An intelligent pre-print image optimization system and method based on deep learning

Abstract

The present invention belongs to the field of computer vision technology, and discloses an intelligent pre-print image optimization system and method based on deep learning; a data collection module is used to collect image parameter data, pre-print device parameter data and user behavior data; a data processing module is used to pre-process the acquired image parameter data, pre-print device parameter data and user behavior data to obtain an image feature data set, a pre-print device feature data set and a user behavior feature data set; the influence of the user behavior feature data set on the image feature data set and the pre-print device feature data set is evaluated to obtain an influence feature data set; an image error prediction module is used to obtain a pre-print image error prediction model based on the influence feature data set, and the pre-print image error is predicted by the pre-print image error prediction model; through accurate error prediction, the system can provide optimization suggestions for users, thereby improving printing quality and efficiency.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and more specifically, to an intelligent pre-print image optimization system and method based on deep learning. Background Art

[0002] Patent application publication number CN113256170A discloses a graphics processing device and method for paper preprinting. Based on the varying suitability of various printing methods in the preprinting process in terms of resolution, color, format, and surface material, the present invention matches the preprinted image with the online corrugated preprinting production terminal. Preprinted graphics processing is then performed based on the properties of the corrugated preprinting terminal. This invention enables automated and efficient graphics processing in the corrugated preprinting process, optimizing resources and time.

[0003] Existing pre-print image optimization systems and methods have the following major problems:

[0004] Failure to consider the sliding window mechanism may cause the model to focus only on local features and fail to effectively capture global information; it will affect the understanding of the overall structure of the image and may lead to insufficient processing of complex scenes; the window size is fixed, and the model's ability to adapt to images with different features will be reduced, especially when processing diverse image datasets, which may lead to poor performance; unreasonable window division may lead to waste of computing resources, especially when processing large images. Failure to dynamically adjust the number of windows may make the model too dispersed in local feature extraction and unable to concentrate resources on extracting important features; the lack of dynamic adaptation to image features may lead to insufficient generalization ability of the model when facing new images, making it difficult to effectively process unseen image types;

[0005] The lack of comprehensive consideration of key parameters may lead to inaccurate image defect assessment results, affecting subsequent processing decisions; the inability to quantify the severity of defects may lead to unreasonable priority decisions of the processing system, and the most significant defects may be ignored, affecting the final image quality; the lack of clear defect assessment standards may lead to computing resources and time being wasted on processing unimportant image characteristics, while ignoring defects that really need attention; the failure to effectively capture the full picture of image defects will reduce the overall efficiency of image processing, affecting the response speed and processing capabilities of the processing system.

[0006] In view of this, the present invention proposes an intelligent pre-print image optimization system and method based on deep learning to solve the above problems. Summary of the Invention

[0007] In order to overcome the above-mentioned defects of the prior art and achieve the above-mentioned objectives, the present invention provides the following technical solutions: an intelligent pre-print image optimization system and method based on deep learning, comprising:

[0008] A data collection module, used to collect image parameter data, pre-printing device parameter data and user behavior data;

[0009] a data processing module, configured to pre-process the acquired image parameter data, pre-printing device parameter data, and user behavior data to obtain an image feature dataset, a pre-printing device feature dataset, and a user behavior feature dataset; and to evaluate the impact of the user behavior feature dataset on the image feature dataset and the pre-printing device feature dataset to obtain an impact feature dataset;

[0010] An image error prediction module is used to obtain a pre-print image error prediction model based on the influencing feature data set training, and to predict the pre-print image error through the pre-print image error prediction model;

[0011] A quality assessment module is used to compare the predicted pre-print image error with a preset pre-print image error threshold to determine whether the pre-print image quality meets the standard;

[0012] The optimization evaluation module: if the pre-printed image quality meets the standards, there is no need to optimize the pre-printed image, and a printing instruction is directly issued through the intelligent pre-printing terminal to print the pre-printed image; if the pre-printed image quality does not meet the standards, the pre-printed image needs to be optimized, and an optimization instruction is issued through the intelligent pre-printing terminal, and the pre-printed image data that needs to be optimized is automatically collected;

[0013] The optimization execution module is used to train and obtain an image defect detection model based on the pre-printed image data that needs to be optimized, and obtain the pre-printed image defect type based on the image defect detection model; obtain the severity of the pre-printed image defect type through the pre-printed image defect degree assessment model, and execute the corresponding image optimization plan according to the severity of the pre-printed image defect type; each module is connected by wired and / or wireless means.

