Deep learning self-calibration system for high-precision color management

Abstract

The invention provides a deep learning self-calibration system for high-precision color management, and relates to the technical field of color calibration, a system architecture comprises an algorithm layer, a system architecture layer and a data processing and optimizing layer, through fusion of multiple advanced algorithms in the algorithm layer, the precision of color correction is effectively improved, and the accuracy of color correction is improved. The high consistency of colors of printed matters in the whole production process is ensured, the quality stability of printed products is greatly improved, data processing is carried out through all modules of the data processing and optimizing layer, real-time data in the printing production process are efficiently processed, a model is dynamically updated, and the printing production efficiency is improved. The design of the system architecture layer provides a comprehensive adaptation scheme for large-scale production, multiple links and multiple equipment materials in the printing industry, the universality and flexibility of the system are enhanced, and the system is suitable for large-scale production, multiple links and multiple equipment materials in the printing industry. The time and labor cost for readjusting the system due to the change of equipment materials is reduced, and the production strain capacity of enterprises is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of color calibration, and in particular to a deep learning self-calibration system for high-precision color management. Background Art

[0002] According to a method for improving the printing accuracy of a color printing system disclosed in China with a publication number of "CN112918116B", the method comprises the following steps: 1. printing an overprint optical identifier; 2. digitally collecting the overprint optical identifier; 3. performing CMYK color channel segmentation on the collected overprint optical identifier; 4. performing edge feature extraction on images of each CMYK channel; 5. geometrically locating the edge features of the images of each CMYK channel, and calculating the nozzle calibration value. The overprint optical identifier positioning algorithm of the present invention has high accuracy and can effectively improve the quality of overprinted printed products; the overprint optical identifier is small in size and can avoid waste of substrates; the present invention has a high degree of automation and reduces labor costs and equipment debugging time.

[0003] According to a printing color detection and control system applied to a printing production line disclosed in China with the publication number "CN118306108B", the present invention obtains the color difference estimation of the current printed product by performing color detection and analysis after the current printed product is printed, and further controls and adjusts the parameters of the printing equipment when triggering the retrieval signaling based on the comparison result of the color difference estimation of the current printed product, and further marks the time point of triggering the retrieval signaling as the abnormality start time, and the time point of adjusting the relevant working parameters of the printing equipment as the calibration completion time, extracts all the printed products within the time range, analyzes the color difference estimation of each printed product within the time range, and obtains the change trend index of the printing equipment within the time range. If the change trend index is large, it means that the printing equipment needs to be inspected and maintained, and the existing hidden dangers and problems are dealt with in time, thereby improving the overall production quality and intelligence level of the printing production line.

[0004] The above patent documents and prior art have the following technical problems when used:

[0005] Problem 1: Traditional printing color management technology often relies on simple color matching algorithms or manual adjustments based on experience, which makes it difficult to conduct a comprehensive and in-depth analysis of the complex color characteristics of printed images. When faced with differences in raw materials for different batches of printing, such as paper whiteness, subtle changes in ink hue, fluctuations in printing equipment characteristics, and changes in printing processes, such as when adjusting the order of multi-color printing overprinting, it is impossible to accurately predict and correct color deviations, which leads to poor color consistency of printed products. The same design has obvious color differences in different batches of printing, which is difficult to meet high-quality printing needs, such as high-end commercial printing and art printing. The scrap rate is high, which increases production costs and wastes resources;

[0006] Problem 2: The printing industry has the characteristics of large-scale production, multiple links and multiple equipment and materials involved. The existing technology lacks an effective system architecture to integrate these complex factors. On the one hand, it is inefficient in large-scale data processing. The traditional centralized computing architecture is difficult to cope with the rapid processing and model training of massive printing color data, resulting in a longer production cycle. On the other hand, each production link is relatively isolated, and there is a lack of effective information feedback and coordination mechanism before, during and after printing. The transmission and accumulation of color deviations between different links cannot be discovered and corrected in time. At the same time, when printing equipment or materials are replaced, it takes a lot of time and manpower to readjust the color management parameters. There is a lack of adaptive adjustment capabilities, which seriously affects production efficiency and corporate competitiveness.

[0007] Problem three: The existing data processing methods for printing color management are relatively simple. The amount of training data obtained is limited and lacks diversity, making it difficult to cover various color deviations that may occur in actual printing, resulting in weak generalization capabilities of deep learning models. Traditional model training methods are mostly static training, which cannot adapt to data distribution changes and color fluctuations in the printing production process in real time. Once production conditions change, model performance drops sharply and requires frequent retraining, which not only interrupts production, but also increases model training costs and time costs, which is not conducive to the intelligent and efficient management of printing production. Summary of the invention

[0008] Technical issues solved

[0009] In view of the shortcomings of the prior art, the present invention provides a deep learning self-calibration system for high-precision color management, which solves the following problems:

[0010] 1. The problem of insufficient printing color correction accuracy and consistency leading to high scrap rate and increased consumables costs;

[0011] 2. It is difficult to adapt to the complex and changing production environment and processes of the printing industry, resulting in low production efficiency;

[0012] 3. To address the inconvenience in production management caused by limited image data processing capabilities and poor model adaptability during printing.

[0013] Technical Solution

[0014] To achieve the above objectives, the present invention is implemented through the following technical solutions: a deep learning self-calibration system for high-precision color management, the system architecture includes an algorithm layer, a system architecture layer and a data processing and optimization layer, wherein:

[0015] The algorithm layer is used to perform deep feature extraction, deviation prediction and correction strategy generation on the color of the printed image. The algorithm layer includes a multi-algorithm fusion model, which includes a convolutional neural network CNN and a recurrent neural network RNN ​​combination unit, a fusion generative adversarial network GAN and an autoencoder combination unit, and a deep belief network DBN and a restricted Boltzmann machine RBM combination unit;

[0016] The system architecture layer is used to build an architecture system that adapts to large-scale production, multiple links and multiple equipment and materials in the printing industry, including a distributed deep learning architecture, a multi-level feedback self-calibration architecture and an adaptive color calibration architecture. The distributed deep learning architecture distributes the training tasks of the deep learning model to multiple computing nodes for parallel processing. Different nodes focus on data processing of different printing product lines or different color spaces. The central coordinator integrates the results of each node. After the multi-level feedback self-calibration architecture performs preliminary correction on the image color in the pre-press stage, the correction result is input into the system again in the printing stage for secondary detection and analysis. The deep learning model parameters or correction strategies are adjusted according to feedback information at different levels, and a feedback channel is established from the post-press quality inspection stage to the pre-press design and printing stages. The adaptive color calibration architecture uses a pre-established printing equipment and material feature library to quickly identify and match the corresponding model configuration during printing production and automatically adjust the structure and parameters of the deep learning model.

[0017] The data processing and optimization layer is used to preprocess, expand and dynamically update the printed color data. The data processing and optimization layer includes a data enhancement and synthesis module, a semantic segmentation data processing module and a real-time data streaming processing and incremental learning module. The data enhancement and synthesis module expands the training data set by performing transformation operations such as rotation, scaling, contrast adjustment, etc. on the existing printed color images and artificially synthesizing image data with specific color deviation characteristics. The semantic segmentation data processing module uses semantic segmentation technology to divide different areas in the printed image according to color categories, object categories or functional areas, and performs color deviation analysis and correction respectively; the real-time data streaming processing and incremental learning module quickly processes and extracts features of the color data generated in real time during the printing production process, and uses incremental learning technology to adapt the model to the new printing data distribution and color change trends.

[0018] Preferably, the convolutional neural network CNN and recurrent neural network RNN ​​combination unit is used to deeply extract the color pixel distribution, pattern texture and other features of the printed image, and analyze the change law of color deviation in the entire printed page and between different printed batches based on the spatial and temporal dependencies between the printed image sequences. In the fusion generative adversarial network GAN and autoencoder combination unit, the GAN generator generates printed image simulation data similar to the target printing color standard according to the target printing color standard, and the discriminator distinguishes between the real printed image and the generated image. The adversarial training of the two improves the quality of the generated image. The autoencoder compresses and reconstructs the input printed color image and learns the potential color representation of the image. The two work together to correct the printed color deviation and ensure color consistency. The deep belief network DBN and restricted Boltzmann machine RBM combination unit performs unsupervised pre-training on the printed color data through RBM, learns the basic feature distribution and structure of the printed color data, and then stacks it to construct a DBN, thereby mining the high-level abstract features hidden in the complex printed color data, providing an accurate basis for subsequent automatic correction.

[0019] Preferably, each computing node in the distributed deep learning architecture has an independent data storage unit and a computing unit. The data storage unit is used to temporarily store the assigned specific printing color data subset. The computing unit performs local training operations of the deep learning model based on the stored data. The computing nodes are connected by high-speed data transmission channels to ensure the efficiency of data interaction and model integration.

[0020] Preferably, the feedback information in the multi-level feedback self-calibration architecture includes color deviation values, deviation distribution areas, color space coordinate differences, and corresponding printing link identification information, wherein the color deviation values ​​are accurate to three decimal places, the deviation distribution areas are represented by image pixel coordinate ranges or printed product physical size areas, the color space coordinate differences are quantitatively described based on color spaces commonly used in the printing industry such as CMYK or LAB, and the printing link identification information clearly distinguishes the sources of pre-press, mid-press, and post-press links.

[0021] Preferably, the data stored in the printing equipment and material feature library in the adaptive color calibration architecture covers printing press brand and model, printing press mechanical parameters such as pressure, speed, dot gain curve, paper types such as coated paper, offset paper, cardboard, including their grammage, roughness, whiteness, etc., ink brand and model and characteristics such as hue, saturation, drying speed, viscosity and other information, and the data is stored in the form of a structured database to facilitate rapid retrieval and model configuration matching.

