achieving the above-mentioned effect. On the basis of the defect recognition deep learning model improving the real-time, accuracy and reliability of defect recognition, it lays the foundation for the system to achieve the above-mentioned flexibility, reliability, intelligence and real-time performance, and also plays a key role.

[0108] In an embodiment of the present invention, a defect recognition deep learning model includes: an input layer, a basic convolutional layer, a dynamic attention module, a multi-scale feature extraction module, a convolution module and a fully connected layer; the basic convolutional layer includes a first convolutional layer with 64 3x3 convolutional kernels and a second convolutional layer with 128 3x3 convolutional kernels; the first convolutional layer is used to extract a first feature of a preprocessed image, and the second convolutional layer is used to perform feature extraction on the first feature to obtain a second feature; the dynamic attention module is used to generate a first feature map based on the second feature, obtain an attention weight based on the first feature map, and obtain a third feature based on the attention weight and the first feature map; the multi-scale feature extraction module is used to perform convolution operations on the third feature through scale spaces of multiple different scales to obtain multiple second feature maps of the same scale; multiple second feature maps are spliced ​​to obtain a third feature map; the convolution module is used to perform feature extraction on the third feature map through multiple convolution blocks of different scales to obtain a fourth feature; the fully connected layer is used to flatten the fourth feature, and obtain the type, location and severity of the printing defect based on the fourth feature.

[0109] Specific:

[0110] Input layer: Input image size: 224x224x3. Convert the size of the preprocessed image to 224x224x3.

[0111] Basic convolutional layer:

[0112] The first convolution layer: 64 3x3 convolution kernels, stride 1, ReLU activation, followed by a 2x2 maximum pooling layer; used to extract the first feature of the preprocessed image, that is, the first convolution layer convolves the preprocessed image with 64 3x3 convolution kernels with a stride of 1 to obtain the first convolution feature, and then performs a pooling operation on the first convolution feature through a 2x2 maximum pooling layer to obtain the first feature. During the convolution process, the activation function used is the ReLU function.

[0113] The second convolutional layer: 128 3x3 convolution kernels with a stride of 1, using ReLU activation, followed by a 2x2 maximum pooling layer; used to extract the first feature and obtain the second feature, that is, convolve the first feature with 128 3x3 convolution kernels with a stride of 1 to obtain the second convolution feature, and then pool the second convolution feature through a 2x2 maximum pooling layer to obtain the second feature.

[0114] Dynamic Attention Module:

[0115] Feature map generation: The output of convolution layer 2 (second feature) is passed into a 1x1 convolution layer to generate multiple feature maps (e.g., 128). In this step, the generated feature map is the first feature map.

[0116] Dynamic attention mechanism: Calculate the importance of each feature map (first feature map) and use global average pooling to generate a feature vector; Generate attention weights based on the feature vector through a fully connected layer (with ReLU activation function) and Sigmoid activation function; Multiply the attention weights by the feature map element by element to obtain important features (third features).

[0117] Multi-scale feature extraction module:

[0118] Scale 1: Perform 3x3 convolution on the output feature map (third feature) to generate 128 feature maps (second feature map);

[0119] Scale 2: Perform 5x5 convolution on the output feature map (third feature) to generate 128 feature maps (second feature map);

[0120] Scale 3: Perform 7x7 convolution on the output feature map (third feature) to generate 128 feature maps (second feature map);

[0121] The feature maps of the three scales (the second feature map) are concatenated to form 384 multi-scale feature maps (the third feature map);

[0122] Convolutional module:

[0123] Convolutional block 1: 384 3x3 convolution kernels, activation function is ReLU, followed by a 2x2 max pooling layer, output feature map;

[0124] Convolutional block 2: 256 3x3 convolution kernels, ReLU activation function, followed by a 2x2 max pooling layer, output feature map;

[0125] Convolutional block 3: 128 3x3 convolution kernels, ReLU activation function, followed by a 2x2 max pooling layer, output feature map;

[0126] That is, convolution block 1 performs a convolution operation on the third feature map through 384 3x3 convolutions to obtain a first convolution map. During the convolution operation, the activation function is the ReLU function. A 2x2 maximum pooling layer is used to perform a pooling operation on the first convolution map to output the first pooled feature map; convolution block 2 performs a convolution operation on the first pooled feature map through 256 3x3 convolution kernels. During the convolution operation, the activation function used is the ReLU function to obtain a second convolution map; then a 2x2 maximum pooling layer is used to perform a reddening operation on the second convolution map to obtain a second pooled feature map; finally, convolution block 3 performs a convolution operation on the second pooled feature map through 128 3x3 convolution kernels, and the activation function used is the ReLU function to obtain a third convolution feature map; then a 2x2 maximum pooling layer is used to perform a pooling operation on the third convolution feature map to obtain the fourth feature.

