A method of defect evaluation for a flexographic first piece
By constructing an electronic sample character library and combining deep learning with fine defect detection, the problem of low efficiency in manual quality inspection in flexographic printing first inspection is solved, and efficient and accurate defect assessment of flexographic first articles is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGDONG UNIV OF TECH
- Filing Date
- 2022-11-04
- Publication Date
- 2026-05-12
AI Technical Summary
Manual quality inspection during the initial inspection of flexographic printing is inefficient and prone to false detections, especially in character detection and dirt detection. The false detection rate is particularly high due to the large number of characters, the complexity of the language, and the visual differences in the fabric substrate.
We constructed an electronic proof character library and a flexographic first-article character library, combined a deep evaluation network for character detection, used time-frequency conversion and Blob analysis for contamination detection, and utilized deep learning and fine defect detection technologies to improve the accuracy and efficiency of detection.
By deeply evaluating the robustness of the network for detection, the impact of visual differences in production processes is reduced, the false detection rate is decreased, and the detection accuracy and efficiency are improved, thus achieving accurate defect assessment of the first piece of flexographic printing.
Smart Images
Figure CN115601760B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of textile industry testing technology, and more specifically, to a defect assessment method for first-piece flexographic printing. Background Technology
[0002] Flexographic printing is a printing method that uses a flexible letterpress plate. During printing, ink is first delivered to the inking roller by the inking feed roller, which then evenly and stably coats the ink onto the surface of the printing plate cylinder. As the substrate passes through the printing plate cylinder and impression cylinder, the image on the printing plate is transferred to the substrate, resulting in a clear image. Flexographic printing boasts high production efficiency, with speeds typically reaching 100 meters per second, far exceeding that of letterpress printing. It also supports a wide range of substrates, including fabrics and other composite materials.
[0003] The flexographic printing process inevitably produces defects due to machine malfunctions and ink failures. Therefore, before mass production of flexographic labels, workers must undergo a first-print inspection process after each shift change, production change, and equipment setup and adjustment. This means that the first or first few flexographic products printed must be compared and inspected with the electronic proof to check the printing quality of the flexographic products.
[0004] Currently, the first inspection of flexographic printing uses a manual three-inspection system: self-inspection, mutual inspection, and specialized inspection. However, the characters on the labels involve languages from all over the world, and each label has an average of more than 200 characters. This not only makes manual quality inspection inefficient, but also makes it very easy to miss or mis-inspect.
[0005] In addition, the substrate of flexographic labels is fabric with a large number of tiny pores. The uneven thickness of the substrate material leads to uneven printing force, which causes slight deformation of the characters and patterns on the sample. Furthermore, when the flexographic plate is printed by transferring ink through the anilox roller, the characters on the printed product are visually different from the electronic proof due to the limitations of the production process, which can easily lead to false detections. Summary of the Invention
[0006] This invention provides a defect assessment method for the first piece of flexographic printing to overcome the technical shortcomings of current manual quality inspection, which is prone to false detection.
[0007] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:
[0008] A defect assessment method for flexographic printing first articles includes character detection and dirt detection, wherein the character detection includes the following steps:
[0009] S1: Construct a character library for the electronic proof and a character library for the flexographic first print based on the character position correspondence between the electronic proof and the flexographic first print.
[0010] S2: Input the characters in the electronic proof and the characters in the flexographic first print into the pre-constructed deep evaluation network to obtain the character detection results and complete the character detection;
[0011] The contamination detection includes the following steps: performing time-frequency conversion, Gaussian filtering, and Blob analysis on the image of the flexographic first piece to obtain the contamination detection result and complete the contamination detection;
[0012] The defect assessment results of the first flexographic print are obtained based on the character detection results and the dirt detection results.
[0013] In the above solution, the problem of inconsistent selection margins is avoided by constructing character libraries for electronic proofs and flexographic first-piece documents, eliminating large-area false defects that may exist around the edges. Furthermore, a deep evaluation network is used to perform more comprehensive defect detection on corresponding characters in the flexographic first-piece document and electronic proof, significantly improving robustness, reducing the impact of visual differences objectively present in the production process, lowering the false detection rate, and improving detection accuracy. Simultaneously, by detecting characters and dirt in different ways—using deep learning for character content and fine-grained defect detection for textured substrates—precise location is achieved, thereby improving detection efficiency while maintaining accuracy.
[0014] Preferably, the character positions of the electronic draft are obtained by projection.