[0022] Preferably, the color data synthesis algorithm in the data enhancement and synthesis module is constructed based on the printing color physical model and the mathematical statistical model. When synthesizing image data with specific color deviation characteristics, it can accurately generate a simulated image according to the deviation parameters set by the user, such as the deviation degree, deviation direction, and deviation distribution pattern. The generated image data format is consistent with the real printed image data format and can be directly used for deep learning model training.

[0023] Preferably, the semantic segmentation technology in the semantic segmentation data processing module adopts a deep learning semantic segmentation model. The model structure is improved based on the U-Net architecture, and a multi-scale feature fusion module is added. It can accurately identify text, image body, background and other areas in printed images at different resolutions. The segmentation accuracy reaches the pixel level, and it has good adaptability to different types of printed images, such as posters, book pages, packaging boxes, etc.

[0024] Preferably, the data streaming processing algorithm in the real-time data streaming processing and incremental learning module adopts sliding window technology and distributed caching mechanism. The sliding window size is dynamically adjusted according to the printing data flow and processing speed to ensure that data is not lost and can be processed in time during data flow peaks. The distributed caching mechanism caches recently processed data in multiple nodes to facilitate the incremental learning module to quickly obtain data for model updates. The validity period of the cached data is dynamically set according to the continuity of the printing task and the timeliness of the data.

[0025] Preferably, The DBN model training in the deep belief network DBN and restricted Boltzmann machine RBM combination unit The training adopts a combination of layer-by-layer pre-training and overall fine-tuning. In the layer-by-layer pre-training stage, the complex The number of training iterations for each layer is set by complexity, and the number of iterations is gradually reduced from the bottom layer to the top layer. The overall fine-tuning stage adopts The gradient descent optimization algorithm adjusts the learning rate and momentum parameters, and takes the accuracy of printing color deviation correction as the optimization target. Standard function.

[0026] Preferably, The autoencoder in the fusion generation adversarial network GAN and autoencoder combination unit adopts symmetric The encoder-decoder structure consists of multiple convolutional layers and pooling layers, which is used to compress printed color images into low-resolution images. The decoder consists of multiple layers of deconvolution layers and upsampling layers to reconstruct the low-dimensional feature vector into the original image. The image is of the same size as the original one and a skip connection is set between the encoder and the decoder to retain the original feature information of the image and improve the weight The quality of the image.

[0027] Preferably, The RNN model in the convolutional neural network CNN and recurrent neural network RNN ​​combination unit adopts a long The short-term memory network LSTM structure contains a forget gate, an input gate, a cell state update gate, and an output gate. The gating mechanism controls the flow and update of information, which can effectively handle the long-distance dependencies in the printed image sequence. The number of layers of the LSTM network is dynamically adjusted according to the complexity of the printed image and the accuracy requirements of the color deviation analysis. The initial weights are set using the Xavier initialization method .

[0028] Beneficial Effects

[0029] The present invention provides a deep learning self-calibration system for high-precision color management. It has the following beneficial effects:

[0030] 1. The system of the present invention effectively ensures high-precision color correction and consistency. Through the integration of multiple advanced algorithms in the algorithm layer, such as the combination of convolutional neural network CNN and recurrent neural network RNN, the color pixel distribution, texture characteristics and color deviation rules between different printing batches of printed images can be accurately analyzed. The fusion of generative adversarial network GAN and autoencoder combination can generate high-quality simulated images and coordinately correct color deviations. The deep belief network DBN and restricted Boltzmann machine RBM combination mines deep-level features to provide accurate correction basis. These algorithms work together to effectively improve the accuracy of color correction and ensure the high consistency of color of printed products throughout the production process. Whether it is the same batch or different batches of printed products, they can achieve extremely close color standards, greatly improving the quality stability of printed products, meeting the needs of high-end printing businesses with strict color requirements, such as the production of printed products such as art albums and brand promotion posters, reducing the scrap rate caused by color deviation, reducing production costs and enhancing corporate reputation.

[0031] 2. The system of the present invention can efficiently adapt to the complex production environment and processes of the printing industry, and uses the design of the system architecture layer to provide a comprehensive adaptation solution for large-scale production, multiple links and multiple equipment and materials in the printing industry. The distributed deep learning architecture uses multiple computing nodes to process massive printing color data in parallel, accelerates the model training process, and improves production efficiency; the multi-level feedback self-calibration architecture runs through the pre-press, mid-press, and post-press links to form a closed-loop feedback control system, which can timely detect and correct color deviations and optimize the entire printing process. The adaptive color calibration architecture is based on a rich printing equipment and material feature library, and can quickly adjust the model configuration according to different printing machines, paper, ink and other conditions, so that the system can easily cope with the complex situation of frequent equipment and material changes and different production tasks in the printing industry, enhance the versatility and flexibility of the system, reduce the time and labor costs required to readjust the system due to changes in equipment and materials, and improve the company's production resilience.

[0032] 3. The system of the present invention optimizes data processing and model performance improvement, and uses various modules of the data processing and optimization layer to process data. The data enhancement and synthesis module expands the training data set based on the physical and mathematical models of printing color. The rich data diversity enables the deep learning model to better learn various color deviations and improve the generalization ability of the model, so that it can accurately predict and correct deviations when facing complex and changeable color data in actual printing. The semantic segmentation data processing module uses an improved deep learning semantic segmentation model to perform refined regional division and color analysis on the printed image, and corrects according to the characteristics of different regions to further improve the accuracy and visual effect of color management. The real-time data streaming processing and incremental learning module uses sliding window technology and distributed caching mechanism to achieve efficient processing of real-time data in the printing production process and dynamic model update, ensuring that the model can adapt to the distribution of printing data and color change trends in a timely manner, always maintain good correction performance, and do not need to frequently interrupt production for large-scale model retraining, thereby improving production continuity and equipment utilization, and helping enterprises to achieve intelligent and efficient printing production management. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a system architecture diagram of the present invention;

[0034] Figure 2 It is a flow chart of the system operation steps of the present invention;

[0035] Figure 3 It is the system data dynamic flow diagram of the present invention. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Specific embodiment one:

[0038] like Figure 1-3 As shown in the figure, the deep learning self-calibration system for high-precision color management has a system architecture including an algorithm layer, a system architecture layer, and a data processing and optimization layer, where:

[0039] The algorithm layer is used for deep feature extraction, deviation prediction and correction strategy generation of printed image colors. The algorithm layer includes a multi-algorithm fusion model, which includes a convolutional neural network CNN and a recurrent neural network RNN ​​combination unit, a fusion generative adversarial network GAN and an autoencoder combination unit, a deep belief network DBN and a restricted Boltzmann machine RBM combination unit. The convolutional neural network CNN and the recurrent neural network RNN ​​combination unit are used to deeply extract the color pixel distribution, pattern texture and other features of the printed image, and analyze the change law of color deviation in the entire printed page and between different printing batches based on the spatial and temporal dependency between the printed image sequences. It can accurately analyze the color pixel distribution, texture features and color deviation laws of the printed image between different printing batches. In the fusion generative adversarial network GAN and the autoencoder combination unit, the GAN generator generates similar printed image simulation data according to the target printing color standard, and the discriminator distinguishes between the real printed image and the generated image. The adversarial training of the two improves the quality of the generated image. The autoencoder performs adversarial training on the input printed image. The color image is compressed and reconstructed, and the potential color representation of the image is learned. The two work together to correct the printing color deviation and ensure color consistency. They can generate high-quality simulated images and coordinately correct color deviations. The deep belief network DBN and the restricted Boltzmann machine RBM combination unit performs unsupervised pre-training of the printing color data through RBM, learns the basic feature distribution and structure of the printing color data, and then stacks it to construct a DBN, which mines the high-level abstract features hidden in the complex printing color data, provides an accurate basis for subsequent automatic correction, and mines deep features to provide an accurate correction basis. These algorithms work together to effectively improve the accuracy of color correction and ensure the high consistency of color of printed products throughout the production process. Whether it is the same batch or different batches of prints, they can achieve extremely close color standards, greatly improving the quality stability of printed products and meeting the needs of high-end printing businesses with strict color requirements, such as the production of art albums, brand promotion posters and other printed products, reducing the scrap rate caused by color deviation, reducing production costs and enhancing corporate reputation.

[0040] The system architecture layer is used to build an architecture system that adapts to the large-scale production, multiple links and various equipment and materials in the printing industry. It provides a comprehensive adaptation solution for the large-scale production, multiple links and various equipment and materials in the printing industry, including distributed deep learning architecture, multi-level feedback self-calibration architecture and adaptive color calibration architecture. The distributed deep learning architecture distributes the training tasks of the deep learning model to multiple computing nodes for parallel processing. Different nodes focus on data processing of different printing product lines or different color spaces. The central coordinator integrates the results of each node. The multi-level feedback self-calibration architecture performs preliminary correction of the image color in the pre-press stage, and then re-enters the correction result into the system for secondary detection and analysis in the mid-press stage. The deep learning model parameters or correction strategies are adjusted according to the feedback information at different levels, and a feedback channel is established from the post-press quality inspection stage to the pre-press design and printing stages. The adaptive color calibration architecture is pre- The printing equipment and material feature library established first can quickly identify and match the corresponding model configuration during printing production, and automatically adjust the structure and parameters of the deep learning model. The distributed deep learning architecture uses multiple computing nodes to process massive printing color data in parallel, accelerate the model training process, and improve production efficiency. The multi-level feedback self-calibration architecture runs through the pre-press, mid-press, and post-press links to form a closed-loop feedback control system, which can timely detect and correct color deviations and optimize the entire printing process. The adaptive color calibration architecture is based on a rich printing equipment and material feature library, and can quickly adjust the model configuration according to different printing machines, paper, ink and other conditions, so that the system can easily cope with the complex situation of frequent equipment and material changes in the printing industry and different production tasks, enhancing the versatility and flexibility of the system, reducing the time and labor costs required to readjust the system due to changes in equipment and materials, and improving the company's production resilience.