[0014] Furthermore, the image parameter data includes image attribute data, image format data, color calibration data and image noise data; the image attribute data includes image resolution data, color space data, bit depth, compression rate, metadata, image brightness data and contrast data;

[0015] Pre-printing device parameter data includes printing resolution, scanning resolution, color mode parameters, and paper type and size parameters; user behavior data includes image adjustment behavior and user preference data;

[0016] Pre-printed device parameter data and user behavior data are acquired through embedded sensors and user terminals. Image parameter data is acquired through the Swin Transformer. After a multi-head self-attention mechanism, a sliding window mechanism, and patch merging operations, a feature matrix is ​​ultimately output. The final output feature matrix is ​​further analyzed and extracted to obtain image parameter data. The specific method is as follows:

[0017] S21, input original pre-printed image I∈R H×W×C , where I is the original pre-printed image, R is the real space where the original pre-printed image is located, H is the image height, W is the image width, and C is the number of channels; the original pre-printed image is divided into q non-overlapping image blocks of size n×n, each image block contains n 2 ×C pixels; the pixels in each image patch are flattened to form a feature vector; where n is the size of the image patch, and the value of n is adjusted by the size adjustment formula;

[0018] The size adjustment formula is specifically:

[0019] Where Pmin is the minimum number of pixels contained in each image block; k is a control factor that controls the number of image blocks in the horizontal and vertical directions, and is usually greater than or equal to 1 to ensure that the image can be fully covered; is the rounding function, which means rounding down to ensure that n is an integer;

[0020] S22. Apply linear projection to each image patch, projecting it from the flattened feature vector to a higher-dimensional feature space to obtain a feature vector of dimension d; integrate the feature vectors after the projection of all image patches to form a matrix X∈R d ;

[0021] S23. Use Swin Transformer Block to process image patches and capture local dependency information through the local window attention mechanism. Divide the image patch into V windows of size M×M. For each image patch in the window, use the multi-head self-attention mechanism to capture local features.

[0022] The number V of image windows is adjusted by the window number adjustment formula, which is: Where q is the number of image patches; n is the size of the image patches;

[0023] Patch Merging is used to perform downsampling and merge adjacent blocks; a sliding window mechanism is used to capture the global features of small image blocks; the window is slidingly displaced, and the sliding operation of each window in the tth layer is S t, then the window feature after sliding is X′=S t (X); After the above steps, the final output feature matrix X∈RH′×W′×d′; where H′ and W′ are the spatial dimensions after multiple patch merging, and d′ is the number of channels;

[0024] S24. Further analyze and extract the feature matrix to obtain image parameter data.

[0025] Furthermore, the method of preprocessing the acquired image parameter data, preprinting device parameter data, and user behavior data to obtain an image feature dataset, a preprinting device feature dataset, and a user behavior feature dataset includes:

[0026] The acquired image parameter data, pre-printing device parameter data, and user behavior data were cleaned, interpolation was used to fill missing values, and clustering algorithms were used to detect and remove outliers. Z-score standardization was used to transform the image parameter data, pre-printing device parameter data, and user behavior data into a standard normal distribution with a mean of 0 and a standard deviation of 1 to eliminate dimensionality effects.

[0027] After the preprocessing in the above steps, the image feature dataset, pre-printed device feature dataset and user behavior feature dataset are obtained.

[0028] Furthermore, the method of evaluating the impact of the user behavior feature dataset on the image feature dataset and the pre-printed device feature dataset to obtain the impact feature dataset includes:

[0029] Taking image features, pre-printed device features and user behavior features as input variables, a joint regression model is constructed. By constructing the joint regression model, the impact of the user behavior feature dataset on the image feature dataset and the pre-printed device feature dataset is analyzed, and finally the impact feature dataset is obtained.

[0030] Furthermore, the training method of the pre-printed image error prediction model includes:

[0031] The dataset is divided into training set, validation set and test set to train the model and evaluate the model performance; the sample set is a subset of the dataset, and each sample set includes a historical impact feature dataset and the corresponding preprint image error;

[0032] Constructing a pre-print image error prediction model, which includes an input layer, a convolution layer, a pooling layer, a batch normalization layer, a fully connected layer, and an output layer; the input layer is used to input a historical impact feature dataset, and the output layer is used to output the pre-print image error; the pre-print image error prediction model is a 3D convolutional neural network model;

[0033] The mean absolute error (MAE) loss function is used to measure the error between the model's predicted value and the actual value. The model is trained using the training set data and the Adam optimizer is used to minimize the loss function. The model performance is evaluated using the validation set, and the model's hyperparameters are tuned until the model performance no longer improves significantly or the preset stopping condition is met.

[0034] The test set is used to evaluate the performance of the model in the prediction task, and the current influencing feature dataset is input into the trained pre-print image error prediction model to obtain the pre-print image error.

[0035] Furthermore, the method of comparing the predicted pre-print image error with a preset pre-print image error threshold to determine whether the pre-print image quality meets the standard includes:

[0036] If the predicted pre-print image error is less than the preset pre-print image error threshold, the pre-print image quality is determined to be up to standard;

[0037] If the predicted pre-print image error is greater than or equal to a preset pre-print image error threshold, it is determined that the pre-print image quality does not meet the standard.

[0038] Furthermore, the pre-printed image data that needs to be optimized includes image resolution, color space, image format, color calibration rate, image noise, image brightness and contrast.

[0039] Furthermore, the training method of the image defect detection model includes:

[0040] The dataset is divided into a training set, a validation set, and a test set to train the model and evaluate its performance. The sample set is a subset of the dataset, and each sample set includes historical pre-printed image data that needs to be optimized and the corresponding pre-printed image defect type.