[0015] Preferably, the steps for constructing the electronic draft character library are as follows:
[0016] Establish a Cartesian coordinate system on the electronic proof, with the uniform distribution direction of the characters as the horizontal axis. On the vertical axis, count the number of pixels with a value of 0 along the horizontal axis to obtain a vertical axis histogram. Calculate the position of each character in each row based on the distribution of the vertical axis histogram. Then, for each row of characters, count the number of pixels with a value of 0 along the vertical axis to obtain a horizontal axis histogram. Obtain the width of the character in the corresponding row and the distance of the character from the vertical axis. Based on the character's position, width, and distance from the vertical axis, obtain the coordinates of the upper left and lower right corners of the character's bounding rectangle. Use the coordinates of the bounding rectangle to extract and save the characters on the bounding rectangle, thus constructing the electronic proof character library.
[0017] Preferably, a character positioning network is used to obtain the character positions of the flexographic printing head. The character positioning network is used to detect individual characters and the connection relationships between characters, and to determine the character positions based on the detection results.
[0018] Preferably, the character localization network includes a feature extraction network and a decoder;
[0019] The feature extraction network includes a first convolutional stage, a second convolutional stage, a third convolutional stage, a fourth convolutional stage, and a fifth convolutional stage connected in sequence. Each convolutional stage includes a convolutional operation and a max pooling operation.
[0020] The decoder includes a first upsampling module, a second upsampling module, a third upsampling module, a fourth upsampling module, and a convolution module;
[0021] In the character localization network, the output of the fifth convolution stage is stacked with itself and then input into the fourth upsampling module. The output of the fourth upsampling module is stacked with the output of the fourth convolution stage and then input into the third upsampling module. The output of the third upsampling module is stacked with the output of the third convolution stage and then input into the second upsampling module. The output of the second upsampling module is stacked with the output of the second convolution stage and then input into the first upsampling module. After completing the last upsampling operation, the output is input into the convolution module. Finally, the predicted character region score and connection relationship score are output. The character position is determined based on the predicted character region score and connection relationship score, and the character image of the flexographic printhead is captured.
[0022] Preferably, the confidence mapping s of the character positioning network conf (w) is:
[0023]
[0024] The image pixel confidence S is obtained from the confidence map. c (p):
[0025]
[0026] The objective function L for constructing the character localization network is:
[0027]
[0028] Where l(w) represents the length of character w, l c (w) represents the length of the predicted character w, p represents the number of pixels in R(w), R(w) represents the bounding box region where character w is located, and S r (p) represents the predicted character region score. S represents the actual character region score. a (p) represents the predicted connectivity score. This represents the actual connection score.
[0029] Preferably, the depth evaluation network includes two VGG-16 networks, used for the detection of characters in the electronic proof and characters in the flexographic first print, respectively;
[0030] In the deep evaluation network, the corresponding character images of the electronic proof character library and the flexographic first document character library are input into different VGG-16 networks. Then, the outputs of the two VGG-16 networks are stacked and passed through deconvolution layers and fully connected layers to obtain the segmentation result as the character detection result.
[0031] Preferably, the weights of the two VGG-16 networks are shared.
[0032] Preferably, the loss function Loss of the deep evaluation network is:
[0033] Loss = Dice_loss + CE_loss
[0034] Dice_loss=1-Dice
[0035]
[0036]
[0037] Where Dice represents the Dice coefficient, Dice_loss represents the Dice loss function, N represents the total number of pixels, and y i ={0,1} represents the actual binary tag. This represents the predicted binary label, and CE_loss represents the binary cross-entropy loss function. Indicates that the output belongs to The probability of the label.
[0038] Preferably, the specific steps for dirt detection are as follows:
[0039] First, Fourier transform is used to convert the spatial domain features of the flexographic printing head image into frequency domain features. Then, Gaussian filtering is used to obtain the high-frequency features of the image. Finally, Blob analysis is used to segment the filtered image, calculate the gray-level co-occurrence matrix of each region in the image, filter defects based on energy, and obtain a binary dirt segmentation map as the dirt detection result, thus completing the dirt detection.