[0041] The data processing and optimization layer is used to preprocess, expand and dynamically update the printing color data. The data processing and optimization layer includes a data enhancement and synthesis module, a semantic segmentation data processing module and a real-time data streaming processing and incremental learning module. The data enhancement and synthesis module expands the training data set by performing transformation operations such as rotation, scaling, and contrast adjustment on the existing printed color images and artificially synthesizing image data with specific color deviation characteristics. The semantic segmentation data processing module uses semantic segmentation technology to divide different areas in the printed image according to color categories, object categories or functional areas, and performs color deviation analysis and correction respectively; the real-time data streaming processing and incremental learning module quickly processes and extracts features of the color data generated in real time during the printing production process, and uses incremental learning technology to adapt the model to the new printing data distribution and color change trends. The data enhancement and synthesis module is based on the physical and mathematical models of printing color to expand The rich data diversity enables the deep learning model to better learn various color deviations and improve the generalization ability of the model, so that it can accurately predict and correct deviations when facing complex and changeable color data in actual printing. The semantic segmentation data processing module uses an improved deep learning semantic segmentation model to perform fine-grained regional division and color analysis on printed images, and performs corrections based on the characteristics of different regions to further improve the accuracy and visual effects of color management. The real-time data streaming processing and incremental learning module uses sliding window technology and distributed caching mechanism to achieve efficient processing of real-time data in the printing production process and dynamic model updates, ensuring that the model can adapt to the distribution of printing data and color change trends in a timely manner, and always maintain good correction performance. There is no need to frequently interrupt production for large-scale model retraining, which improves production continuity and equipment utilization, and helps enterprises achieve intelligent and efficient printing production management.

[0042] Each sub-architecture in the entire system architecture layer, when used, further includes the following:

[0043] Each computing node in the distributed deep learning architecture has an independent data storage unit and computing unit. The data storage unit is used to temporarily store the assigned specific printing color data subset. The computing unit performs local training operations of the deep learning model based on the stored data. The computing nodes are connected by high-speed data transmission channels to ensure the efficiency of data interaction and model integration.

[0044] The feedback information in the multi-level feedback self-calibration architecture includes color deviation value, deviation distribution area, color space coordinate difference and corresponding printing link identification information. The color deviation value is accurate to three decimal places, the deviation distribution area is expressed in the image pixel coordinate range or the physical size area of ​​the printed product, the color space coordinate difference is quantitatively described based on the commonly used color space in the printing industry such as CMYK or LAB, and the printing link identification information clearly distinguishes the sources of pre-press, mid-press and post-press.

[0045] The data stored in the printing equipment and material feature library in the adaptive color calibration architecture covers printing press brand and model, printing press mechanical parameters such as pressure, speed, dot gain curve, paper types such as coated paper, offset paper, cardboard, including their grammage, roughness, whiteness, etc., ink brand and model and characteristics such as hue, saturation, drying speed, viscosity and other information. The data is stored in the form of a structured database to facilitate quick retrieval and model configuration matching.

[0046] When used, each module in the entire data processing and optimization layer further includes the following:

[0047] The color data synthesis algorithm in the data enhancement and synthesis module is constructed based on the physical model of printing color and the mathematical statistical model. When synthesizing image data with specific color deviation characteristics, it can accurately generate simulated images based on the deviation parameters set by the user, such as the degree of deviation, direction of deviation, and deviation distribution pattern. The generated image data format is consistent with the real printed image data format and can be directly used for deep learning model training.

[0048] The semantic segmentation technology in the semantic segmentation data processing module adopts a deep learning semantic segmentation model. The model structure is improved based on the U-Net architecture and a multi-scale feature fusion module is added. It can accurately identify text, image body, background and other areas in printed images at different resolutions. The segmentation accuracy reaches the pixel level and it has good adaptability to different types of printed images, such as posters, book pages, packaging boxes, etc.

[0049] The data streaming processing algorithm in the real-time data streaming processing and incremental learning module adopts sliding window technology and distributed caching mechanism. The sliding window size is dynamically adjusted according to the printing data flow and processing speed to ensure that data is not lost and can be processed in time during data flow peaks. The distributed caching mechanism caches recently processed data in multiple nodes to facilitate the incremental learning module to quickly obtain data for model updates. The validity period of the cached data is dynamically set according to the continuity of the printing task and the timeliness of the data. Specific embodiment 2:

[0051] like Figure 1-3 As shown, according to the contents in the above specific embodiments, the following contents are further disclosed:

[0052] In the overall system architecture above, the specific contents of the algorithms for each unit of the algorithm layer are as follows:

[0053] The DBN model training in the combination unit of deep belief network DBN and restricted Boltzmann machine RBM adopts layer-by-layer pre-training. The training is combined with overall fine-tuning. In the layer-by-layer pre-training stage, each layer is set according to the complexity of the printed color data. The number of training iterations is gradually reduced from the bottom layer to the top layer. The overall fine-tuning stage adopts the optimization based on gradient descent. The algorithm is optimized, the learning rate and momentum parameters are adjusted, and the accuracy of printing color deviation correction is used as the optimization objective function.

[0054] The convolutional neural network (CNN) and the recurrent neural network (RNN) are combined into a unit, using the long short-term memory (LSTM) network structure.

[0055] Forget gate: f t =σ(W f ·[h t-1 , x t ]+b f )

[0056] in:

[0057] f t : represents the output of the forget gate at time t, which determines how much past information to forget from the cell state;

[0058] σ: is the activation function, usually the sigmoid function, which maps the input value to the interval between 0 and 1 to control the transmission ratio of information. Here, the sigmoid function processes the linear combination results in the brackets so that the output value is between 0 (completely forgotten) and 1 (completely retained);

[0059] W f : is the weight matrix of the forget gate, which is concatenated with the input vector [h t-1 , x t ] performs matrix multiplication operations to weight the input information to determine the importance of different parts of the information;

[0060] h t-1 : It is the hidden state vector of the previous moment, which contains some state information of the network processing information at the previous moment and serves as a reference for the forget gate decision at the current moment;

[0061] x t : is the input vector at the current moment. When processing a printed image sequence, it may be the pixel feature vector at the current position in the image sequence, etc.

[0062] b f : is the bias vector of the forget gate, which is used to adjust the offset of the linear combination to better adapt to the data characteristics;

[0063] Input gate: i t =σ(W i ·[h t-1 , x t ]+b i )

[0064] The explanation is similar to the forget gate, except that here it determines how much new information to update to the cell state;

[0065] Cell state update gate:

[0066] in:

[0067] represents the processed candidate cell state, which is obtained by weighting the input information (by W C and bias b C ), and then processed by the tanh activation function. The tanh function maps the input value to the interval between -1 and 1, which can better represent the update amplitude of the cell state;

[0068] The explanations of other symbols are similar to the previous ones, which are to weight and process the input information in different ways to obtain the appropriate candidate states for update;

[0069] Cell status update:

[0070] in:

[0071] C t : is the cell state updated at the current moment, which is determined by the forget gate f t The last cell state C that determines the forgotten part t-1 , plus the input gate i t Determine the candidate cell states for the update In this way, the cell state can be reasonably updated according to the new input information and the past state;

[0072] Output gate: o t =σ(W o ·[h t-1 , x t ]+b o )

[0073] Similar to the forget gate, it determines how much information is output from the current cell state as the hidden state at the current moment;

[0074] Hide status update:h t =o t ×tanh(C t )

[0075] in:

[0076] h t : is the hidden state at the current moment, which is the cell state C after the tanh function processing with the output ratio determined by the output gate t The hidden state will be used as part of the input for the network processing at the next moment, and can also be used for subsequent analysis or output tasks. For example, when analyzing the color deviation law of a printed image sequence, it may be used as a basis for judging the color change trend.

[0077] Implementation steps:

[0078] Initialize the weight matrix W f , W i , W C , W o and the bias vector b f , b i , b C , b o , usually random initialization or some specific initialization methods can be used, such as the Xavier initialization method mentioned,

[0079] For each time t, calculate the forget gate f in sequence t , input gate i t , cell state update gate Cell state update C t , output gate o t and the hidden state update h t , perform corresponding matrix multiplication, vector addition and activation function processing operations according to the above formula;

[0080] The hidden state h t Used for subsequent tasks, such as combining with CNN to perform in-depth extraction of features such as color pixel distribution and pattern texture of printed images, as well as analyzing the variation patterns of color deviation across the entire printed page and between different printing batches.

[0081] Operation logic: LSTM controls the flow and update of information through four gating mechanisms: forget gate, input gate, cell state update gate and output gate. When processing printed image sequences, for the input at each moment (such as the pixel feature vector at a certain position in the image sequence), the forget gate determines how much past cell state information to forget, the input gate determines how much new information to update to the cell state, the cell state update gate generates candidate cell state update amounts, and then updates the cell state through the cell state update formula. Finally, the output gate determines how much information to output from the updated cell state as the hidden state at the current moment. In this way, the network can effectively handle long-distance dependencies in printed image sequences, that is, it can take into account the relationship between different positions in the image sequence, so as to better analyze features such as color deviation.

[0082] Beneficial effects: It can effectively handle long-distance dependencies in printed image sequences, which is very important for analyzing the color deviation patterns of printed images across the entire page and between different batches, because the color deviation of a printed image may be related not only to the pixel at the current position, but also to the pixels at the previous or subsequent positions. LSTM can capture these complex relationships through a gating mechanism, thereby more accurately analyzing and predicting color deviations. By deeply extracting features such as color pixel distribution and pattern texture of printed images and combining them with collaborative work with CNN, printed images can be analyzed more comprehensively and accurately, the accuracy of color correction can be improved, and the high consistency of color of printed products throughout the production process can be ensured.