[0041] The model's input data is historical pre-printed image data that needs to be optimized; the model's output label is the pre-printed image defect type; the image defect detection model consists of v decision trees. Bootstrap sampling is used to train each decision tree by randomly selecting a subset from the training set to train the tree, and randomly selecting a subset of features to split nodes until a stopping condition is met; the image defect detection model is a random forest model;

[0042] Use multi-class cross entropy as the model's loss function to measure the difference between the model's predicted value and the actual value. Use the training set to train the image defect detection model and use the backpropagation algorithm to update the model parameters to minimize the loss function. Use the validation set to evaluate the performance of the image defect detection model by calculating the precision metric.

[0043] Using the grid search hyperparameter optimization method, the hyperparameters of the image defect detection model are tuned based on the performance feedback of the validation set. The model hyperparameter combination is adjusted until the model performance no longer improves significantly or the preset stopping condition is reached.

[0044] The test set is used to evaluate the performance of the model in the prediction task, and the trained image defect detection model is used to predict the pre-printed image defect data that currently needs to be optimized to obtain the pre-printed image defect type.

[0045] Furthermore, the method of executing a corresponding image optimization solution according to the severity of the pre-printed image defect type includes:

[0046] Construct a pre-printed image defect degree assessment model, calculate the severity of the pre-printed image defect type through the pre-printed image defect degree assessment model, and execute the corresponding image optimization solution according to the severity of the pre-printed image defect type; the pre-printed image defect degree assessment model is:

[0047]

[0048] Where Sge is the severity of the pre-printed image defect type; R re is the image resolution; C lr is the color space; Fot is the image format; C lb is the color calibration rate; N se is the image noise; L bs is the image brightness; Dst is the contrast; ω1, ω2, ω3, ω4, ω5, ω6, and ω7 are the weight factors of each parameter.

[0049] A deep learning-based intelligent pre-print image optimization method, comprising:

[0050] S1. Collect image parameter data, pre-printing device parameter data and user behavior data;

[0051] S2. Preprocess the acquired image parameter data, pre-printing device parameter data, and user behavior data to obtain an image feature dataset, a pre-printing device feature dataset, and a user behavior feature dataset; evaluate the impact of the user behavior feature dataset on the image feature dataset and the pre-printing device feature dataset to obtain an impact feature dataset;

[0052] S3. Obtain a pre-print image error prediction model based on the influencing feature dataset, and predict the pre-print image error using the pre-print image error prediction model;

[0053] S4, comparing the predicted pre-printed image error with a preset pre-printed image error threshold to determine whether the pre-printed image quality meets the standard;

[0054] S5. If the pre-printed image quality meets the standards, there is no need to optimize the pre-printed image, and a printing instruction is directly issued through the intelligent pre-printing terminal to print the pre-printed image. If the pre-printed image quality does not meet the standards, the pre-printed image needs to be optimized, and an optimization instruction is issued through the intelligent pre-printing terminal, and the pre-printed image data that needs to be optimized is automatically collected.

[0055] S6. An image defect detection model is obtained by training the pre-printed image data to be optimized as needed, and the pre-printed image defect type is obtained based on the image defect detection model; the severity of the pre-printed image defect type is obtained through a pre-printed image defect degree assessment model, and a corresponding image optimization solution is executed according to the severity of the pre-printed image defect type.

[0056] The technical effects and advantages of the intelligent pre-printed image optimization system and method based on deep learning of the present invention are as follows:

[0057] The present invention can dynamically adjust the size of image patches according to different image features through a size adjustment formula; adaptive partitioning can not only effectively process images of different resolutions and sizes, but also ensure that the images can be fully covered to avoid information loss; at the same time, the size of the image patches is dynamically adjusted according to the control factor, ensuring flexible processing of different images; the sliding window mechanism can not only obtain local information, but also capture the global features of the image through the sliding operation of the window; enhance the ability to understand the overall image and avoid the problem of global information loss that may be caused by the local attention mechanism; the sliding window mechanism can also shift at different levels, capturing more global features layer by layer, further improving the global perception ability of the model; by adaptively adjusting the number of windows, the model's attention on local and global features can be reasonably distributed;

[0058] The pre-printed image defect severity assessment model can quantify the severity of pre-printed image defect types. The model combines multiple key parameters that affect image quality, such as resolution, color space, image format, color calibration rate, image noise, brightness, and contrast. It can effectively capture the full picture of image defects and ensure that the assessment results are highly accurate and reliable. By quantifying the severity of defects, the model can help the image processing system prioritize the most significant defect problems, avoiding unnecessary global optimization. It not only improves the efficiency of image processing, but also reduces the waste of computing resources that are not optimized for other image characteristics, saving time and computing overhead. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a schematic diagram of the structure of an intelligent pre-printed image optimization system based on deep learning of the present invention;

[0060] Figure 2 This is a flow chart of an intelligent pre-printed image optimization method based on deep learning according to the present invention. DETAILED DESCRIPTION

[0061] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0062] Example 1

[0063] See also Figure 1 As shown, the intelligent pre-print image optimization system and method based on deep learning described in this embodiment include:

[0064] A data collection module, used to collect image parameter data, pre-printing device parameter data and user behavior data;

[0065] a data processing module, configured to pre-process the acquired image parameter data, pre-printing device parameter data, and user behavior data to obtain an image feature dataset, a pre-printing device feature dataset, and a user behavior feature dataset; and to evaluate the impact of the user behavior feature dataset on the image feature dataset and the pre-printing device feature dataset to obtain an impact feature dataset;

[0066] An image error prediction module is used to obtain a pre-print image error prediction model based on the influencing feature data set training, and to predict the pre-print image error through the pre-print image error prediction model;

[0067] A quality assessment module is used to compare the predicted pre-print image error with a preset pre-print image error threshold to determine whether the pre-print image quality meets the standard;

[0068] The optimization evaluation module: if the pre-printed image quality meets the standards, there is no need to optimize the pre-printed image, and a printing instruction is directly issued through the intelligent pre-printing terminal to print the pre-printed image; if the pre-printed image quality does not meet the standards, the pre-printed image needs to be optimized, and an optimization instruction is issued through the intelligent pre-printing terminal, and the pre-printed image data that needs to be optimized is automatically collected;

[0069] The optimization execution module is used to train and obtain an image defect detection model based on the pre-printed image data that needs to be optimized, and obtain the pre-printed image defect type based on the image defect detection model; obtain the severity of the pre-printed image defect type through the pre-printed image defect degree assessment model, and execute the corresponding image optimization plan according to the severity of the pre-printed image defect type; each module is connected by wired and / or wireless means.

[0070] Image parameter data includes image attribute data, image format data, color calibration data, and image noise data; image attribute data includes image resolution data, color space data, bit depth, compression rate, metadata, image brightness data, and contrast data;

[0071] Pre-printing device parameter data includes printing resolution, scanning resolution, color mode parameters, and paper type and size parameters; user behavior data includes image adjustment behavior and user preference data;

[0072] Pre-printed device parameter data and user behavior data are acquired through embedded sensors and user terminals; image parameter data is acquired through the Swin Transformer; after the multi-head self-attention mechanism, sliding window mechanism, and patch merging operations, the feature matrix is ​​finally output; the final output feature matrix is ​​further analyzed and extracted to obtain image parameter data. The specific method is as follows:

[0073] S21, input original pre-printed image I∈R H×W×C , where I is the original pre-printed image, R is the real space where the original pre-printed image is located, H is the image height, W is the image width, and C is the number of channels; the original pre-printed image is divided into q non-overlapping image blocks of size n×n, each image block contains n 2 ×C pixels; the pixels in each image patch are flattened to form a feature vector; where n is the size of the image patch, and the value of n is adjusted by the size adjustment formula;

[0074] The sizing formula is:

[0075] Where Pmin is the minimum number of pixels contained in each image block; k is a control factor that controls the number of image blocks in the horizontal and vertical directions, and is usually greater than or equal to 1 to ensure that the image can be fully covered; is the rounding function, which means rounding down to ensure that n is an integer;

[0076] For example, assuming the original image size is H=1000, W=800, C=3; each image block must contain at least 256 pixels, that is, P min =256; adjustment factor k=2; then

[0077]

[0078] Compare the two values ​​and take the smaller one as the value of n, then n=min(16,400)=16; since W<H,

[0079] In summary, the original pre-printed image is divided into 63 image patches of size 16×16;

[0080] S22. Apply linear projection to each image patch, projecting it from the flattened feature vector to a higher-dimensional feature space to obtain a feature vector of dimension d; integrate the feature vectors after the projection of all image patches to form a matrix X∈R d ;

[0081] S23. Use Swin Transformer Block to process image patches and capture local dependency information through the local window attention mechanism. Divide the image patch into V windows of size M×M. For each image patch in the window, use the multi-head self-attention mechanism to capture local features.

[0082] The number of image windows V is adjusted by the window number adjustment formula, which is: Where q is the number of image patches; n is the size of the image patches;

[0083] For example, when the original pre-printed image is divided into 63 image blocks of size 16×16, the number of image windows is:

[0084] Patch Merging is used to perform downsampling and merge adjacent blocks; a sliding window mechanism is used to capture the global features of small image blocks; the window is slidingly displaced, and the sliding operation of each window in the tth layer is S t , then the window feature after sliding is X′=S t (X); After the above steps, the final output feature matrix X∈RH′×W′×d′; where H′ and W′ are the spatial dimensions after multiple patch merging, and d′ is the number of channels;

[0085] S24, further analyzing and extracting the feature matrix to obtain image parameter data;

[0086] Capturing dependencies over a wider range through a sliding window mechanism across windows, thereby extracting global parameter data such as tonal balance, overall lighting, and texture variations in complex images;

[0087] For example, in even layers, the window is slightly moved by a number of pixels in the horizontal and vertical directions, so that the small image blocks at the previous window boundary participate in the new window attention calculation, and the features across windows are fused, thereby capturing the global long-range dependency information in the image.