[0040] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:
[0041] This invention provides a defect assessment method for flexographic printing first drafts. By constructing character libraries for both electronic proofs and flexographic printing first drafts, it avoids the problem of inconsistent selection margins and eliminates large-area false defects that may exist around the edges. Furthermore, by using a deep evaluation network to perform more comprehensive defect detection on corresponding characters in the flexographic printing first draft and electronic proofs, robustness is greatly improved, the impact of visual differences objectively present in the production process is reduced, the false detection rate is lowered, and the detection accuracy is improved. Simultaneously, by detecting characters and contaminants in different ways—using deep learning for character content detection and fine-grained defect detection for textured substrates—precise localization is achieved, thereby improving detection efficiency while maintaining accuracy. Attached Figure Description
[0042] Figure 1 This is a flowchart illustrating the implementation of the technical solution of the present invention;
[0043] Figure 2 This is a schematic diagram illustrating the method of obtaining character positions using projection in this invention;
[0044] Figure 3 This is an example diagram of the ordinate histogram obtained by projection method in this invention;
[0045] Figure 4 This is a schematic diagram of the character positioning network architecture in this invention;
[0046] Figure 5 This is a schematic diagram illustrating the acquisition of character positions via a character positioning network in this invention.
[0047] Figure 6 This is a schematic diagram of the architecture of the deep evaluation network in this invention;
[0048] Figure 7 This is a schematic diagram of the architecture of the upsampling module in this invention;
[0049] Figure 8 This is a schematic diagram of the architecture of the convolution module in this invention. Detailed Implementation
[0050] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.
[0051] To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions.
[0052] It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.
[0053] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0054] Example 1
[0055] like Figure 1 As shown, a defect assessment method for flexographic printing first articles includes character detection and dirt detection, wherein the character detection includes the following steps:
[0056] S1: Construct a character library for the electronic proof and a character library for the flexographic first print based on the character position correspondence between the electronic proof and the flexographic first print.
[0057] S2: Input the characters in the electronic proof and the characters in the flexographic first print into the pre-constructed deep evaluation network to obtain the character detection results and complete the character detection;
[0058] The contamination detection includes the following steps: performing time-frequency conversion, Gaussian filtering, and Blob analysis on the image of the flexographic first piece to obtain the contamination detection result and complete the contamination detection;
[0059] The defect assessment results of the first flexographic print are obtained based on the character detection results and the dirt detection results.
[0060] In the specific implementation process, by constructing character libraries for electronic proofs and flexographic first-piece documents, the problem of inconsistent selection margins is avoided, eliminating potentially large-area false defects around the edges. Furthermore, a deep evaluation network is used to perform more comprehensive defect detection on corresponding characters in the flexographic first-piece document and electronic proof, significantly improving robustness, reducing the impact of visual differences inherent in the production process, lowering the false detection rate, and improving detection accuracy. Simultaneously, by detecting characters and dirt in different ways—using deep learning for character content and fine-grained defect detection for textured substrates—precise location is achieved, thereby improving detection efficiency while maintaining accuracy.
[0061] Example 2
[0062] A defect assessment method for flexographic printing first articles includes character detection and dirt detection, wherein the character detection includes the following steps:
[0063] S1: Construct a character library for the electronic proof and a character library for the flexographic first print based on the character position correspondence between the electronic proof and the flexographic first print.
[0064] More specifically, such as Figure 2 As shown, the character positions of the electronic draft are obtained using the projection method.
[0065] More specifically, the steps for constructing the electronic draft character library are as follows:
[0066] Establish a Cartesian coordinate system on the electronic proof, with the direction of the uniform distribution of characters as the horizontal axis; on the vertical axis, count the number of pixels with a value of 0 along the horizontal axis to obtain a vertical histogram, as shown below. Figure 3As shown, the position of each character in each row is calculated based on the distribution of the vertical axis histogram; then, for each character in each row, the number of pixels with a pixel value of 0 is counted along the vertical axis to obtain the horizontal axis histogram, and the width of the character in the corresponding row and the distance of the character from the vertical axis are obtained; based on the position of the character, the width of the character, and the distance of the character from the vertical axis, the coordinates of the upper left and lower right corners of the character's bounding rectangle are obtained, and the characters on the bounding rectangle are extracted and saved based on the coordinates of the bounding rectangle to construct the electronic draft character library.
[0067] More specifically, a character positioning network is used to obtain the character positions of the flexographic printing head. The character positioning network is used to detect individual characters and the connection relationships between characters, and to determine the character positions based on the detection results.
[0068] More specifically, such as Figure 4 As shown, the character localization network includes a feature extraction network and a decoder;
[0069] The feature extraction network includes a first convolutional stage, a second convolutional stage, a third convolutional stage, a fourth convolutional stage, and a fifth convolutional stage connected in sequence. Each convolutional stage includes a convolutional operation and a max pooling operation.