[0083] The autoencoder in the combined unit of the generative adversarial network GAN and the autoencoder adopts a symmetric encoder-decoder The encoder structure consists of multiple convolutional layers and pooling layers, which are used to compress the printed color image into a low-dimensional feature vector. The decoder consists of multiple layers of deconvolution layers and upsampling layers to reconstruct the low-dimensional feature vector into a graph of the original image size. The image is reconstructed and a jump connection is set between the encoder and the decoder to retain the original feature information of the image and improve the quality of the reconstructed image. quantity.

[0084] The fusion generative adversarial network GAN and autoencoder combination unit further includes the following:

[0085] Convolutional Neural Network based generator:

[0086] G(z)=σ(W g1 tanh(W g2 ·z+b g2 )+b g1 )

[0087] in:

[0088] G(z): is the output of the generator, i.e. the generated simulated printing image data;

[0089] σ: activation function, such as the sigmoid function, is used to map the result of the internal processing of the generator to a suitable interval to generate image data that meets the requirements. Here, the result after multi-layer processing is mapped to between 0 and 1 (if it is processing image pixel values, etc.);

[0090] W g1 , W g2 : It is the weight matrix of different layers in the generator, which is used to weight the input to determine the importance of different parts of information and gradually shape the characteristics of the image in the process of generating the image;

[0091] b g1 , b g2 : is the bias vector of different layers in the generator, which is used to adjust the offset of the linear combination to better adapt to the data characteristics;

[0092] z: is the random noise vector input to the generator. Based on this random noise, the generator generates simulated printing image data through internal multi-layer processing (for example, here it is processed by and then by the tanh activation function, and finally by and again);

[0093] Convolutional neural network based discriminator:

[0094] D(x)=σ(W d1 tanh(W d2 ·x+b d2 )+b d1 )

[0095] in:

[0096] D(x): is the output of the discriminator. Its value is between 0 and 1, indicating the probability that the input image (x) is a real printed image or a simulated image generated by the generator. If the output is close to 1, it is considered to be a real image; if the output is close to 0, it is considered to be a generated image.

[0097] Similar to the generator, W d1 , W d2 is the weight matrix of the discriminator, b d1 、b d2 is the bias vector of the discriminator, is the image input to the discriminator, which can be a real printed image or a simulated image generated by the generator;

[0098] Autoencoder:

[0099] Encoder: h = f e nc(x)=σ(W enc ·x+b enc )

[0100] in:

[0101] h: is the output of the encoder, that is, the compressed low-dimensional feature vector;

[0102] f e nc(x): represents the encoding function of the encoder for the input image;

[0103] W enc : is the weight matrix of the encoder, which is used to weight the input image for compression purposes;

[0104] b enc : is the bias vector of the encoder, which is used to adjust the offset of the linear combination to better adapt to the data characteristics;

[0105] Decoder: x′=f dec (h)=σ(W dec ·h+b dec)

[0106] in:

[0107] x′: is the output of the decoder, i.e. the reconstructed image, which is expected to be as close as possible to the original input image;

[0108] f dec (h): represents the decoding function of the decoder for the low-dimensional feature vector;

[0109] W dec : is the weight matrix of the decoder, which is used to weight the low-dimensional feature vector for reconstruction;

[0110] b dac : is the bias vector of the decoder, which is used to adjust the offset of the linear combination to better adapt to the data characteristics.

[0111] Implementation steps:

[0112] Initialize the weight matrix W of the generator and discriminator g1 , W g2 , W d1 , W d2 and the bias vector b g1 、b g2 、b d1 、b d2 , and initialize the weight matrix W of the autoencoder enc , W dec and the bias vector b enc 、b dec , usually using random initialization or other appropriate initialization methods;

[0113] During the training phase:

[0114] For the generator: a random noise vector z is sampled from a random noise distribution, z is input to the generator G(z), and processed according to the above formula to obtain the generated simulated printing image data G(z);

[0115] For the discriminator:

[0116] Select a real printed image x from the real printed image data set, and use the simulated printed image data G(z) generated by the generator as input. The real printed image x and the simulated printed image data G(z) can be selected alternately or in a certain proportion. Input the real printed image x and the simulated printed image data G(z) into the discriminator D(x) respectively, and process them according to the above formula to obtain the discriminator's discrimination result D(x) for the real image and the generated image;

[0117] For the autoencoder:

[0118] Select the input image x from the real printed image dataset and input x into the encoder f enc (x), process x according to the above formula to obtain a low-dimensional feature vector h, and input h into the decoder f dec (h), process according to the above formula to obtain the reconstructed image x′;

[0119] According to the discrimination results of the discriminator and the reconstruction results of the autoencoder, the weight matrices and bias vectors of the generator, discriminator and autoencoder are updated. Usually, an optimization algorithm based on gradient descent is used, such as stochastic gradient descent (SGD) or its variants (such as Adam, etc.), so as to achieve the purpose of generating high-quality simulated images, accurately distinguishing the authenticity of images, and reconstructing images well.

[0120] Operation logic: The generator attempts to generate printed image simulation data close to the target printing color standard from a random noise vector, and the discriminator distinguishes between real printed images and generated images through continuous adversarial training, that is, the generator continuously generates better simulated images to "cheat" the discriminator, and the discriminator continuously improves its own discrimination ability to "see through" the generator's "scam". In this process, the performance of both the generator and the discriminator is improved, and finally the generator can generate high-quality simulated images.

[0121] The autoencoder compresses and reconstructs the input printed color image. The encoder compresses the image into a low-dimensional feature vector, and then the decoder reconstructs the low-dimensional feature vector into an image of the original image size. In this process, the autoencoder learns the latent color representation of the image, and by setting a jump connection between the encoder and the decoder, the original feature information of the image is retained to improve the quality of the reconstructed image. Then, the generator and the autoencoder work together to correct the printed color deviation and ensure color consistency. For example, the simulated image generated by the generator can be used as a reference to correct the color deviation of the real image. The latent color representation learned by the autoencoder can help better understand and process the color characteristics of the image.

[0122] Beneficial effects: The generator and the discriminator improve the quality of generated images through adversarial training, so that the generated simulated printing images can be closer to the target printing color standard. This is very helpful for analyzing and correcting printing color deviations, because by comparing with real images, the color deviation can be more accurately found. The autoencoder learns the potential color representation of the image and can reconstruct the image better. By working together with the generator, it can more comprehensively guarantee the correction of printing color deviations and color consistency, improve the high consistency of color of printed products throughout the production process, and meet the needs of high-end printing business with strict color requirements.

[0123] The RNN model in the combination unit of convolutional neural network CNN and recurrent neural network RNN ​​adopts long short-term memory network LSTM structure, the LSTM unit contains a forget gate, an input gate, a cell state update gate and an output gate, which are controlled by a gating mechanism. The flow and update of control information can effectively handle the long-distance dependencies in the printed image sequence, and the layers of the LSTM network The number is dynamically adjusted according to the complexity of the printed image and the accuracy requirements of the color deviation analysis. The initial weight of the network is Xavier initialization method to set .

[0124] Deep Belief Network DBN and Restricted Boltzmann Machine RBM Combination Unit

[0125] Restricted Boltzmann Machine RBM energy function:

[0126]

[0127] in:

[0128] E(v, h): is the energy function of RBM, which is a function of the visible unit state vector v and the hidden unit state vector h. Training aims to reduce its value.

[0129] a i : The bias vector element of the visible unit, adjusting the energy contribution of the visible unit;

[0130] b j : The bias vector element of the hidden unit adjusts the energy contribution of the hidden unit;

[0131] w ij : The weight matrix elements connecting visible units and hidden units determine the degree of mutual influence;

[0132] v i : The i-th element of the visible unit state vector, such as the printed color data eigenvalue;

[0133] h j : The jth element of the hidden unit state vector, resulting from the interaction with the visible unit.

[0134] Conditional probability distribution:

[0135] The conditional probability of a visible unit given a hidden unit:

[0136]

[0137] in:

[0138] P(v i |h): Given the hidden unit state vector h, the i-th element v of the visible unit state vector i The conditional probability of

[0139] σ: sigmoid function, which maps the input to a value between 0 and 1 to represent probability;

[0140] The conditional probability of a hidden unit given a visible unit:

[0141]

[0142] in:

[0143] P(h j |v): Given the visible unit state vector v, the jth element h of the hidden unit state vector j The conditional probability of

[0144] Deep belief network DBN, composed of multiple RBM stacked:

[0145] Assume that the k-1th RBM has been trained and its hidden unit state vector h k-1 As the input of the kth RBM;

[0146] The training adopts a combination of layer-by-layer pre-training and overall fine-tuning:

[0147] Layer-by-layer pre-training: For each RBM (from bottom to top), pre-train according to the RBM training method (by minimizing the energy function and using conditional probability distribution training), and set the number of training iterations for each layer in a way that gradually decreases from bottom to top;

[0148] Overall fine-tuning: Adopting the optimization algorithm based on gradient descent (such as Adam), adjusting the learning rate and momentum parameters, taking the accuracy of printing color deviation correction as the optimization objective function, fine-tuning the entire DBN, mining high-level abstract features, and providing accurate basis for correction.

[0149] Implementation steps:

[0150] Initialize the RBM weight matrix w ij , visible unit bias vector a i and the hidden unit bias vector b j Random initialization or other suitable methods;

[0151] Train a restricted Boltzmann machine (RBM):

[0152] Randomly initialize the visible unit state vector and the hidden unit state vector h,

[0153] Multiple iterations of training: Using conditional probability distribution to calculate P(h j |v), update the hidden unit state vector h according to probability sampling (such as Gibbs sampling), and use the conditional probability distribution to calculate P(v i |h), update the visible unit state vector v according to probability sampling, calculate the gradient of the energy function with respect to the parameters based on the new v and h, and update the weight matrix w ij , visible unit bias vector a i and the hidden unit bias vector b j , until the training stop condition is met, such as reaching a preset number of iterations or the energy function converges;

[0154] Training Deep Belief Network DBN:

[0155] Layer-by-layer pre-training: The bottom layer uses the original printed color data as the visible unit state vector v of the first RBM, and trains the hidden unit state vector h1. The subsequent k-th RBM (k>1) uses the hidden unit state vector h of the previous RBM. k-1 is the visible unit state vector v, which is trained according to the principle of decreasing the number of iterations from the bottom layer to the top layer.