[0127] Fully connected layer:

[0128] Flatten layer: Flattens the output (fourth feature) of the convolution block. Specifically, the Flatten layer flattens the input four-dimensional tensor into a two-dimensional tensor, that is, performs a dimensionality reduction operation. It is usually converted into a tensor of shape (M, H xW x C). In this way, the feature map of each sample is flattened into a one-dimensional vector for input to the fully connected layer. Among them, M is the batch size (batchsize), which indicates the number of fourth features. H represents the height of the fourth feature (height). W represents the width of the fourth feature (width). C represents the number of channels of the fourth feature (channels, usually the number of convolution kernels).

[0129] Example: Assume the output shape of the convolutional block is (32, 7, 7, 128), where:

[0130] 32: batch size.

[0131] 7: Feature map height (height).

[0132] 7: Feature map width (width).

[0133] 128: Number of channels.

[0134] After passing through the Flatten layer, the output will become:

[0135] The flattened shape is: (32,7x 7x 128) = (32,6272).

[0136] That is, the input of the Flatten layer is the output of the convolution block, that is, a multi-dimensional feature map with a shape of (M,H,W,C)$.

[0137] Flattened output: converted to a two-dimensional tensor of shape (M, H x W x C) for subsequent fully connected layer processing.

[0138] Fully connected layer 1: 512 neurons, ReLU activation;

[0139] Dropout layer: Use a ratio of 0.5 to prevent overfitting;

[0140] Fully connected layer 2: Softmax layer of the output category number. That is, the type, location and severity of printing defects are calculated through the Softmax function.

[0141] By combining the dynamic attention mechanism and multi-scale feature extraction, the model's ability to focus on important features is enhanced, while providing rich context to improve the accuracy of printing defect recognition and classification.

[0142] In the embodiment of the present invention, the printing defect types include:

[0143] Color-related defects:

[0144] Color deviation: The color of the printed product is inconsistent with the design or color sample. It is indicated by mark 1.

[0145] Color unevenness: Color density is uneven in some areas, with spots or color bands. This is indicated by the label 2.

[0146] Position-related defects:

[0147] Misalignment: The printed pattern is not aligned correctly, causing the pattern to shift. This is indicated by callout 3.

[0148] Overlap and gaps: There are overlaps or unnecessary gaps between different patterns. This is indicated by a 4.

[0149] Printing quality defects:

[0150] Blurred: The printed pattern is unclear and the edges are blurred. This is indicated by a 5.

[0151] Missing printing: Some areas are missing ink, resulting in missing patterns. Indicated by callout 6.

[0152] Stains: Stains or foreign matter appear on the printed surface, affecting the appearance. This is indicated by a mark 7.

[0153] Surface treatment defects:

[0154] Bubbling: Bubbles appear on the surface of the ink layer. Indicated by mark 8.

[0155] Shedding: The ink is not firmly attached and is easy to fall off or peel off. It is indicated by the mark 9.

[0156] Texture and pattern defects:

[0157] Uneven texture: The texture of the printed material surface does not match the design. Indicated by mark 10.

[0158] Pattern distortion: Pattern deformation due to improper printing pressure or speed. Indicated by mark 11.

[0159] Mechanical related defects:

[0160] Uneven embossing: Uneven pressure during printing leads to different printing effects. This is indicated by 12.

[0161] Machine Failure: Printing problems caused by equipment failure, such as cylinder wear or transport system problems. Indicated by callout 13.

[0162] Drying and curing defects:

[0163] Poor drying: The ink fails to dry in time, causing the pattern to be blurred or sticky. This is indicated by mark 14.

[0164] Incomplete curing: Problems during UV printing or thermal curing, resulting in the ink not being completely cured. Indicated by mark 15.

