Cold-rolled strip post-grinding roller mark defect detection method and system based on deep learning

Abstract

The present invention provides a method and system for detecting roller marks defects after grinding of cold-rolled strip steel based on deep learning, comprising: step M1: collecting defect images reflected by the cold-rolled strip steel after grinding by an automatic grinding system; step M2: preprocessing the defect images, and segmenting the preprocessed defect images into pictures of preset sizes; step M3: manually annotating the pictures of preset sizes after segmentation to create a data set; step M4: using the Hrnet pre-training model based on the cityscapes data set to train the data set with a preset number of times and learning rate, so that the defect recognition rate annotated in the data set reaches a preset value, and a trained model is obtained; step M5: deploying the trained model in the cold-rolled strip steel production line to monitor the defect recognition results online; after the present invention is put into use, the generation of roller mark defect circles can be effectively reduced, and the yield rate can be improved. At the same time, the automation of the roller mark recognition station is realized, and the labor intensity of existing staff is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of cold-rolled strip grinding roll mark defect detection based on deep learning, and specifically, to a cold-rolled strip grinding roll mark defect detection method and system based on deep learning. More specifically, it relates to an online detection method that can replace manual detection, can be used in conjunction with a strip automatic grinding system, and can be widely used in a strip grinding roll mark defect detection system at the outlet of a cold rolling unit in a steel plant. Background Art

[0002] During the production process of the cold rolling unit, the rollers in contact with the strip steel produce point-like protrusions or adhere to fine impurities due to wear, which in turn causes periodic convex or concave marks on the surface of the strip steel. Once this happens, batches of hard rolled coils are often downgraded due to unqualified surface quality, resulting in large economic losses. This type of defect accounts for more than 90% of all post-grinding defects, and the defects are usually very small, with the smallest being as small as a needle tip. Directly irradiating the cold-rolled strip steel with light generally cannot identify this defect, and it needs to be polished with a quality inspection oilstone to make it visible.

[0003] Although after grinding, the camera can capture the roller mark defects under the right light angle. However, due to the damage caused to the strip surface by the grinding process (such as the attached Figure 1 ), the roller mark defect is hidden in the complex background of the wear mark, which increases the difficulty of computer vision recognition. Some roller mark defects with low contrast compared with the background wear mark are often not recognized by traditional machine vision algorithms. In addition, some roller marks are very small, occupying only a few pixels, which increases the difficulty of recognition.

[0004] Roller mark defects are low-frequency events in the production process of cold rolling mills, so the number of roll mark defect images that can be collected is very limited. Some deep learning networks that need to be trained with a large amount of sample data to obtain useful essential features cannot be used.

[0005] Therefore, it is necessary to design an online detection method for roller mark defects after grinding of cold-rolled strip based on deep learning to overcome the above problems.

[0006] Patent document CN108021938A (application number: 201711224806.9) discloses an online detection method for surface defects of cold-rolled strip, characterized in that it includes the following steps: step S1, using light to irradiate the cold-rolled strip and receive the image reflected by the cold-rolled strip; step S2, preprocessing the image; step S3, constructing a deep learning network for surface defect detection of cold-rolled strip, using the deep learning network for defect detection of cold-rolled strip to extract defect features of the defect image, and identify the defect type of the defect features; step S4, determining the quality grade of the cold-rolled strip according to the defect type, severity, number and distribution of the defect features.

[0007] Patent document CN108021938A (application number: 201711224806.9) discloses an online detection method for surface defects of cold-rolled strip steel, but the following three points cannot meet the detection requirements of roller mark defects:

[0008] 1. It does not include the detection method after the grinding process is completed. In the detection method designed, the image is obtained by irradiating the cold-rolled strip with light and receiving the defect image reflected by the cold-rolled strip. Therefore, its deep learning network model does not take into account the complex strip surface background after grinding.

[0009] 2. In the process of establishing a machine learning model, it is necessary to train with a large amount of sample data to obtain useful essential features and improve the accuracy of defect classification and detection. However, roller printing defects cannot provide a large amount of sample data.

[0010] 3. It cannot solve the problem of imbalanced categories between roller prints and backgrounds. Summary of the invention

[0011] In view of the defects in the prior art, the object of the present invention is to provide a method and system for detecting roller mark defects after grinding of cold-rolled strip based on deep learning.