[0156] Overall fine-tuning: After completing the layer-by-layer pre-training, use an optimization algorithm based on gradient descent (such as Adam), set appropriate learning rate and momentum parameters, take the accuracy of printed color deviation correction as the objective function, calculate its gradient with respect to all DBN parameters, adjust the parameters to mine high-level abstract features, adapt to the characteristics of printed color data, and provide a basis for correction.

[0157] Operation logic: In the restricted Boltzmann machine (RBM), the operation is based on the energy function and conditional probability distribution. The energy function value is reduced by adjusting parameters such as weights, so that the conditional probability distribution of hidden units when the visible unit state is given and vice versa is consistent with the actual data distribution, so as to capture the intrinsic characteristics and associations of color data. For example, Gibbs sampling is used to update the visible and hidden unit states to achieve weight training.

[0158] The deep belief network DBN is based on multiple RBM stacks. During pre-training layer by layer, each RBM performs feature transformation on the input data (starting from the original color data and then the state vector of the upper hidden unit) to mine more abstract features. The features become more abstract as the number of layers increases. In the overall fine-tuning stage, the accuracy of printing color deviation correction is used as the objective function to adjust all network parameters so that the network adapts to the needs of printing color data and mines high-level abstract features to provide an accurate basis for correction. These features help to reveal the deep-seated causes of printing color deviation and the intrinsic connection between regional color features.

[0159] Beneficial effects: RBM performs unsupervised pre-training on printing color data and automatically learns the basic feature distribution and structure without the need for a large amount of manually labeled data, laying the foundation for DBN construction and training, and facilitating the printing color management system to obtain preliminary data features. After multiple RBMs are stacked into a DBN, high-level abstract features in complex printing color data can be mined to more accurately reflect the essence of the data, such as the deep causes of color deviations in different batches and regional color associations, providing an accurate basis for subsequent automatic correction, improving the color consistency of printed products, reducing the scrap rate, and meeting the needs of high-end printing business. In addition, by combining layer-by-layer pre-training with overall fine-tuning, the generalization ability of the network is improved, enabling it to cope with complex and changeable color data in actual printing and accurately predict and correct deviations. Specific embodiment three:

[0161] like Figure 1-3As shown, according to the contents in the above specific embodiments, the following contents are further disclosed:

[0162] The entire operation steps in the above system architecture are as follows:

[0163] Sp1: Data acquisition and input: The original printing image data is obtained from the image acquisition devices on the printing production line, such as scanners, cameras, etc. These data contain the color information of the printed product and the pixel distribution of the image. At the same time, the system reads the relevant data in the printing equipment and material feature library, including the brand model of the printing machine, mechanical parameters, paper type and characteristics, ink brand model and characteristics, etc. These data serve as the basic configuration information for the system operation;

[0164] Sp2: Data processing and optimization layer operations:

[0165] Sp2.1: The data enhancement and synthesis module processes the original printed image data. First, the original data set is expanded by performing transformation operations such as rotation, scaling, and contrast adjustment on the existing printed color image. Then, according to the deviation parameters set by the user, such as deviation degree, deviation direction, and deviation distribution pattern, the color data synthesis algorithm based on the printed color physical model and mathematical statistical model is used to artificially synthesize image data with specific color deviation characteristics. The generated image data format is consistent with the real printed image data format. The processed data is sent to the algorithm layer for subsequent analysis.

[0166] Sp2.2: The semantic segmentation data processing module uses a deep learning semantic segmentation model based on the improved U-Net architecture and adds a multi-scale feature fusion module to divide the printed image according to color category, object category or functional area, such as dividing the image into text area, image main area, background area, etc. The color features are extracted for different areas, and the feature data of these areas and the corresponding color deviation analysis requirement data are transmitted to the algorithm layer;

[0167] Sp2.3: The real-time data streaming processing and incremental learning module uses sliding window technology and distributed cache mechanism to process the color data generated in real time during the printing production process. The sliding window size is dynamically adjusted according to the printing data flow and processing speed to ensure that data is not lost and can be processed in time during the peak of data flow. The distributed cache mechanism caches the recently processed data in multiple nodes, and then extracts the data features and transmits them to the algorithm layer. At the same time, the incremental learning module uses the optimization algorithm based on gradient descent to adjust the learning rate and momentum parameters of the model according to the newly incoming data and cached data, and uses the accuracy of printing color deviation correction as the optimization objective function to dynamically update the model to adapt the model to the new printing data distribution and color change trend;

[0168] Sp3: Algorithm layer operations:

[0169] Sp3.1: The convolutional neural network (CNN) and recurrent neural network (RNN) combination unit receives image data from the data processing and optimization layer. CNN deeply extracts features such as color pixel distribution and pattern texture of the printed image. RNN analyzes the variation patterns of color deviation in the entire printed page and between different printing batches based on the spatial and temporal dependencies between printed image sequences, and generates color deviation analysis results and preliminary correction strategy recommendation data, which are then transmitted to the system architecture layer.

[0170] Sp3.2: The combined unit of the generative adversarial network GAN and the autoencoder receives image data. The generator of the GAN generates simulated printing image data similar to the target printing color standard according to the target printing color standard. The discriminator distinguishes between the real printing image and the generated image. The adversarial training of the two improves the quality of the generated image. The autoencoder compresses and reconstructs the input printing color image and learns the potential color representation of the image. The two work together to generate color correction data and color consistency evaluation data, which are transmitted to the system architecture layer.

[0171] Sp3.3: After receiving the data, the deep belief network DBN and restricted Boltzmann machine RBM combination unit performs unsupervised pre-training on the printing color data through RBM, learns the basic feature distribution and structure of the printing color data, and then stacks it to construct a DBN to mine the high-level abstract features hidden in the complex printing color data, generate accurate color deviation feature data and correction parameter data, and transmit them to the system architecture layer.

[0172] Sp4: System architecture layer operations:

[0173] Sp4.1: The distributed deep learning architecture integrates various data transmitted from the algorithm layer and the data processed by each computing node. Each computing node has an independent data storage unit and a computing unit. The data storage unit temporarily stores the specific printing color data subset assigned to it. The computing unit performs local training operations on the deep learning model based on the stored data. The computing nodes are connected by high-speed data transmission channels to ensure the efficiency of data interaction and model integration. Finally, the overall color correction model data is generated and transmitted to the multi-level feedback self-calibration architecture.

[0174] Sp4.2: Multi-level feedback self-calibration architecture In the pre-press stage, the image color is initially corrected according to the received color correction model data, and the relevant correction information and color deviation data are recorded, including the color deviation value accurate to three decimal places, the deviation distribution area is represented by the image pixel coordinate range or the physical size area of ​​the printed product, and the color space coordinate difference is quantitatively described according to the commonly used color space in the printing industry, such as CMYK or LAB, and the corresponding printing link identification information clearly distinguishes the pre-press stage. In the printing stage, the correction result is input into the system again for secondary detection and analysis, and the deep learning model parameters or correction strategies are adjusted according to the feedback information at different levels. The relevant data of the printing stage is transmitted to the post-press quality inspection stage, the algorithm layer, and the data processing and optimization layer for further optimization and adjustment. The post-press quality inspection stage detects the printed product, and transmits the detected color deviation data and the corresponding printed product information (including the printing link identification information to clearly distinguish the post-press stage) back to the pre-press design and printing stage, the algorithm layer, and the data processing and optimization layer to form a closed-loop feedback control system to optimize the entire printing process;

[0175] Sp4.3: The adaptive color calibration architecture uses the pre-established printing equipment and material feature library to quickly identify the current printing equipment and material information during printing production, match the corresponding model configuration, and automatically adjust the structure and parameters of the deep learning model to ensure that the system can adapt to the color management requirements under different printing machines, paper, ink and other conditions, and transmit the adjusted model data to the distributed deep learning architecture for integration processing;

[0176] Sp5: Output and application: After a series of system processing and correction, the final output is high-precision color-corrected printing image data or printing color control parameter data. These data are used in the printing production process to directly control the color output of the printing equipment, ensure that the printed products meet high-precision color standards, meet the needs of high-end printing business, improve the quality stability of printed products, reduce scrap rate, reduce production costs and enhance corporate reputation. Specific embodiment four:

[0178] like Figure 1-3 As shown, according to the contents in the above specific embodiments, the following contents are further disclosed:

[0179] In the above system architecture, the entire data processing trend is as follows:

[0180] Data source: The original data mainly comes from the printing image data obtained by the image acquisition equipment on the printing production line, as well as the equipment and material related data in the printing equipment and material feature library.

[0181] Data processing and optimization layer:

[0182] Input: original printing image data, printing equipment and material feature library data.

[0183] Output: expanded image data set, regional feature data and color deviation analysis requirement data, real-time data features and dynamically updated model parameter data.

[0184] Algorithm layer:

[0185] Input: Data output by the data processing and optimization layer.

[0186] Output: color deviation analysis results and preliminary correction strategy recommendation data, color correction data and color consistency evaluation data, precise color deviation feature data and correction parameter data.

[0187] System architecture layer:

[0188] Input: data output by the algorithm layer and data transmitted between its own architecture modules.

[0189] Output: overall color correction model data, adjusted model data, closed-loop feedback information data.

[0190] Data flow: The original data flows from the printing image acquisition device and feature library to the data processing and optimization layer. The data processed by the data processing and optimization layer flows to the algorithm layer. The data processed by the algorithm layer flows to the system architecture layer. The architecture modules within the system architecture layer transmit data to each other, and the final processing result data flows to the printing production equipment to control color output or feedback to other related links for optimization and adjustment.