[0165] Material related defects:

[0166] Material defects: The printing material itself has defects, such as wrinkles, bubbles, etc. It is indicated by mark 16.

[0167] Material mismatch: The ink is incompatible with the material, resulting in poor adhesion. Indicated by callout 17.

[0168] Design-related flaws:

[0169] Design error: There are problems with the design itself, such as insufficient resolution or incorrect color settings. Indicated by callout 18.

[0170] File format problem: The file format is incorrect, resulting in poor printing output. This is indicated by annotation 19.

[0171] The above defect types may affect the printing quality and the appearance of the final product. By monitoring and analyzing these defects, combined with deep learning and automatic adjustment technology, the printing quality can be effectively improved and the defect rate can be reduced.

[0172] In order to improve the recognition accuracy of the defect recognition deep learning model, it is necessary to train the defect recognition deep learning model. The training methods of the defect recognition deep learning model include:

[0173] A training set is obtained, the training set includes multiple training printed images, and the defect types are marked in the training printed images. The defect types are marked in the training printed images, for example, color deviation is marked 1, and color unevenness is marked 2. The marking can be specifically performed by referring to the above representation method.

[0174] The defect recognition deep learning model is trained using the training set, and the loss function of the defect recognition deep learning model is calculated in each iteration.

[0175] When the loss function converges, it is determined that the training of the defect recognition deep learning model is completed.

[0176] Among them, when training the defect recognition deep learning model, it is necessary to continuously calculate the loss function to determine whether the recognition accuracy of the defect recognition deep learning model training meets the standard. When the loss function converges, it is determined that the recognition accuracy of the defect recognition deep learning model training meets the standard, and the training of the defect recognition deep learning model is stopped.

[0177] In the embodiment of the present invention, a new loss function design is provided by combining classification, bounding box regression, feature reconstruction and dynamic attention mechanism. This design not only improves the multi-task learning ability of the model, but also enhances the richness of feature expression, making the defect recognition deep learning model perform better on complex data sets. In this way, the model can better capture key features and improve the accuracy and robustness of defect recognition.

[0178] The loss function of the defect recognition deep learning model is:

[0179] L total =L cls +λ1L bbox +λ2L feat +λ3L att

[0180] Among them, L total is the loss function of the deep learning model for defect recognition, L cls is the cross entropy loss function, L bboxis the bounding box regression loss function, L feat is the feature reconstruction loss function, L att is the dynamic attention loss function, λ1, λ2 and λ3 are hyperparameters, and the values ​​of λ1, λ2 and λ3 are 1.0, 1.0 and 0.5 respectively.

[0181] The cross entropy loss function is calculated as:

[0182]

[0183] Where N is the number of training print images in the training set, C is the number of sample categories, and y i,c is the true label of sample i in category c, The predicted probability of sample i in category c, for The logarithmic function of y. i represents the sample number, c represents the label number, and the value of c is 1, 2, 3, ..., K, where K represents the number of defect types. i,c It is pre-labeled in the training samples (training printed images). The probability of the printing defect type is calculated by the fully connected layer 2 in the above-mentioned defect recognition deep learning model through the Softmax function. The probability also represents the defect type.

[0184] The bounding box regression loss function is calculated as:

[0185]

[0186] Among them, t i,j is the true bounding box parameter of the printing defect; is the predicted bounding box parameter of the printing defect. In the embodiment of the present invention, for the training sample, the defect border parameters (upper left corner and lower right corner coordinates) are also pre-marked, that is, the real bounding box parameters. At the same time, after the sample is detected to have defects, the bounding box parameters of the defect are identified by the convolutional Bible network or the Hough detection algorithm, that is, the predicted bounding box parameters.

[0187] SmoothL1() is a smooth loss function:

[0188]

[0189] express In this embodiment of the present invention, It also represents the Euclidean distance from the true bounding box parameters to the predicted bounding box parameters.

[0190] The feature reconstruction loss function is calculated as:

[0191]

[0192] Among them, F i is the true feature of the training printed images in the training set, is the feature reconstructed by the network, express The norm of . Indicates F i minus In the embodiment of the present invention, F i is the second feature output by the base convolutional layer, is the fourth feature output by the convolution block. In the embodiment of the present invention, when the sizes of the second feature and the fourth feature do not meet the requirement, F i minus When performing the operation, the second feature and the fourth feature need to be aligned in size. Specifically, the feature with a smaller size is dimensionally expanded, and the value of the expanded dimension is 0, so that the dimensions of the second feature and the fourth feature are the same.