[0012] According to the present invention, a method for detecting roller marks on cold-rolled strip after grinding based on deep learning is provided, comprising:

[0013] Step M1: collecting defect images reflected by the cold-rolled steel strip after being polished by the automatic polishing system;

[0014] Step M2: preprocessing the defect image, and dividing the preprocessed defect image into images of a preset size;

[0015] Step M3: manually label the segmented images of preset size and create a data set;

[0016] Step M4: Use the Hrnet pre-trained model based on the cityscapes data set to train the data set for a preset number of times and learning rate, so that the defect recognition rate marked in the data set reaches a preset value, and obtain a trained model;

[0017] Step M5: deploy the trained model on the cold-rolled strip production line to monitor defect recognition results online;

[0018] The Hrnet pre-training model is a parallel structure that maintains high-resolution representation at all times, rather than simply restoring high-resolution representation from low-resolution representation.

[0019] Preferably, the step M1 comprises: after the cold-rolled steel strip is polished by the matching automatic steel strip polishing system, the cold-rolled steel strip is irradiated with a light source at a preset angle, and the camera receives a defect image emitted by the cold-rolled steel strip at a preset angle;

[0020] The step M2 comprises:

[0021] Step M2.1: Cut the invalid edge of the defect image and obtain the ROI area;

[0022] Step M2.2: Use an equalization algorithm to equalize the brightness of the ROI area to obtain a ROI area with equalized brightness;

[0023] Step M2.3: Divide the ROI area with balanced brightness into images of preset sizes.

[0024] Preferably, the step M3 comprises:

[0025] Step M3.1: manually label the segmented images of preset size and generate PNG tags in the form of legends, representing the roller print and the background respectively;

[0026] Step M3.2: Divide the images and PNG labels into training set, validation set and test set according to the preset ratio;

[0027] Step M3.3: After classifying the images and PNG labels, create a dataset;

[0028] Step M3.4: Check the data set according to the preset requirements. If the check fails, repeat steps M3.2 to M3.3 until the data set meets the preset requirements.

[0029] Preferably, the step M4 comprises:

[0030] Step M4.1: Load the Hrnet pre-trained model based on the cityscapes dataset;

[0031] Step M4.2: The loss function uses the dice loss function and the bce loss function to evenly distribute the categories of the roller print and the background, and obtain a evenly distributed data set;

[0032] Step M4.3: Use the Hrnet pre-trained model based on the cityscapes dataset to train the uniformly distributed dataset with a preset number of times and learning rate so that the defect recognition rate marked in the test dataset reaches the preset value.

[0033] Preferably, step M5 includes: after the test is completed, exporting the training model, deploying the training model on the cold-rolled strip production line to perform online detection of defect recognition results, and when the defect recognition result meets the preset value, outputting the defect location; when the defect recognition result does not meet the preset value, repeating steps M1 to M5 until the defect recognition result meets the preset value.

[0034] According to the present invention, a cold-rolled strip grinding roll mark defect detection system based on deep learning is provided, comprising:

[0035] Module M1: Collect defect images reflected by the cold-rolled steel strip after being polished by the automatic polishing system;

[0036] Module M2: preprocessing the defect image and dividing the preprocessed defect image into images of a preset size;

[0037] Module M3: Manually annotate the segmented images of preset size and create a data set;

[0038] Module M4: Use the Hrnet pre-trained model based on the cityscapes dataset to train the dataset with a preset number of times and learning rate, so that the defect recognition rate marked in the dataset reaches the preset value, and obtain the trained model;

[0039] Module M5: Deploy the trained model on the cold-rolled strip production line to monitor defect recognition results online;

[0040] The Hrnet pre-training model is a parallel structure that maintains high-resolution representation at all times, rather than simply restoring high-resolution representation from low-resolution representation.

[0041] Preferably, the module M1 comprises: after the cold-rolled steel strip is polished by the matching automatic steel strip polishing system, the cold-rolled steel strip is irradiated with a light source at a preset angle, and the camera receives the defect image emitted by the cold-rolled steel strip at a preset angle;

[0042] The module M2 comprises:

[0043] Module M2.1: Cut the invalid edges of the defect image and obtain the ROI area;

[0044] Module M2.2: Use the equalization algorithm to equalize the brightness of the ROI area to obtain a ROI area with equalized brightness;

[0045] Module M2.3: Divide the ROI area with balanced brightness into images of preset sizes.