[0191] Final result: Obtain high-precision color-corrected printing image data or printing color control parameter data, achieve high color consistency of printed products throughout the entire production process, improve the quality stability of printed products, reduce scrap rates, enhance the company's production resilience, and achieve intelligent and efficient printing production management. Specific embodiment five:

[0193] like Figure 1-3 As shown, according to the contents in the above specific embodiments, the following contents are further disclosed:

[0194] In the above system architecture, in actual use, the corresponding hardware structure is as follows:

[0195] The hardware architecture of the algorithm layer is as follows:

[0196] The hardware of the convolutional neural network CNN and recurrent neural network RNN ​​combination unit is as follows:

[0197] Computing core: High-performance graphics processing units (GPUs), such as the NVIDIA A100 series GPUs, are used. They have a large number of CUDA cores that can process convolution operations in CNN and sequence calculations in RNN in parallel, accelerating feature extraction and deviation law analysis. For example, when processing high-resolution printed images, the powerful computing power of A100 can quickly complete the color pixel distribution and pattern texture feature extraction tasks, ensuring that the color deviation change law between different printing batches can be analyzed in a short time;

[0198] Memory configuration: Equipped with large-capacity high-speed video memory, such as 32GB or higher GDDR6 video memory, which helps store intermediate data in the CNN and RNN operations, including image feature matrices, RNN hidden states and other information, to avoid data transmission delays and reduced computing efficiency due to insufficient memory;

[0199] Data transmission interface: Use PCIe4.0 or higher version interface to connect to the host system to ensure high-speed data transmission between the host memory and GPU video memory. For example, when transferring printed image data from the host to the GPU for calculation, the PCIe4.0 interface can achieve a data transmission rate of tens of GB per second, quickly loading the image data into the GPU for processing.

[0200] The hardware of the combined unit of the fusion generative adversarial network GAN and the autoencoder is as follows:

[0201] Dual GPU architecture: Two GPUs work together, one for the training calculation of the GAN generator and discriminator, and the other for the compression and reconstruction calculation of the autoencoder. For example, for GAN training to generate high-quality printed image simulation data, the GPU can use its powerful parallel computing capabilities to accelerate the adversarial training process between the generator and the discriminator; while the autoencoder runs independently on another GPU to ensure the efficiency of image compression and reconstruction. The two interact with each other through high-speed data buses (such as NVLink) to achieve synergy.

[0202] High-speed storage system: Equipped with a solid-state drive array based on the NVMe protocol, which is used to store a large amount of image data during the GAN training process, model parameters of the generator and discriminator, and feature vector data of the autoencoder. The read and write speed of the NVMe solid-state drive can reach several GB per second, which can quickly load and save data, reduce data reading waiting time, and improve the operating efficiency of the entire combined unit;

[0203] Network interface card: Install a high-performance network interface card that supports a network transmission rate of 10Gbps or higher. This helps to achieve fast data transmission when synchronizing data and sharing model parameters between GAN and autoencoder, ensuring that the two can work closely together to improve the effect of color correction and consistency assurance.

[0204] The hardware architecture of the deep belief network DBN and restricted Boltzmann machine RBM combination unit is as follows:

[0205] Cluster computing architecture: Build a cluster consisting of multiple servers, each equipped with a multi-core CPU, such as Intel Xeon processors, and a certain number of GPU accelerator cards. In the RBM pre-training stage, use the multi-core advantage of the CPU to perform parallel processing of large-scale data to accelerate the learning process of data feature distribution and structure. When building DBN stacks and training the whole, use the parallel computing capabilities of GPU accelerator cards to perform deep feature extraction and model optimization. For example, a cluster of 10 servers, each equipped with two 24-core CPUs and four GPU accelerator cards, can efficiently process massive amounts of printing color data and mine the high-level abstract features hidden in it.

[0206] Distributed storage system: Use distributed file systems, such as Ceph or GlusterFS, to store printing color data and model parameters of DBN and RBM. Distributed file systems can store data in multiple storage nodes, providing high reliability and high scalability, ensuring secure storage and fast access to data during large-scale data processing;

[0207] High-speed Internet network: Deploy a high-speed InfiniBand network inside the cluster to achieve low-latency, high-bandwidth data transmission between servers. During the DBN model training process, data and model parameters need to be frequently exchanged between servers. The InfiniBand network can provide a transmission rate of up to 100Gbps or even higher, reducing data transmission time and improving the training efficiency of the entire combined unit.

[0208] The hardware architecture of the system architecture layer is as follows:

[0209] The distributed deep learning architecture hardware is as follows:

[0210] Computing node cluster: It consists of multiple computing nodes. Each computing node uses an industrial standard server, equipped with a dual-core CPU (such as Intel Xeon Gold series), a large-capacity memory (such as 128GB or higher) and multiple GPU accelerator cards (such as NVIDIA T4 or higher models). Different computing nodes are connected through high-speed Ethernet switches to form an internal high-speed network to achieve data communication and task collaboration between nodes. For example, a distributed deep learning architecture may contain 50 computing nodes. The GPU accelerator card of each node can process data of different printing product lines or color spaces in parallel, accelerate the model training process, and improve the whole system's ability to process large-scale printing data;

[0211] Central coordination server: Equipped with high-performance servers, multi-core CPUs and large-capacity memory, running distributed task scheduling software and data integration programs, the central coordination server is responsible for allocating the training tasks of the deep learning model to each computing node, monitoring the running status of the nodes, collecting and integrating the calculation results of each node, and ensuring the stable operation and efficient collaboration of the entire distributed system. For example, the central coordination server reasonably allocates printing color data processing tasks according to the load conditions and task priorities of each computing node to ensure full utilization of system resources;

[0212] Storage Area Network (SAN): Build a SAN based on Fibre Channel or iSCSI protocol to connect all computing nodes and the central coordination server. SAN is used to centrally store massive amounts of printing color data, model training data, and parameter files of deep learning models. The high-speed data storage and access capabilities it provides can meet the high requirements of distributed systems for data reading and writing, and ensure the rapid sharing and transmission of data between nodes. For example, the storage capacity of SAN can reach tens of TB or even higher, which can store a large amount of color data and model files accumulated by printing companies over the years.

[0213] The hardware architecture of the multi-level feedback self-calibration architecture is as follows:

[0214] Online testing equipment: High-precision color testing equipment, such as spectrophotometers or color measuring instruments, are deployed in the pre-press, mid-press, and post-press stages of the printing production line. These devices can collect color data of printed products in real time, including information such as color spectral reflectance and chromaticity coordinates, and transmit the data to the back-end processing system. For example, the testing equipment in the pre-press stage can perform color testing on the input images or films to ensure the accuracy of the color of the materials entering the printing process; the testing equipment in the mid-press stage can monitor the color changes during the printing process and detect deviations in a timely manner; the testing equipment in the post-press stage performs the final color quality inspection on the finished product to provide a basis for feedback and adjustment;

[0215] Data transmission network: Industrial Ethernet is used to build a stable data transmission network from online detection equipment to the back-end processing system. The network has a redundant design to ensure the reliability of data transmission and prevent color deviation data loss or transmission delay due to network failure. For example, the network adopts a ring topology. When a link fails, the data can automatically switch to the backup link for continued transmission to ensure the continuous operation of the system.

[0216] Feedback control server: A dedicated feedback control server is set up to receive color deviation data from detection equipment in each link, and analyze and process it according to preset algorithms and strategies. The feedback control server runs feedback control software, which can determine whether it is necessary to adjust the deep learning model parameters or correction strategy based on the color deviation value, deviation distribution area, color space coordinate difference, and printing link identification information, and send adjustment instructions to the corresponding links, such as prepress design software, printing machine control system, etc. For example, when post-press detection finds that the color deviation exceeds the standard, the feedback control server can analyze the cause based on the deviation information, and then send instructions to the prepress design software to modify the image color parameters, or send instructions to the printing machine control system to adjust parameters such as ink supply and printing pressure, to achieve closed-loop feedback control.

[0217] The adaptive color calibration architecture hardware is shown below:

[0218] Equipment and material identification system: Sensors and identification devices are installed on the printing equipment to read the brand and model of the printing press, mechanical parameters (such as pressure sensors, speed sensors, dot gain meters, etc.), and related information of paper and ink (such as paper type identification sensors, ink viscosity meters, etc.). These sensors and identification devices transmit the collected data to the adaptive control unit. For example, by installing an ink viscosity meter in the ink supply system of the printing press, the viscosity change of the ink can be monitored in real time, providing important parameter basis for adaptive color calibration;

[0219] Adaptive control unit: An industrial control computer or embedded controller is used as the hardware platform of the adaptive control unit to run the adaptive color calibration algorithm software. The unit receives data from the equipment and material identification system, quickly matches the corresponding color calibration model configuration by querying the pre-established printing equipment and material feature library, and generates control instructions to send to the color control system or data processing system of the printing press. For example, when it is recognized that the paper type has been changed, the adaptive control unit selects a suitable color calibration model from the feature library based on the paper's grammage, roughness, whiteness and other parameters, and adjusts the color output parameters of the printing press to ensure the accuracy of printed colors under different paper conditions;

[0220] Feature library storage server: equipped with high-performance servers as the storage medium for the printing equipment and material feature library. The server uses redundant array of disks (RAID) technology to ensure data security and reliability. The feature library storage server stores a large amount of detailed information on equipment and materials such as printing machines, paper, inks, and the corresponding color calibration model parameters, providing fast data retrieval and model matching services for the adaptive control unit. For example, the feature library stores dot gain curve data for hundreds of different brands and models of printing machines, optical property data for thousands of papers, and color property data for various inks, which can meet the diverse production needs of printing companies.