[0193] The dynamic attention loss function is calculated as:

[0194]

[0195] in, is the attention weight, A tar get is the preset target attention weight, A i is the importance score; β is the control term, β = 0.1.

[0196] By adopting such a loss function to train the defect recognition deep learning model, the model can better capture key features and improve the accuracy and robustness of the model in identifying printing defects. On the basis of the defect recognition deep learning model improving the real-time, accuracy and reliability of defect recognition, it lays the foundation for the system to achieve the above-mentioned flexibility, reliability, intelligence and real-time performance, and also plays a key role.

[0197] The digital printing enhancement method provided by the embodiment of the present invention achieves the following objectives:

[0198] Improved printing quality: Through real-time monitoring and correction, the printing defect rate is reduced by more than 30% and the yield rate is significantly improved.

[0199] Improved production efficiency: Increased production efficiency by 20% to 50%, depending on product type and ### Expected effect (continued)

[0200] Improved production efficiency: Through optimized production scheduling and automated deviation correction, production efficiency is increased by 20% to 50%, depending on the product type and production environment. Downtime caused by equipment failure or human error is significantly reduced, ensuring the continuity and stability of the production line.

[0201] Cost reduction: Overall production costs are expected to be reduced by 15% to 30% due to reduced scrap and manual intervention. Operating costs will be further reduced and profit margins will be improved by improving raw material utilization.

[0202] Improved customer satisfaction: Provide higher quality printing products to meet customers' needs for precision and beauty, enhance customer trust and loyalty. Improve customer satisfaction by quickly responding to market needs and shortening delivery time.

[0203] Application examples:

[0204] In order to verify the effectiveness of the embodiments of the present invention, the following are several application cases:

[0205] Packaging and printing enterprise case:

[0206] Through real-time monitoring and automatic correction, the printing defect rate was reduced from 5% to 1%. Production efficiency was increased by 40%, while manual intervention on the production line was reduced.

[0207] Commercial printing company case:

[0208] The new production process has shortened the delivery time of customer orders by 30%, and the quality issues reported by customers have been reduced by 50% through real-time monitoring of the system.

[0209] Publishing and printing industry case:

[0210] The printing quality has been greatly improved, and the clarity of the text and images of the books has been significantly improved. The reader satisfaction survey shows that customer satisfaction has increased by more than 20%.

[0211] In summary, the digital printing enhancement system provided by the embodiment of the present invention not only improves the degree of automation of printing, but also ensures the stability of printing quality and production efficiency through real-time analysis and feedback mechanism, and is suitable for the intelligent development needs of modern printing enterprises. The application of this technology will bring revolutionary impact to the printing industry and promote the industry to a higher level.

[0212] The algorithm and display provided herein are not inherently related to any particular computer, virtual system or other device. Various general purpose systems can also be used together with the teachings based on this. According to the above description, it is obvious that the structure required for constructing such systems. In addition, the present invention is not directed to any specific programming language either. It should be understood that various programming languages ​​can be utilized to realize the content of the present invention described herein, and the description of the above specific languages ​​is for disclosing the best mode of the present invention.

[0213] In the description provided herein, a large number of specific details are described. However, it is understood that embodiments of the present invention can be practiced without these specific details. In some instances, well-known methods, structures and techniques are not shown in detail so as not to obscure the understanding of this description.

[0214] The various component embodiments of the present invention may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. It should be understood by those skilled in the art that a microprocessor or digital signal processor (DSP) may be used in practice to implement some or all of the functions of some or all of the components in the apparatus according to an embodiment of the present invention. The present invention may also be implemented as a device or device program (e.g., a computer program and a computer program product) for executing part or all of the methods described herein. Such a program implementing the present invention may be stored on a computer-readable medium, or may have the form of one or more signals. Such a signal may be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