[0046] Preferably, the module M3 comprises:

[0047] Module M3.1: Manually annotate the segmented images of preset sizes and generate PNG labels in the form of legends to represent the roller print and background respectively;

[0048] Module M3.2: Divide the images and PNG labels into training set, validation set and test set according to the preset ratio;

[0049] Module M3.3: Create a dataset after classifying the images and PNG tags;

[0050] Module M3.4: Check the data set according to the preset requirements. If the check fails, repeatedly trigger modules M3.2 to M3.3 until the data set meets the preset requirements.

[0051] Preferably, the module M4 comprises:

[0052] Module M4.1: Load the Hrnet pre-trained model based on the cityscapes dataset;

[0053] Module M4.2: The loss function uses the dice loss function and the bce loss function to evenly distribute the categories of the roller print and the background, and obtain a evenly distributed data set;

[0054] Module M4.3: Use the HrNet pre-trained model based on the cityscapes dataset to train the uniformly distributed dataset with a preset number of times and learning rate, so that the defect recognition rate marked in the test dataset reaches the preset value.

[0055] Preferably, the module M5 includes: after the test is completed, exporting the training model, deploying the training model on the cold-rolled strip production line to perform online detection of defect recognition results, and when the defect recognition result meets the preset value, outputting the defect location; when the defect recognition result does not meet the preset value, repeatedly triggering modules M1 to M5 to execute until the defect recognition result meets the preset value.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] 1. The present invention can be generally applied to the intelligent identification system behind the automatic strip grinding system at the production exit of the cold rolling and acid rolling unit;

[0058] 2. After the present invention is put into use, the generation of roller mark defect circles can be effectively reduced, and the yield rate can be improved. At the same time, the automation of the roller mark recognition station can be realized, and the labor intensity of existing staff can be reduced;

[0059] 3. The present invention reduces the number of missed inspections and product downgrades of roller printing on the one hand, and reduces the workload of personnel on the other hand, bringing about the optimization of human resources, and is expected to generate benefits of more than 10 million yuan. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0061] Figure 1 is a flow chart of a defect detection method according to an embodiment of the present invention;

[0062] Figure 2 Schematic diagram of the process of step S2 in an embodiment of the present invention;

[0063] Figure 3 Schematic diagram of the process of step S3 in an embodiment of the present invention;

[0064] Figure 4 Schematic diagram of the process of step S4 in an embodiment of the present invention;

[0065] Figure 5 This is a schematic diagram of the HrNet model structure;

[0066] Figure 6 The surface image of the steel strip after grinding;

[0067] Figure 7 These are 5 measured pictures after deployment in an embodiment of the present invention; (the left side is the original picture, and the right side is the recognition result) DETAILED DESCRIPTION

[0068] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0069] According to the present invention, a method for detecting roller marks on cold-rolled strip after grinding based on deep learning is provided, comprising:

[0070] Step M1: collecting defect images reflected by the cold-rolled steel strip after being polished by the automatic polishing system;

[0071] Specifically, the step M1 comprises: after the cold-rolled steel strip is polished by the matching steel strip automatic polishing system, a light source is used to illuminate the cold-rolled steel strip at an angle of 40° to 50°, and a camera receives a defect image emitted by the cold-rolled steel strip at an angle of 5° to 15°;

[0072] Step M2: preprocessing the defect image, and dividing the preprocessed defect image into images of a preset size;

[0073] like Figure 2 As shown, the step M2 includes:

[0074] Step M2.1: Cut the invalid edge of the defect image and obtain the ROI area;

[0075] Step M2.2: Use an equalization algorithm to equalize the brightness of the ROI area to obtain an ROI area with balanced brightness; due to the light source scattering problem and the "vignetting" problem of the optical lens, the middle area of ​​the acquired strip image is bright and the two side areas are dark. In order to avoid defect recognition problems caused by uneven brightness, use an equalization algorithm to equalize the image brightness.

[0076] Step M2.3: Divide the ROI area with balanced brightness into images of size 1024*1000.