[0088] The method of preprocessing the acquired image parameter data, pre-printing device parameter data, and user behavior data to obtain an image feature dataset, a pre-printing device feature dataset, and a user behavior feature dataset includes:

[0089] The acquired image parameter data, pre-printing device parameter data, and user behavior data were cleaned, interpolation was used to fill missing values, and clustering algorithms were used to detect and remove outliers. Z-score standardization was used to transform the image parameter data, pre-printing device parameter data, and user behavior data into a standard normal distribution with a mean of 0 and a standard deviation of 1 to eliminate dimensionality effects.

[0090] After the preprocessing in the above steps, the image feature dataset, pre-printed device feature dataset and user behavior feature dataset are obtained.

[0091] The method for evaluating the impact of the user behavior feature dataset on the image feature dataset and the pre-printed device feature dataset to obtain the impact feature dataset includes:

[0092] Taking image features, pre-printed device features and user behavior features as input variables, a joint regression model is constructed. By constructing the joint regression model, the impact of the user behavior feature dataset on the image feature dataset and the pre-printed device feature dataset is analyzed, and finally the impact feature dataset is obtained.

[0093] The training method of the pre-print image error prediction model includes:

[0094] The dataset is divided into training set, validation set and test set to train the model and evaluate the model performance; the sample set is a subset of the dataset, and each sample set includes a historical impact feature dataset and the corresponding preprint image error;

[0095] Constructing a pre-print image error prediction model, which includes an input layer, a convolution layer, a pooling layer, a batch normalization layer, a fully connected layer, and an output layer; the input layer is used to input a historical impact feature dataset, and the output layer is used to output the pre-print image error; the pre-print image error prediction model is a 3D convolutional neural network model;

[0096] The mean absolute error is used as the loss function to measure the error between the model's predicted value and the actual value; the mean absolute error loss function is: Wherein, MAE is the mean absolute error; n′ is the number of datasets; yi is the true value of the error of the i-th preprinted image; is the predicted value of the error of the i-th pre-printed image;

[0097] Use the training set data to train the model and minimize the loss function using the Adam optimizer. Use the validation set to evaluate the model's performance and tune the model's hyperparameters until the model performance no longer improves significantly or reaches the preset stopping condition.

[0098] The test set is used to evaluate the performance of the model in the prediction task, and the current influencing feature dataset is input into the trained pre-print image error prediction model to obtain the pre-print image error.

[0099] The method of comparing the predicted pre-print image error with a preset pre-print image error threshold to determine whether the pre-print image quality meets the standard includes:

[0100] If the predicted pre-print image error is less than the preset pre-print image error threshold, the pre-print image quality is determined to be up to standard;

[0101] If the predicted pre-print image error is greater than or equal to a preset pre-print image error threshold, it is determined that the pre-print image quality does not meet the standard.

[0102] Pre-print image data that needs to be optimized includes image resolution, color space, image format, color calibration rate, image noise, image brightness and contrast.

[0103] The training methods for image defect detection models include:

[0104] The dataset is divided into a training set, a validation set, and a test set to train the model and evaluate its performance. The sample set is a subset of the dataset, and each sample set includes historical pre-printed image data that needs to be optimized and the corresponding pre-printed image defect type.

[0105] The model's input data is historical pre-printed image data that needs to be optimized; the model's output label is the pre-printed image defect type; the image defect detection model consists of v decision trees. Bootstrap sampling is used to train each decision tree by randomly selecting a subset from the training set to train the tree, and randomly selecting a subset of features to split nodes until a stopping condition is met; the image defect detection model is a random forest model;

[0106] Use multi-class cross entropy as the loss function of the model to measure the difference between the model's predicted value and the actual value; the multi-class cross entropy loss function is:

[0107] Where L is the average loss of the dataset; N is the total number of samples in the dataset; B is the number of preprint image optimization types; p i′b′ Optimize the true label of the type for the b′th preprinted image for the i′th sample; p i′b′ The probability that the i′th sample predicted by the model belongs to the optimized type of the b′th pre-printed image;

[0108] Use the training set to train the image defect detection model and use the backpropagation algorithm to update the model parameters to minimize the loss function. Use the validation set to evaluate the performance of the image defect detection model by calculating the precision metric.

[0109] Using the grid search hyperparameter optimization method, the hyperparameters of the image defect detection model are tuned based on the performance feedback of the validation set. The model hyperparameter combination is adjusted until the model performance no longer improves significantly or the preset stopping condition is reached.

[0110] The test set is used to evaluate the performance of the model in the prediction task, and the trained image defect detection model is used to predict the pre-printed image defect data that currently needs to be optimized to obtain the pre-printed image defect type.

[0111] The method for executing a corresponding image optimization solution according to the severity of the pre-printed image defect type includes:

[0112] A pre-printed image defect assessment model is constructed. The severity of the pre-printed image defect type is calculated using the pre-printed image defect assessment model, and the corresponding image optimization solution is executed according to the severity of the pre-printed image defect type. The pre-printed image defect assessment model is:

[0113]

[0114] Where Sge is the severity of the pre-printed image defect type; R re is the image resolution; C lr is the color space; Fot is the image format; C lb is the color calibration rate; N se is the image noise; L bs is the image brightness; Dst is the contrast; ω1, ω2, ω3, ω4, ω5, ω6, and ω7 are the weight factors of each parameter.