Claims

1. A digital printing enhancement system, characterized in that: The system comprises: Image capture module: uses a camera to capture the printed image in real time during the printing process; Data preprocessing module: denoising, enhancing and normalizing the printed image to obtain a preprocessed image; Deep learning analysis module: Use the trained defect recognition deep learning model to analyze the pre-processed image and identify the type of printing defects; Among them, the defect recognition deep learning model includes: an input layer, a basic convolution layer, a dynamic attention module, a multi-scale feature extraction module, a convolution module and a fully connected layer; the basic convolution layer includes a first convolution layer with 64 3x3 convolution kernels and a second convolution layer with 128 3x3 convolution kernels; the first convolution layer is used to extract the first feature of the preprocessed image, and the second convolution layer is used to extract the first feature to obtain the second feature; the dynamic attention module is used to generate a first feature map based on the second feature, obtain the attention weight based on the first feature map, and obtain the third feature based on the attention weight and the first feature map; the multi-scale feature extraction module is used to perform convolution operations on the third feature through scale spaces of multiple different scales to obtain multiple second feature maps of the same scale; multiple second feature maps are spliced ​​to obtain a third feature map; the convolution module is used to extract the third feature map through multiple convolution blocks of different scales to obtain a fourth feature; the fully connected layer is used to flatten the fourth feature and obtain the printing defect type based on the fourth feature.

2. The digital printing enhancement system according to claim 1, characterized in that: The system also includes an automatic real-time deviation correction module and a slitting and flattening guarantee module. The automatic real-time deviation correction module obtains an adjustment strategy based on the deep learning analysis results, and automatically adjusts the parameters of the printing press according to the strategy to correct the deviation in real time; The slitting smoothness guarantee module includes: Slitting monitoring submodule: uses sensors and cameras to monitor the slitting position and paper edges; Feedback adjustment submodule: By analyzing the image after slitting, the cutting effect is judged, and the parameters of the cutting machine are adjusted in real time to ensure that the paper is flat and correct after slitting.

3. The digital printing enhancement system according to claim 1, characterized in that: The system also includes a multi-model and multi-roll synchronous production module, which includes: Production task scheduling submodule: manage production tasks for multiple models and multiple volumes, optimize scheduling through algorithms, and improve production efficiency; Dynamic adjustment submodule: Automatically adjust the settings of the printing press according to the production requirements of different styles.

4. The digital printing enhancement system according to claim 1, characterized in that: The system also includes a data monitoring and feedback module, which includes: Real-time data recording submodule: real-time recording of various data in the production process, including printing speed, equipment status, and environmental conditions; Abnormal alarm submodule: when abnormal conditions are detected, alarm information is issued and corresponding solutions are generated; Visualization module: visualize real-time data and historical records.

5. The digital printing enhancement system according to claim 1, characterized in that: The training methods for the defect recognition deep learning model include: A training set is obtained, where the training set includes a plurality of training printed images, and defect types are annotated in the training printed images; The defect recognition deep learning model is trained using the training set, and the loss function of the defect recognition deep learning model is calculated for each iteration; When the loss function converges, it is determined that the training of the defect recognition deep learning model is completed.

6. The digital printing enhancement system according to claim 5, characterized in that: The loss function of the defect recognition deep learning model is: L total =L cls +λ1L bbox +λ2L feat +λ3L att Among them, L total is the loss function of the deep learning model for defect recognition, L cls is the cross entropy loss function, L bbox is the bounding box regression loss function, L feat is the feature reconstruction loss function, L att is the dynamic attention loss function, λ1, λ2 and λ3 are hyperparameters, and the values ​​of λ1, λ2 and λ3 are 1.0, 1.0 and 0.5 respectively.

7. The digital printing enhancement system according to claim 6, characterized in that: The cross entropy loss function is calculated as: Where N is the number of training print images in the training set, C is the number of sample categories, and y i,c is the true label of sample i in category c, The predicted probability of sample i in category c, for The logarithmic function of .

8. The digital printing enhancement system according to claim 7, characterized in that: The bounding box regression loss function is calculated as: Among them, t i,j is the true bounding box parameter of the printing defect; is the predicted bounding box parameter of printing defects; SmoothL1() is the smoothing loss function, express The absolute value of .

9. The digital printing enhancement system according to claim 8, characterized in that: The feature reconstruction loss function is calculated as: Among them, F i is the true feature of the training printed images in the training set, is the feature reconstructed by the network, express The norm of .

10. The digital printing enhancement system according to claim 9, characterized in that: The dynamic attention loss function is calculated as: in, is the attention weight, A target is the preset target attention weight, A i is the importance score; β is the control term, β = 0.1.

Images