[0070] The decoder includes a first upsampling module, a second upsampling module, a third upsampling module, a fourth upsampling module, and a convolution module;
[0071] In the character localization network, the output of the fifth convolutional stage is stacked with itself and then input into the fourth upsampling module. The output of the fourth upsampling module is stacked with the output of the fourth convolutional stage and then input into the third upsampling module. The output of the third upsampling module is stacked with the output of the third convolutional stage and then input into the second upsampling module. The output of the second upsampling module is stacked with the output of the second convolutional stage and then input into the first upsampling module to complete the final upsampling operation. Finally, the output is fed into the convolutional module, and the predicted character region score and connectivity score are output. Based on the predicted character region score and connectivity score, the character position is determined, and the character image of the flexographic printhead is captured. Figure 5 As shown, where Figure 5 The left half shows the character region score, and the right half shows the bounding box selection result of the character localization network.
[0072] S2: Input the characters in the electronic proof and the characters in the flexographic first print into the pre-constructed deep evaluation network to obtain the character detection results and complete the character detection;
[0073] More specifically, such as Figure 6 As shown, the deep evaluation network includes two VGG-16 networks, which are used for the detection of characters in electronic proofs and characters in flexographic first prints, respectively.
[0074] In the deep evaluation network, the corresponding character images of the electronic proof character library and the flexographic first document character library are input into different VGG-16 networks. Then, the outputs of the two VGG-16 networks are stacked and passed through deconvolution layers and fully connected layers to obtain the segmentation result as the character detection result.
[0075] The contamination detection includes the following steps: performing time-frequency conversion, Gaussian filtering, and Blob analysis on the image of the flexographic first piece to obtain the contamination detection result and complete the contamination detection;
[0076] More specifically, the steps for dirt detection are as follows:
[0077] First, Fourier transform is used to convert the spatial domain features of the flexographic printing head image into frequency domain features. Then, Gaussian filtering is used to obtain the high-frequency features of the image. Finally, Blob analysis is used to segment the filtered image, calculate the gray-level co-occurrence matrix of each region in the image, filter defects based on energy, and obtain a binary dirt segmentation map as the dirt detection result, thus completing the dirt detection.
[0078] The defect assessment results of the first flexographic print are obtained based on the character detection results and the dirt detection results.
[0079] Example 3
[0080] A defect assessment method for flexographic printing first articles includes character detection and dirt detection, wherein the character detection includes the following steps:
[0081] S1: Construct a character library for the electronic proof and a character library for the flexographic first print based on the character position correspondence between the electronic proof and the flexographic first print.
[0082] More specifically, the character positions of the electronic draft are obtained using a projection method.
[0083] More specifically, the steps for constructing the electronic draft character library are as follows:
[0084] Establish a Cartesian coordinate system on the electronic proof, with the uniform distribution direction of the characters as the horizontal axis. On the vertical axis, count the number of pixels with a value of 0 along the horizontal axis to obtain a vertical axis histogram. Calculate the position of each character in each row based on the distribution of the vertical axis histogram. Then, for each row of characters, count the number of pixels with a value of 0 along the vertical axis to obtain a horizontal axis histogram. Obtain the width of the character in the corresponding row and the distance of the character from the vertical axis. Based on the character's position, width, and distance from the vertical axis, obtain the coordinates of the upper left and lower right corners of the character's bounding rectangle. Use the coordinates of the bounding rectangle to extract and save the characters on the bounding rectangle, thus constructing the electronic proof character library.
[0085] More specifically, a character positioning network is used to obtain the character positions of the flexographic printing head. The character positioning network is used to detect individual characters and the connection relationships between characters, and to determine the character positions based on the detection results.
[0086] More specifically, the character localization network includes a feature extraction network and a decoder;
[0087] The feature extraction network includes a first convolutional stage, a second convolutional stage, a third convolutional stage, a fourth convolutional stage, and a fifth convolutional stage connected in sequence. Each convolutional stage includes a convolutional operation and a max pooling operation.
[0088] This embodiment uses VGG-16 BN as the feature extraction network;
[0089] The decoder includes a first upsampling module, a second upsampling module, a third upsampling module, a fourth upsampling module, and a convolution module;
[0090] Among them, such as Figure 7 As shown, each upsampling module includes two convolutional layers and two batch normalization layers. The kernel size of the convolutional layers is 3*3, and the number of output channels is out_ch. The decoder adopts a top-down feature aggregation method; the convolutional module is as follows: Figure 8 As shown, each layer includes a convolutional layer and a ReLU layer.