[0077] Step M3: Manually annotate the segmented images of preset size, generate PNG labels in the form of legends, classify the images and PNG labels, and create a data set;

[0078] like Figure 3 As shown, specifically, the step M3 includes:

[0079] Step M3.1: manually label the segmented images of preset size and generate PNG labels in the form of legends, where 1 represents roller printing and 0 represents background;

[0080] Step M3.2: Divide the images and PNG labels into a training set train, a validation set val, and a test set test according to a preset ratio; the validation set is mainly used to verify whether it is overfitting and to adjust training parameters, etc. The ratio of the number of images is train:val:test=8:1:1;

[0081] Step M3.3: After classifying the images and PNG labels, create a dataset;

[0082] Step M3.4: Check the data set according to the preset requirements. If the check fails, repeat steps M3.2 to M3.3 until the data set meets the preset requirements.

[0083] The steps for checking the data set are: 1. Get the maximum width and height of the image; 2. Calculate the maximum aspect ratio of the image; 3. Get the number of channels of the image; 4. Determine whether the label is a grayscale image; 5. Verify whether the size of the image and the annotation match; 6. Verify the format of the annotated image, and count the categories and number of pixels of the annotated image; 7. Count the format, category and number of pixels of each category on all annotated images; 8. Verify the annotated image and output the verification results; 9. Determine the relationship between the height and width of the set data and the images of the validation set and the test set; 10. Verify whether the format of the image is consistent with the setting; 11. Output the shape verification result; 12. Check whether the delimiter meets the requirements.

[0084] Step M4: Use the Hrnet pre-trained model based on the cityscapes data set to train the data set for a preset number of times and learning rate, so that the defect recognition rate marked in the data set reaches a preset value, and obtain a trained model;

[0085] like Figure 4 As shown, specifically, the step M4 includes:

[0086] To solve the problem of fewer available training images, load the HrNet pre-trained model based on the cityscapes dataset. The HrNet neural network has a unique parallel structure that can maintain high-resolution representation at all times, rather than just restoring high-resolution representation from low-resolution representation.

[0087] In image segmentation, the category prediction of each pixel is usually evaluated, and then all pixels are averaged. This is essentially learning each pixel in the image equally. This creates a problem. If there is an unbalanced representation of multiple categories in the image, then the training will be dominated by the most mainstream category. In the recognition of roller marks, the network will tend to learn about the background (cold-rolled strip grinding marks), reducing the network's ability to extract foreground targets (roller mark defects on the strip). The dice loss (dice coefficient loss) is calculated by dividing the intersection of the prediction and annotation by their total pixels. It considers all pixels of a category as a whole and calculates the proportion of the intersection in the total, so it is not affected by a large number of background pixels and can achieve better results.

[0088] In practical applications, dice loss is often used in combination with bce loss (binary cross entroy loss) to improve the stability of model training.

[0089] Step M4.1: Load the Hrnet pre-trained model based on the cityscapes dataset;

[0090] Step M4.2: The loss function uses the dice loss function and the bce loss function to evenly distribute the categories of the roller print and the background, and obtain a evenly distributed data set;

[0091] Step M4.3: Use the Hrnet pre-trained model based on the cityscapes dataset to train the uniformly distributed dataset with a preset number of times and learning rate, so that the accuracy of the roller marks in the test set, i.e. the roller marks in the test dataset, reaches the preset value. That is, the roller marks identified in the test set are similar to those manually marked. The consistency test Kappa of the test is greater than 0.8.

[0092] Test, the test set is tested. If the roller print defects identified by the test set are similar to those manually labeled and / or the consistency detection Kappa of the test is greater than 0.8, the test effect is ideal, otherwise the number of iterations training or the learning rate is redefined.

[0093] Step M5: deploy the trained model on the cold-rolled strip production line to monitor defect recognition results online;

[0094] The Hrnet pre-training model is a parallel structure that maintains high-resolution representation at all times, rather than simply restoring high-resolution representation from low-resolution representation.

[0095] Specifically, step M5 includes: after the test is completed, the training model is exported, and the training model is deployed in the cold-rolled strip production line to perform online detection of defect recognition results. When the defect recognition result meets the preset value, the defect position is output. When the defect recognition result does not meet the preset value, the number of samples and sample diversification are continuously increased, and steps M1 to M5 are repeated until the defect recognition result meets the preset value. Improve the model recognition rate.