[0115] For example, the image resolution R re 300dpi, color space C lr is 90, the image format Fot is 1.2 (the weight of a certain format), and the color calibration rate C lb is 0.8, image noise N se is 50, image brightness L bs is 100, contrast Dst is 1.5, ω1 is 0.5, ω2 is 0.3, ω3 is 0.7, ω4 is 0.4, ω5 is 0.6, ω6 is 0.2, and ω7 is 0.9. The severity of the pre-printed image defect type is

[0116] The preset pre-print image error threshold is set by a staff member, and different pre-print image errors are collected through an intelligent pre-print terminal, and an average value of multiple pre-print image errors is taken as the preset pre-print image error threshold.

[0117] In this embodiment, the size of the image patch can be dynamically adjusted according to the different features of the image through the resizing formula; the adaptive partitioning can not only effectively process images of different resolutions and sizes, but also ensure that the image can be fully covered to avoid information loss; at the same time, the size of the image patch is dynamically adjusted according to the control factor, ensuring flexible processing of different images; the sliding window mechanism can not only obtain local information, but also capture the global features of the image through the sliding operation of the window, thereby enhancing the ability to understand the overall image and avoiding the problem of global information loss that may be caused by the local attention mechanism; the sliding window mechanism can also shift at different levels, capturing more global features layer by layer, further improving the global perception ability of the model; by adaptively adjusting the number of windows, the model's attention on local and global features can be reasonably distributed;

[0118] The pre-printed image defect severity assessment model can quantify the severity of pre-printed image defect types. The model combines multiple key parameters that affect image quality, such as resolution, color space, image format, color calibration rate, image noise, brightness, and contrast. It can effectively capture the full picture of image defects and ensure that the assessment results are highly accurate and reliable. By quantifying the severity of defects, the model can help the image processing system prioritize the most significant defect problems, avoiding unnecessary global optimization. It not only improves the efficiency of image processing, but also reduces the waste of computing resources that are not optimized for other image characteristics, saving time and computing overhead.

[0119] Example 2

[0120] See also Figure 2 As shown, for the parts not described in detail in this embodiment, please refer to the description of Example 1. A method for intelligent pre-printed image optimization based on deep learning is provided, comprising:

[0121] S1. Collect image parameter data, pre-printing device parameter data and user behavior data;

[0122] S2. Preprocess the acquired image parameter data, pre-printing device parameter data, and user behavior data to obtain an image feature dataset, a pre-printing device feature dataset, and a user behavior feature dataset; evaluate the impact of the user behavior feature dataset on the image feature dataset and the pre-printing device feature dataset to obtain an impact feature dataset;

[0123] S3. Obtain a pre-print image error prediction model based on the influencing feature dataset, and predict the pre-print image error using the pre-print image error prediction model;

[0124] S4, comparing the predicted pre-printed image error with a preset pre-printed image error threshold to determine whether the pre-printed image quality meets the standard;

[0125] S5. If the pre-printed image quality meets the standards, there is no need to optimize the pre-printed image, and a printing instruction is directly issued through the intelligent pre-printing terminal to print the pre-printed image. If the pre-printed image quality does not meet the standards, the pre-printed image needs to be optimized, and an optimization instruction is issued through the intelligent pre-printing terminal, and the pre-printed image data that needs to be optimized is automatically collected.

[0126] S6. An image defect detection model is obtained by training the pre-printed image data to be optimized as needed, and the pre-printed image defect type is obtained based on the image defect detection model; the severity of the pre-printed image defect type is obtained through a pre-printed image defect degree assessment model, and a corresponding image optimization solution is executed according to the severity of the pre-printed image defect type.

[0127] Since the electronic device introduced in this embodiment is an electronic device used to implement a deep learning-based intelligent pre-printed image optimization system and method in the embodiment of this application, based on the deep learning-based intelligent pre-printed image optimization system and method introduced in the embodiment of this application, those skilled in the art can understand the specific implementation of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of this application will not be described in detail here. As long as those skilled in the art implement the electronic device used in the deep learning-based intelligent pre-printed image optimization system and method in the embodiment of this application, they all fall within the scope of protection of this application.

[0128] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field according to actual conditions.