[0091] In the character localization network, the output of the fifth convolution stage is stacked with itself and then input into the fourth upsampling module. The output of the fourth upsampling module is stacked with the output of the fourth convolution stage and then input into the third upsampling module. The output of the third upsampling module is stacked with the output of the third convolution stage and then input into the second upsampling module. The output of the second upsampling module is stacked with the output of the second convolution stage and then input into the first upsampling module. After completing the last upsampling operation, the output is input into the convolution module. Finally, the predicted character region score and connection relationship score are output. The character position is determined based on the predicted character region score and connection relationship score, and the character image of the flexographic printhead is captured.
[0092] More specifically, the confidence mapping s of the character localization network conf (w) is:
[0093]
[0094] In this embodiment, a confidence mapping is obtained by utilizing the number of characters to prevent erroneous labels generated under weak supervision from biasing the network.
[0095] The image pixel confidence S is obtained from the confidence map. c (p):
[0096]
[0097] The objective function L for constructing the character localization network is:
[0098]
[0099] Where l(w) represents the length of character w, l c (w) represents the length of the predicted character w, p represents the number of pixels in R(w), R(w) represents the bounding box region where character w is located, and S r (p) represents the predicted character region score. S represents the actual character region score. a (p) represents the predicted connectivity score. This represents the actual connection score. A high confidence level is indicated when the result matches the text length; otherwise, a low confidence level is indicated.
[0100] In this embodiment, after obtaining the output of the character positioning network, a threshold calculation is performed on the output score probability map, and region connection is performed using Connected Component Labeling (CCL), followed by processing.
[0101] S2: Input the characters in the electronic proof and the characters in the flexographic first print into the pre-constructed deep evaluation network to obtain the character detection results and complete the character detection;
[0102] More specifically, the deep evaluation network includes two VGG-16 networks, used for the detection of characters in electronic proofs and characters in flexographic first prints, respectively;
[0103] In the deep evaluation network, the corresponding character images of the electronic proof character library and the flexographic first document character library are input into different VGG-16 networks. Then, the outputs of the two VGG-16 networks are stacked and passed through deconvolution layers and fully connected layers to obtain the segmentation result as the character detection result.
[0104] In the specific implementation, due to the small size of the objects to be detected, only the first three stages and the last three convolutional layers of VGG-16 are used, with weights shared between the two networks. The corresponding electronic proof characters and flexographic first-issue characters are input into the depth evaluation network. The feature vectors output from the last convolutional layer of the two VGG-16 networks are stacked and then input into a deconvolutional layer for upsampling, reducing the feature dimension to half that of the stacked vector while still including features from both the electronic proof and the first-issue characters. The output of the first convolutional layer of stage 3 is stacked with the upsampled feature vector, and similarly, the feature vector after two upsamplings is stacked with the output of the first convolutional layer of stage 2. When the size of the upsampled feature vector matches the size of the input image, the segmentation result is obtained after passing through a fully connected (softmax) layer.
[0105] More specifically, the two VGG-16 networks share weights.
[0106] More specifically, the loss function Loss of the deep evaluation network is:
[0107] Loss = Dice_loss + CE_loss
[0108] Dice_loss=1-Dice
[0109]
[0110]
[0111] Where Dice represents the Dice coefficient, Dice_loss represents the Dice loss function, N represents the total number of pixels, and y i ={0,1} represents the actual binary tag. This represents the predicted binary label, and CE_loss represents the binary cross-entropy loss function. Indicates that the output belongs to The probability of the label.
[0112] In practice, the binary cross-entropy function not only converges quickly but also accumulates the loss of each pixel across the entire batch. The Dice coefficient is used as the loss function to make the predicted value as close as possible to the actual value.
[0113] The contamination detection includes the following steps: performing time-frequency conversion, Gaussian filtering, and Blob analysis on the image of the flexographic first piece to obtain the contamination detection result and complete the contamination detection;
[0114] More specifically, the steps for dirt detection are as follows:
[0115] First, Fourier transform is used to convert the spatial domain features of the flexographic printing head image into frequency domain features. Then, Gaussian filtering is used to obtain the high-frequency features of the image. Finally, Blob analysis is used to segment the filtered image, calculate the gray-level co-occurrence matrix of each region in the image, filter defects based on energy, and obtain a binary dirt segmentation map as the dirt detection result, thus completing the dirt detection.