[0221] The hardware architecture of the data processing and optimization layer is as follows:

[0222] The data enhancement and synthesis module hardware is as follows:

[0223] Workstation host: A high-performance workstation is used as the operating platform for the data enhancement and synthesis module. The workstation is equipped with a multi-core CPU (such as Intel Core i9 series), a large-capacity memory (such as 64GB or higher) and a mid-to-high-end GPU (such as NVIDIA GeForce RTX series). The CPU is responsible for executing the logic control and data preprocessing tasks in the data enhancement and synthesis algorithm, and the GPU is mainly used to accelerate image transformation operations and color data synthesis calculations. For example, when performing transformation operations such as rotation and scaling on printed images, the CPU first reads the image and sets basic parameters, and then the GPU uses its parallel computing capabilities to quickly complete the transformation calculation of image pixels, thereby improving the efficiency of data enhancement;

[0224] Data storage devices: connect to large-capacity external storage devices, such as network attached storage (NAS) or direct attached storage (DAS). These storage devices are used to store original printing image data, enhanced image data sets, and synthesized image data with specific color deviation characteristics. NAS devices can share data through the network, which is convenient for multiple workstations or system modules to access; DAS devices provide higher local data read and write speeds to meet the needs of fast data storage and reading during data enhancement and synthesis. For example, a 10TB NAS device can store a large amount of image data resources of a printing company, which can be called and processed by the data enhancement and synthesis module at any time.

[0225] The hardware of the semantic segmentation data processing module is as follows:

[0226] Deep learning server: Use a dedicated deep learning server to run the semantic segmentation model. The server is equipped with multiple high-performance GPUs (such as NVIDIA A100 or V100 series) and large-capacity memory. The GPU is used to accelerate the training and reasoning process of the semantic segmentation model, and can quickly process pixel-level semantic segmentation tasks of printed images. For example, when performing semantic segmentation on high-resolution printed poster images, multiple A100 GPUs can process different areas of the image in parallel, accurately identify text, image subject, background and other areas, and the segmentation accuracy can reach pixel level.

[0227] Image acquisition card: If you need to directly collect image data from the printing production line for semantic segmentation processing, install a high-speed image acquisition card in the server. The image acquisition card can quickly convert the analog or digital signals collected by the printing image sensor and transmit them to the server memory to ensure the real-time and accuracy of the data. For example, an image acquisition card that supports 4K resolution and 60 frames per second can meet the requirements of high-speed printing production lines for image data acquisition and provide timely data input for semantic segmentation.

[0228] The hardware of the real-time data streaming processing and incremental learning module is as follows:

[0229] Edge computing devices: Edge computing devices, such as industrial-grade micro servers or high-performance embedded computers, are deployed near printing production equipment. These edge computing devices are responsible for preliminary processing and feature extraction of color data generated in real time during the printing production process. They use low-power, high-performance processors (such as multi-core processors with ARM architecture) and a certain amount of memory. For example, edge computing devices can perform real-time filtering, data compression and other pre-processing operations on the color data collected by the online detection equipment of the printing press, reduce the amount of data transmission, and improve the real-time performance of data processing;

[0230] Cloud computing platform: Incremental learning and model update tasks are deployed on the cloud computing platform. The cloud computing platform consists of a large-scale data center server cluster with powerful computing power and massive storage resources. The edge computing device uploads the processed feature data to the cloud computing platform through the network. The cloud computing platform uses its rich computing resources to run the incremental learning algorithm and update the deep learning model. For example, the cloud computing platform can dynamically allocate computing resources according to the printing data flow and processing requirements, and automatically expand server instances when the data flow peaks to ensure the timely completion of incremental learning tasks, so that the model can quickly adapt to the printing data distribution and color change trends;

[0231] Network communication equipment: Build a high-speed and stable network communication link between edge computing devices and cloud computing platforms, such as 5G wireless networks or high-speed fiber optic networks. Network communication equipment ensures that real-time data can be transmitted quickly and reliably between the two, meeting the timeliness and accuracy requirements of data streaming and incremental learning. For example, 5G wireless networks can provide low-latency, high-bandwidth data transmission services, enabling edge computing devices to upload data to the cloud computing platform in a timely manner. At the same time, the cloud computing platform can also quickly download updated model parameters to the edge computing devices, realizing real-time updates and applications of the model. Specific embodiment six:

[0233] like Figure 1-3 As shown, according to the contents in the above specific embodiments, the following contents are further disclosed:

[0234] In addition to being used in the printing industry, the above system architecture can also be used in other industries for color management and automatic calibration. Specific examples are as follows:

[0235] Textile printing and dyeing industry

[0236] Application scenarios and industry needs: In the textile printing and dyeing industry, color management is crucial from the pattern design of fabrics to the dyeing and processing of the final product. Different batches of dyes, fabrics of different materials, and different printing and dyeing processes will lead to color deviations. A high-precision color management system can ensure the consistency of fabric color, meet the strict requirements of fashion brands for product color standardization, reduce the defective rate caused by color difference, and improve production efficiency and product quality.

[0237] System architecture changes:

[0238] Algorithm layer:

[0239] In terms of data input, it is necessary to collect and organize data on textile fabric images and printing and dyeing process parameters. For example, the texture of the fabric, fiber material information, and parameters such as temperature, time, and dye concentration during the printing and dyeing process should be included in the data scope.

[0240] In terms of algorithm adjustment, the CNN-RNN combination unit can focus on analyzing the relationship between fabric texture and color penetration law, as well as the distribution changes of colors on fabrics during different batches of printing and dyeing. When generating simulated printing and dyeing effect images, the GAN-Autoencoder combination unit needs to consider factors such as the absorption characteristics of textile materials to dyes, so as to generate a color correction reference that is more in line with actual production. The DBN-RBM combination unit performs pre-training and feature mining on the data feature distribution unique to the textile printing and dyeing industry, such as learning the basic color presentation patterns of different dye combinations on different fabrics.

[0241] System architecture layer:

[0242] The task allocation of computing nodes in the distributed deep learning architecture is divided according to the type of textile products (such as cotton, silk, chemical fiber, etc.) and the type of printing and dyeing process (such as dip dyeing, pad dyeing, printing, etc.);

[0243] In addition to conventional information such as color deviation values, the feedback information of the multi-level feedback self-calibration architecture must also include fabric material information, printing and dyeing process identification, and dye batch information, so as to more accurately trace and adjust the source of color deviation;

[0244] The feature library storage data of the adaptive color calibration architecture is expanded to include textile fabric type characteristics (such as water absorption, color fastness, etc.), printing and dyeing equipment parameters (such as printing machine accuracy, dye tank capacity, etc.), and dye brand models and characteristics (such as solubility, dye uptake, etc.);

[0245] Data processing and optimization layer:

[0246] When synthesizing image data with specific color deviation characteristics, the data enhancement and synthesis module needs to build a model based on the physical and chemical principles of textile printing and dyeing, such as considering the color change law during the dye diffusion process;

[0247] The semantic segmentation data processing module focuses on dividing textile fabric images into pattern areas, background areas, and areas where different materials are stitched together, so as to analyze and correct the color management requirements of different areas.

[0248] The real-time data streaming processing and incremental learning modules need to adapt to the relatively slow data generation speed and long production cycle in the textile printing and dyeing production process, and adjust the calculation logic of the sliding window size and the cache data validity period.

[0249] The operation steps are as follows:

[0250] Sp1: Data collection and input: Obtain color data from textile fabric design images and online detection equipment in the printing and dyeing process, and read textile fabric, printing and dyeing equipment and dye feature library information at the same time;

[0251] Sp2: Data processing and optimization layer operations:

[0252] Sp2.1: The data enhancement and synthesis module expands and synthesizes the original image data combined with the printing and dyeing process parameters to generate a variety of simulated printing and dyeing data;

[0253] Sp2.2: The semantic segmentation data processing module divides the textile fabric image into regions, extracts the color features of each region and transmits them to the algorithm layer;

[0254] Sp2.3: The real-time data streaming processing and incremental learning module processes the real-time data of the printing and dyeing process, extracts features, and performs incremental learning and updating of the model based on the characteristics of the production cycle;

[0255] Sp3: Algorithm layer operations

[0256] Sp3.1: CNN-RNN combination unit analyzes fabric images and printing and dyeing process data, predicts color deviation patterns and generates correction strategy recommendations;

[0257] Sp3.2: GAN-Autoencoder combination unit generates printing and dyeing effect simulation images and collaboratively corrects color deviations to evaluate color consistency;

[0258] Sp3.3: The DBN-RBM combination unit mines the deep features of printing and dyeing data and provides accurate correction parameters.

[0259] Sp4: System architecture layer operations:

[0260] Sp4.1: Distributed deep learning architecture integrates data from each node, generates an overall color correction model and allocates resources among different textile printing and dyeing tasks;

[0261] Sp4.2: Multi-level feedback self-calibration architecture performs color detection and feedback before, during and after printing and dyeing, and adjusts model parameters and printing and dyeing process parameters;

[0262] Sp4.3: Adaptive color calibration architecture matches model configuration according to fabric, equipment and dye information, and adjusts model structure and parameters;

[0263] Sp5: Output and application: Output high-precision color-corrected printing and dyeing process parameters or fabric color control instructions, which are applied to the textile printing and dyeing production process to ensure product color consistency.

[0264] Digital advertising and display manufacturing industry

[0265] Application scenarios and industry needs: In the field of digital advertising, whether it is a large outdoor display or an indoor high-definition advertising screen, the accuracy and consistency of color directly affects the advertising effect and brand image. During the display manufacturing process, different panel materials, backlight technology, and driver chips will affect the display color. This system can be used to ensure the color calibration accuracy of the display during the production process, as well as the color consistency of advertising content under different ambient light conditions.