[0129] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that for users of ordinary skill in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An intelligent pre-print image optimization system based on deep learning, characterized in that: include: A data collection module, used to collect image parameter data, pre-printing device parameter data and user behavior data; The image parameter data includes image attribute data, image format data, color calibration data and image noise data; the image attribute data includes image resolution data, color space data, bit depth, compression rate, metadata, image brightness data and contrast data; Pre-printing device parameter data includes printing resolution, scanning resolution, color mode parameters, and paper type and size parameters; user behavior data includes image adjustment behavior and user preference data; Pre-printed device parameter data and user behavior data are obtained through embedded sensors and user terminals; Image parameter data is obtained through the Swin Transformer; after the multi-head self-attention mechanism, sliding window mechanism and PatchMerging operation, the feature matrix is ​​finally output; Further analyze and extract the final output feature matrix to obtain image parameter data; a data processing module, configured to pre-process the acquired image parameter data, pre-printing device parameter data, and user behavior data to obtain an image feature dataset, a pre-printing device feature dataset, and a user behavior feature dataset; and to evaluate the impact of the user behavior feature dataset on the image feature dataset and the pre-printing device feature dataset to obtain an impact feature dataset; An image error prediction module is used to obtain a pre-print image error prediction model based on the influencing feature data set training, and to predict the pre-print image error through the pre-print image error prediction model; A quality assessment module is used to compare the predicted pre-print image error with a preset pre-print image error threshold to determine whether the pre-print image quality meets the standard; The optimization evaluation module: if the pre-printed image quality meets the standards, there is no need to optimize the pre-printed image, and a printing instruction is directly issued through the intelligent pre-printing terminal to print the pre-printed image; if the pre-printed image quality does not meet the standards, the pre-printed image needs to be optimized, and an optimization instruction is issued through the intelligent pre-printing terminal, and the pre-printed image data that needs to be optimized is automatically collected; An optimization execution module is used to train and obtain an image defect detection model based on the optimized pre-printed image data, and obtain the pre-printed image defect type based on the image defect detection model; The severity of the pre-printed image defect type is obtained through the pre-printed image defect degree assessment model, and the corresponding image optimization plan is executed according to the severity of the pre-printed image defect type; each module is connected through wired and / or wireless means.

2. The deep learning-based intelligent pre-print image optimization system according to claim 1, characterized in that: The method for obtaining image parameter data further includes: S21, input original pre-printed image I∈R H×W×C , where I is the original pre-printed image, R is the real space where the original pre-printed image is located, H is the image height, W is the image width, and C is the number of channels; the original pre-printed image is divided into q non-overlapping image blocks of size n×n, each image block contains n 2 ×C pixels; the pixels in each image patch are flattened to form a feature vector; where n is the size of the image patch, and the value of n is adjusted by the size adjustment formula; The size adjustment formula is specifically: Among them, P min The minimum number of pixels contained in each image block; k is a control factor that controls the number of image blocks in the horizontal and vertical directions, and is usually greater than or equal to 1 to ensure that the image can be completely covered; is the rounding function, which means rounding down to ensure that n is an integer; S22. Apply linear projection to each image patch, projecting it from the flattened feature vector to a higher-dimensional feature space to obtain a feature vector of dimension d; integrate the feature vectors after the projection of all image patches to form a matrix X∈R d ; S23. Use Swin Transformer Block to process image patches and capture local dependency information through the local window attention mechanism. Divide the image patch into V windows of size M×M. For each image patch in the window, use the multi-head self-attention mechanism to capture local features. The number V of image windows is adjusted by the window number adjustment formula, which is: Where q is the number of image patches; n is the size of the image patches; Patch Merging is used to perform downsampling and merge adjacent blocks; a sliding window mechanism is used to capture the global features of small image blocks; the window is slidingly displaced, and the sliding operation of each window in the tth layer is S t , then the window feature after sliding is X′=S t (X); After the above steps, the final output feature matrix X∈R H′×W′×d′ ; Where H′ and W′ are the spatial sizes after multiple patch mergings, and d′ is the number of channels; S24. Further analyze and extract the feature matrix to obtain image parameter data.

3. The deep learning-based intelligent pre-print image optimization system according to claim 2, characterized in that: The method of preprocessing the acquired image parameter data, pre-printing device parameter data, and user behavior data to obtain an image feature data set, a pre-printing device feature data set, and a user behavior feature data set includes: The acquired image parameter data, pre-printing device parameter data, and user behavior data were cleaned, interpolation was used to fill missing values, and clustering algorithms were used to detect and remove outliers. Z-score standardization was used to transform the image parameter data, pre-printing device parameter data, and user behavior data into a standard normal distribution with a mean of 0 and a standard deviation of 1 to eliminate dimensionality effects. After the preprocessing in the above steps, the image feature dataset, pre-printed device feature dataset and user behavior feature dataset are obtained.

4. The deep learning-based intelligent pre-print image optimization system according to claim 3, characterized in that: The method of evaluating the influence of the user behavior feature dataset on the image feature dataset and the pre-printed device feature dataset to obtain the influenced feature dataset includes: Taking image features, pre-printed device features and user behavior features as input variables, a joint regression model is constructed. By constructing the joint regression model, the impact of the user behavior feature dataset on the image feature dataset and the pre-printed device feature dataset is analyzed, and finally the impact feature dataset is obtained.