[0096] According to the present invention, a cold-rolled strip grinding roll mark defect detection system based on deep learning is provided, comprising:

[0097] Module M1: Collect defect images reflected by the cold-rolled steel strip after being polished by the automatic polishing system;

[0098] Specifically, the module M1 includes: after the cold-rolled steel strip is polished by the matching steel strip automatic polishing system, the cold-rolled steel strip is irradiated with a light source at an angle of 40° to 50°, and the camera receives the defect image emitted by the cold-rolled steel strip at an angle of 5° to 15°;

[0099] Module M2: preprocessing the defect image and dividing the preprocessed defect image into images of a preset size;

[0100] The module M2 comprises:

[0101] Module M2.1: Cut the invalid edges of the defect image and obtain the ROI area;

[0102] Module M2.2: Use the equalization algorithm to equalize the brightness of the ROI area to obtain a ROI area with balanced brightness; due to the light source scattering problem and the "vignetting" problem of the optical lens, the middle area of ​​the acquired strip image is bright and the two side areas are dark. In order to avoid the defect recognition problem caused by uneven brightness, the equalization algorithm is used to equalize the image brightness.

[0103] Module M2.3: Divide the ROI area with balanced brightness into images of size 1024*1000.

[0104] Module M3: Manually annotate the segmented images of preset size, generate PNG labels in the form of legends, classify the images and PNG labels, and create a dataset;

[0105] Specifically, the module M3 includes:

[0106] Module M3.1: Manually label the segmented images of preset size and generate PNG labels in the form of legends, with 1 representing roller print and 0 representing background;

[0107] Module M3.2: Divide the images and PNG labels into a training set train, a validation set val, and a test set test according to a preset ratio; the validation set is mainly used to verify whether it is overfitting and to adjust training parameters, etc. The ratio of the number of images is train:val:test=8:1:1;

[0108] Module M3.3: Create a dataset after classifying the images and PNG tags;

[0109] Module M3.4: Check the data set according to the preset requirements. If the check fails, repeatedly trigger modules M3.2 to M3.3 until the data set meets the preset requirements.

[0110] The dataset inspection module is as follows: 1. Get the maximum width and height of the image; 2. Calculate the maximum aspect ratio of the image; 3. Get the number of channels of the image; 4. Determine whether the label is a grayscale image; 5. Verify whether the size of the image and the annotation match; 6. Verify the format of the annotated image, and count the categories and number of pixels of the annotated image; 7. Count the format, category and number of pixels of each category on all annotated images; 8. Verify the annotated image and output the verification results; 9. Determine the relationship between the height and width of the set data and the images of the validation set and the test set; 10. Verify whether the format of the image is consistent with the setting; 11. Output the shape verification result; 12. Check whether the delimiter meets the requirements.

[0111] Module M4: Use the Hrnet pre-trained model based on the cityscapes dataset to train the dataset with a preset number of times and learning rate, so that the defect recognition rate marked in the dataset reaches the preset value, and obtain the trained model;

[0112] Specifically, the module M4 includes:

[0113] To solve the problem of fewer available training images, load the HrNet pre-trained model based on the cityscapes dataset. The HrNet neural network has a unique parallel structure that can maintain high-resolution representation at all times, rather than just restoring high-resolution representation from low-resolution representation.

[0114] In image segmentation, the category prediction of each pixel is usually evaluated, and then all pixels are averaged. This is essentially learning each pixel in the image equally. This creates a problem. If there is an unbalanced representation of multiple categories in the image, then the training will be dominated by the most mainstream category. In the recognition of roller marks, the network will tend to learn about the background (cold-rolled strip grinding marks), reducing the network's ability to extract foreground targets (roller mark defects on the strip). The dice loss (dice coefficient loss) is calculated by dividing the intersection of the prediction and annotation by their total pixels. It considers all pixels of a category as a whole and calculates the proportion of the intersection in the total, so it is not affected by a large number of background pixels and can achieve better results.

[0115] In practical applications, dice loss is often used in combination with bce loss (binary cross entroy loss) to improve the stability of model training.