[0266] System architecture changes:

[0267] Algorithm layer:

[0268] The data input includes information such as the physical parameters of the display panel (such as resolution, color gamut, contrast, etc.), the spectral characteristics of the backlight source, and the color conversion function of the driver chip;

[0269] The CNN-RNN combination unit focuses on analyzing the color display change rules of video image sequences on different display devices, considering the impact of factors such as panel pixel arrangement and refresh rate on color deviation. The GAN-Autoencoder combination unit generates simulated images that meet the target advertising color standards and adapt to the characteristics of specific display screens. The autoencoder learns the potential representation of the display color to optimize the color correction effect. The DBN-RBM combination unit mines the features behind the complex data in the manufacturing and use of display screens, such as the impact of different backlight aging degrees on color;

[0270] System architecture layer:

[0271] The distributed deep learning architecture computing nodes divide tasks according to the display type (such as LCD, OLED, etc.) and application scenario (such as outdoor large screen, indoor small screen, etc.);

[0272] The feedback information of the multi-level feedback self-calibration architecture includes the display device number, ambient light sensor data, and the distribution of color deviation in different areas of the display (such as edge area and center area), so as to accurately adjust the color calibration parameters of the display;

[0273] The adaptive color calibration architecture feature library stores display panel supplier information, backlight model and characteristics, driver chip parameters, and color correction models under different ambient light conditions.

[0274] Data processing and optimization layer:

[0275] The data enhancement and synthesis module synthesizes simulated display data with different color deviations according to the display principle of the display screen and the characteristics of the advertising image, such as data simulating the change of color saturation in a low-brightness environment;

[0276] The semantic segmentation data processing module divides the advertising image into regions, such as the main character region, background region, and text region, and performs color deviation analysis based on the display characteristics of the display screen, for example, focusing on clarity and color contrast in the text region;

[0277] The real-time data streaming processing and incremental learning modules need to adapt to the high-speed data update requirements when the display screen plays videos or dynamic advertisements, optimize the data processing algorithm and model update mechanism, and ensure that a large amount of image frame data can be processed in a short time and the model is updated in a timely manner.

[0278] Run steps:

[0279] Sp1: Data collection and input: Obtain data from advertising image materials, display production test data, and ambient light sensors, and read display-related feature library information.

[0280] Sp2: Data processing and optimization layer operations

[0281] Sp2.1: The data enhancement and synthesis module processes the advertising image data in combination with the display characteristics, expands the data set and synthesizes simulated display deviation data.

[0282] Sp2.2: The semantic segmentation data processing module divides the advertising image into regions, extracts regional color features and transmits them to the algorithm layer.

[0283] Sp2.3: The real-time data streaming processing and incremental learning module processes the real-time image data during the display screen playback, extracts features and quickly updates the model.

[0284] Sp3: Algorithm layer operations

[0285] Sp3.1: The CNN-RNN combination unit analyzes the color display rules of the advertising image on the display screen, predicts the deviation and generates a correction strategy.

[0286] Sp3.2: The GAN-Autoencoder combination unit generates high-quality simulated display images, jointly corrects color deviations and evaluates consistency.

[0287] Sp3.3: The DBN-RBM combination unit mines the deep features of display data and provides accurate correction parameters.

[0288] Sp4: System architecture layer operations

[0289] Sp4.1: The distributed deep learning architecture integrates the data of each node, generates color correction models suitable for different display screens and assigns tasks.

[0290] Sp4.2: Multi-level feedback self-calibration architecture performs color detection and feedback during display production testing, installation and debugging, and use, and adjusts model and display parameters.

[0291] Sp4.3: The adaptive color calibration architecture matches the model configuration according to the display device information and ambient light conditions, and adjusts the model structure and parameters.

[0292] Sp5: Output and application: Output high-precision color-corrected display driving parameters or advertising image color adjustment instructions, which are applied to display manufacturing and advertising delivery processes to ensure color accuracy and consistency.

[0293] The above-mentioned entire system architecture, in actual use, can be applied to other industries that require color management and self-calibration, in addition to the printing industry, textile industry, advertising printing industry, etc. mentioned above. When in use, the data source of the system is adjusted according to the industry, and the overall architecture remains unchanged. The corresponding model and correction parameters are output according to the source data, so that the entire system can be adapted to multiple industry fields, increasing adaptability during use.

[0294] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprising a reference structure" do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0295] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. Deep learning self-calibration system for high-precision color management, characterized by: The system architecture includes an algorithm layer, a system architecture layer, and a data processing and optimization layer, wherein: The algorithm layer is used to perform deep feature extraction, deviation prediction and correction strategy generation on the color of the printed image. The algorithm layer includes a multi-algorithm fusion model, which includes a convolutional neural network CNN and a recurrent neural network RNN ​​combination unit, a fusion generative adversarial network GAN and an autoencoder combination unit, and a deep belief network DBN and a restricted Boltzmann machine RBM combination unit; The system architecture layer is used to build an architecture system that adapts to large-scale production, multiple links and multiple equipment and materials in the printing industry, including a distributed deep learning architecture, a multi-level feedback self-calibration architecture and an adaptive color calibration architecture. The distributed deep learning architecture distributes the training tasks of the deep learning model to multiple computing nodes for parallel processing. Different nodes focus on data processing of different printing product lines or different color spaces. The central coordinator integrates the results of each node. After the multi-level feedback self-calibration architecture performs preliminary correction on the image color in the pre-press stage, the correction result is input into the system again in the printing stage for secondary detection and analysis. The deep learning model parameters or correction strategies are adjusted according to feedback information at different levels, and a feedback channel is established from the post-press quality inspection stage to the pre-press design and printing stages. The adaptive color calibration architecture uses a pre-established printing equipment and material feature library to quickly identify and match the corresponding model configuration during printing production and automatically adjust the structure and parameters of the deep learning model. The data processing and optimization layer is used to preprocess, expand and dynamically update the printed color data. The data processing and optimization layer includes a data enhancement and synthesis module, a semantic segmentation data processing module and a real-time data streaming processing and incremental learning module. The data enhancement and synthesis module expands the training data set by performing transformation operations such as rotation, scaling, contrast adjustment, etc. on the existing printed color images and artificially synthesizing image data with specific color deviation characteristics. The semantic segmentation data processing module uses semantic segmentation technology to divide different areas in the printed image according to color categories, object categories or functional areas, and performs color deviation analysis and correction respectively; the real-time data streaming processing and incremental learning module quickly processes and extracts features of the color data generated in real time during the printing production process, and uses incremental learning technology to adapt the model to the new printing data distribution and color change trends.

2. The deep learning self-calibration system for high-precision color management according to claim 1, characterized in that: The convolutional neural network CNN and recurrent neural network RNN ​​combination unit is used to deeply extract the color pixel distribution, pattern texture and other features of the printed image, and analyze the change law of color deviation in the entire printed page and between different printing batches based on the spatial and temporal dependencies between the printed image sequences. In the fusion generative adversarial network GAN and autoencoder combination unit, the GAN generator generates printed image simulation data similar to the target printing color standard according to the target printing color standard, and the discriminator distinguishes between the real printed image and the generated image. The adversarial training of the two improves the quality of the generated image. The autoencoder compresses and reconstructs the input printed color image and learns the potential color representation of the image. The two work together to correct the printed color deviation and ensure color consistency. The deep belief network DBN and restricted Boltzmann machine RBM combination unit performs unsupervised pre-training on the printed color data through RBM, learns the basic feature distribution and structure of the printed color data, and then stacks it to construct a DBN, thereby mining high-level abstract features hidden in complex printed color data, providing an accurate basis for subsequent automatic correction.

3. The deep learning self-calibration system for high-precision color management according to claim 1, characterized in that: Each computing node in the distributed deep learning architecture has an independent data storage unit and a computing unit. The data storage unit is used to temporarily store the assigned specific printing color data subset. The computing unit performs local training operations of the deep learning model based on the stored data. The computing nodes are connected by high-speed data transmission channels to ensure the efficiency of data interaction and model integration.

4. The deep learning self-calibration system for high-precision color management according to claim 1, characterized in that: The feedback information in the multi-level feedback self-calibration architecture includes color deviation values, deviation distribution areas, color space coordinate differences, and corresponding printing link identification information, wherein the color deviation values ​​are accurate to three decimal places, the deviation distribution areas are represented by image pixel coordinate ranges or printed product physical size areas, the color space coordinate differences are quantitatively described based on color spaces commonly used in the printing industry, such as CMYK or LAB, and the printing link identification information clearly distinguishes the sources of pre-press, mid-press, and post-press links.

5. The deep learning self-calibration system for high-precision color management according to claim 1, characterized in that: The data stored in the printing equipment and material feature library in the adaptive color calibration architecture covers printer brand and model, printer mechanical parameters, paper type, ink brand and model and characteristic information, and the data is stored in the form of a structured database to facilitate rapid retrieval and model configuration matching.

6. The deep learning self-calibration system for high-precision color management according to claim 1, characterized in that: The color data synthesis algorithm in the data enhancement and synthesis module is constructed based on the printing color physical model and the mathematical statistical model. When synthesizing image data with specific color deviation characteristics, it can accurately generate simulated images according to the deviation parameters set by the user. The generated image data format is consistent with the real printing image data format and can be directly used for deep learning model training.

7. The deep learning self-calibration system for high-precision color management according to claim 1, characterized in that: The semantic segmentation technology in the semantic segmentation data processing module adopts a deep learning semantic segmentation model. The model structure is improved based on the U-Net architecture, and a multi-scale feature fusion module is added. It can accurately identify text, image body, background and other areas in printed images at different resolutions. The segmentation accuracy reaches the pixel level and has good adaptability to different types of printed images.

8. The deep learning self-calibration system for high-precision color management according to claim 1, characterized in that: The data streaming processing algorithm in the real-time data streaming processing and incremental learning module adopts sliding window technology and distributed caching mechanism. The sliding window size is dynamically adjusted according to the printing data flow and processing speed to ensure that data is not lost and can be processed in time during data flow peaks. The distributed caching mechanism caches recently processed data in multiple nodes to facilitate the incremental learning module to quickly obtain data for model updates. The validity period of the cached data is dynamically set according to the continuity of the printing task and the timeliness of the data.

Images