5. The deep learning-based intelligent pre-print image optimization system according to claim 4, characterized in that: The training method of the pre-printed image error prediction model includes: The dataset is divided into training set, validation set and test set to train the model and evaluate the model performance; the sample set is a subset of the dataset, and each sample set includes a historical impact feature dataset and the corresponding preprint image error; Constructing a pre-print image error prediction model, which includes an input layer, a convolution layer, a pooling layer, a batch normalization layer, a fully connected layer, and an output layer; the input layer is used to input a historical impact feature dataset, and the output layer is used to output the pre-print image error; the pre-print image error prediction model is a 3D convolutional neural network model; The mean absolute error (MAE) loss function is used to measure the error between the model's predicted value and the actual value. The model is trained using the training set data and the Adam optimizer is used to minimize the loss function. The model performance is evaluated using the validation set, and the model's hyperparameters are tuned until the model performance no longer improves significantly or the preset stopping condition is met. The test set is used to evaluate the performance of the model in the prediction task, and the current influencing feature dataset is input into the trained pre-print image error prediction model to obtain the pre-print image error.

6. The deep learning-based intelligent pre-print image optimization system according to claim 5, characterized in that: The method of comparing the predicted pre-printed image error with a preset pre-printed image error threshold to determine whether the pre-printed image quality meets the standard includes: If the predicted pre-print image error is less than the preset pre-print image error threshold, the pre-print image quality is determined to be up to standard; If the predicted pre-print image error is greater than or equal to a preset pre-print image error threshold, it is determined that the pre-print image quality does not meet the standard.

7. The deep learning-based intelligent pre-print image optimization system according to claim 6, characterized in that: The pre-printed image data that needs to be optimized includes image resolution, color space, image format, color calibration rate, image noise, image brightness and contrast.

8. The deep learning-based intelligent pre-print image optimization system according to claim 7, characterized in that: The training method of the image defect detection model includes: The dataset is divided into a training set, a validation set, and a test set to train the model and evaluate its performance. The sample set is a subset of the dataset, and each sample set includes historical pre-printed image data that needs to be optimized and the corresponding pre-printed image defect type. The model's input data is historical pre-printed image data that needs to be optimized; the model's output label is the pre-printed image defect type; the image defect detection model consists of v decision trees. Bootstrap sampling is used to train each decision tree by randomly selecting a subset from the training set to train the tree, and randomly selecting a subset of features to split nodes until a stopping condition is met; the image defect detection model is a random forest model; Use multi-class cross entropy as the model's loss function to measure the difference between the model's predicted value and the actual value. Use the training set to train the image defect detection model and use the backpropagation algorithm to update the model parameters to minimize the loss function. Use the validation set to evaluate the performance of the image defect detection model by calculating the precision metric. Using the grid search hyperparameter optimization method, the hyperparameters of the image defect detection model are tuned based on the performance feedback of the validation set. The model hyperparameter combination is adjusted until the model performance no longer improves significantly or the preset stopping condition is reached. The test set is used to evaluate the performance of the model in the prediction task, and the trained image defect detection model is used to predict the pre-printed image defect data that currently needs to be optimized to obtain the pre-printed image defect type.

9. The deep learning-based intelligent pre-print image optimization system according to claim 8, characterized in that: The method of executing a corresponding image optimization solution according to the severity of the pre-printed image defect type includes: Construct a pre-printed image defect degree assessment model, calculate the severity of the pre-printed image defect type through the pre-printed image defect degree assessment model, and execute the corresponding image optimization solution according to the severity of the pre-printed image defect type; the pre-printed image defect degree assessment model is: Among them, S ge is the severity of the pre-printed image defect type; R re is the image resolution; C lr is the color space; F ot For image format; C lb is the color calibration rate; N se is the image noise; L bs is the image brightness; D st is the contrast; ω1, ω2, ω3, ω4, ω5, ω6, and ω7 are the weight factors of each parameter.

10. A deep learning-based intelligent pre-print image optimization method, used to implement the deep learning-based intelligent pre-print image optimization system according to any one of claims 1 to 9, characterized in that: include: S1. Collect image parameter data, pre-printing device parameter data and user behavior data; S2. Preprocess the acquired image parameter data, pre-printing device parameter data, and user behavior data to obtain an image feature dataset, a pre-printing device feature dataset, and a user behavior feature dataset; evaluate the impact of the user behavior feature dataset on the image feature dataset and the pre-printing device feature dataset to obtain an impact feature dataset; S3. Obtain a pre-print image error prediction model based on the influencing feature dataset, and predict the pre-print image error using the pre-print image error prediction model; S4, comparing the predicted pre-printed image error with a preset pre-printed image error threshold to determine whether the pre-printed image quality meets the standard; S5. If the pre-printed image quality meets the standards, there is no need to optimize the pre-printed image, and a printing instruction is directly issued through the intelligent pre-printing terminal to print the pre-printed image. If the pre-printed image quality does not meet the standards, the pre-printed image needs to be optimized, and an optimization instruction is issued through the intelligent pre-printing terminal, and the pre-printed image data that needs to be optimized is automatically collected. S6. Train an image defect detection model based on the optimized pre-printed image data as needed, and obtain the pre-printed image defect type based on the image defect detection model; The severity of the pre-printed image defect type is obtained through the pre-printed image defect degree assessment model, and the corresponding image optimization solution is executed according to the severity of the pre-printed image defect type.

Images