[0116] In practical implementation, frequency characteristics refer to the grayscale variation characteristics of the image. Low-frequency characteristics are those with insignificant grayscale changes, such as the overall outline of the image, while high-frequency characteristics are those with drastic grayscale changes, such as image edges and noise. Since most dirt is relatively dark in color and has low grayscale energy, dirty areas can be filtered based on energy.
[0117] The defect assessment results of the first flexographic print are obtained based on the character detection results and the dirt detection results.
[0118] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.
Claims
1. A method for defect assessment of flexographic printing first articles, characterized in that, This includes character detection and contamination detection, wherein the character detection includes the following steps: S1: Construct a character library for the electronic proof and a character library for the flexographic first print based on the character position correspondence between the electronic proof and the flexographic first print. S2: Input the characters in the electronic proof and the characters in the flexographic first print into the pre-constructed deep evaluation network to obtain the character detection results and complete the character detection; The contamination detection includes the following steps: performing time-frequency conversion, Gaussian filtering, and Blob analysis on the image of the flexographic first piece to obtain the contamination detection result and complete the contamination detection; The defect assessment results of the first flexographic print are obtained based on the character detection results and the dirt detection results; The method employs a character localization network to obtain the character positions of the flexographic printing first piece. This network detects individual characters and the connections between them, and determines the character positions based on the detection results. The character localization network includes a feature extraction network and a decoder. The feature extraction network comprises a first convolutional stage, a second convolutional stage, a third convolutional stage, a fourth convolutional stage, and a fifth convolutional stage connected sequentially. Each convolutional stage includes a convolution operation and a max-pooling operation. The decoder includes a first upsampling module, a second upsampling module, a third upsampling module, a fourth upsampling module, and a convolutional module. In the process, the output of the fifth convolution stage is stacked with itself and then input into the fourth upsampling module. The output of the fourth upsampling module is stacked with the output of the fourth convolution stage and then input into the third upsampling module. The output of the third upsampling module is stacked with the output of the third convolution stage and then input into the second upsampling module. The output of the second upsampling module is stacked with the output of the second convolution stage and then input into the first upsampling module to complete the last upsampling operation. Finally, it is input into the convolution module, and the predicted character region score and connection relationship score are output. The character position is determined based on the predicted character region score and connection relationship score, and the character image of the flexographic first piece is captured. The confidence mapping of the character localization network for: Obtain image pixel confidence based on confidence mapping : The objective function for constructing the character localization network : in, Character Length, Characters representing prediction Length, express pixels, Character The character bounding box area it is located in This represents the predicted character region score. This represents the actual character region score. This represents the predicted connectivity score. This represents the actual connection score; The specific steps for dirt detection are as follows: First, Fourier transform is used to convert the spatial domain features of the character image of the flexographic printing first piece into frequency domain features. Then, Gaussian filtering is used to obtain the high-frequency features of the image. Finally, Blob analysis is used to segment the filtered image, calculate the gray-level co-occurrence matrix of each region in the image, filter defects according to energy, and obtain a binary dirt segmentation map as the dirt detection result to complete the dirt detection.
2. The defect assessment method for flexographic printing first articles according to claim 1, characterized in that, The character positions in the electronic draft were obtained using a projection method.
3. The defect assessment method for flexographic printing first articles according to claim 2, characterized in that, The specific steps for constructing the electronic manuscript character library are as follows: Establish a Cartesian coordinate system on the electronic proof, with the uniform distribution direction of the characters as the horizontal axis. On the vertical axis, count the number of pixels with a value of 0 along the horizontal axis to obtain a vertical axis histogram. Calculate the position of each character in each row based on the distribution of the vertical axis histogram. Then, for each row of characters, count the number of pixels with a value of 0 along the vertical axis to obtain a horizontal axis histogram. Obtain the width of the character in the corresponding row and the distance of the character from the vertical axis. Based on the character's position, width, and distance from the vertical axis, obtain the coordinates of the upper left and lower right corners of the character's bounding rectangle. Use the coordinates of the bounding rectangle to extract and save the characters on the bounding rectangle, thus constructing the electronic proof character library.
4. The defect assessment method for flexographic printing first articles according to claim 1, characterized in that, Weight sharing between the two VGG-16 networks.
5. The defect assessment method for flexographic printing first articles according to claim 1, characterized in that, The loss function Loss of the deep evaluation network is: in, Represents the Dice coefficient. This represents the Dice loss function, where N represents the total number of pixels. This represents the actual binary tag. The binary label representing the prediction. This represents the binary cross-entropy loss function. Indicates that the output belongs to The probability of the label.