[0116] Module M4.1: Load the Hrnet pre-trained model based on the cityscapes dataset;

[0117] Module M4.2: The loss function uses the dice loss function and the bce loss function to evenly distribute the categories of the roller print and the background, and obtain a evenly distributed data set;

[0118] Module M4.3: Use the Hrnet pre-trained model based on the cityscapes dataset to train the uniformly distributed dataset with a preset number of times and learning rate, so that the accuracy of the roller marks in the test set, i.e. the roller marks identified in the test set, is close to the ones manually marked. The consistency test Kappa of the test is greater than 0.8.

[0119] The Hrnet pre-training model is a parallel structure that maintains high-resolution representation at all times, rather than simply restoring high-resolution representation from low-resolution representation.

[0120] Test, the test set is tested. If the roller print defects identified by the test set are similar to those manually labeled and / or the consistency detection Kappa of the test is greater than 0.8, the test effect is ideal, otherwise the number of iterations training or the learning rate is redefined.

[0121] Module M5: Deploy the trained model on the cold-rolled strip production line to monitor defect recognition results online;

[0122] The Hrnet pre-training model is a parallel structure that maintains high-resolution representation at all times, rather than simply restoring high-resolution representation from low-resolution representation.

[0123] Specifically, the module M5 includes: after the test is completed, the training model is exported, and the training model is deployed in the cold-rolled strip production line to perform online detection of defect recognition results. When the defect recognition result meets the preset value, the defect location is output. When the defect recognition result does not meet the preset value, the number of samples and sample diversification are continuously increased, and the execution of modules M1 to M5 is repeatedly triggered until the defect recognition result meets the preset value. Improve the model recognition rate.

[0124] The following preferred examples provide a more detailed description of the present invention:

[0125] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0126] The present invention provides a method for detecting roller mark defects of cold-rolled strip after grinding based on deep learning, comprising the following steps:

[0127] Step S1: Collecting original images

[0128] This step is equipped with an automatic strip grinding system. After grinding, the light source (model HLL-760-W) is used to illuminate the cold-rolled strip at an angle of 50°, and the camera (model LA-GM-04K08A) is used to capture defective images of the cold-rolled strip at an angle of 10°; 63 images containing roller mark defects are collected.

[0129] Step S2: pre-process the original image.

[0130] This step includes:

[0131] S21: Cut the invalid edges of the image and obtain the ROI area.

[0132] S22: Balanced image. Due to the problem of light source scattering, the middle area of ​​the acquired strip image is bright and the two sides are dark. In order to avoid the defect recognition problem caused by uneven brightness, the image brightness is balanced using an equalization algorithm.

[0133] S23: Divide the image into pictures of size 1024*1000.

[0134] Step S3: Process the image using an image segmentation algorithm to create a data set.

[0135] The configuration of the training server in this step: AMAX deep learning server, Intel(R) Xeon(R) Gold 5115CPU@2.40GHz, memory 128G, hard disk: 4T, GPU: Nvidia Titan XP 12G

[0136] This step includes:

[0137] S31: Manually annotate the preprocessed image and generate a PNG label in the form of a legend, where 1 represents roller printing and 0 represents background.

[0138] S32: Classify the images and PNG tags into training set train, validation set val (validation), and test set test. The main function of validation is to verify whether it is overfitting and to adjust the training parameters. The number of images is 40 for train, 14 for val, and 9 for test.

[0139] S33: Create a dataset.

[0140] S34: Data set check. If the check fails, the process returns to S32 for inspection.

[0141] Step S4: training.

[0142] This step includes:

[0143] S41: To solve the problem of limited available training images, load the Hrnet pre-trained model based on the cityscapes dataset, such as Figure 5 As shown in the figure, the hrNet neural network has a unique parallel structure that can maintain high-resolution representation at all times, instead of just restoring high-resolution representation from low-resolution representation.

[0144] S42: The loss function selects dice loss + bce loss.

[0145] The algorithm used in this step is to solve the problem of imbalanced categories between roller prints and backgrounds.

[0146] S43: The number of training iterations is set to 600.

[0147] Step S5: testing.

[0148] The results after this training and testing are:

[0149] #testimage=9acc=0.9998IoU=0.8431

[0150] Category IoU:[0.9998(background)0.6864(roller print)]

[0151] Category Acc:[1.0000(background)0.7198(roller print)]

[0152] Kappa: 0.8140

[0153] Step S6: Deployment

[0154] After the test, the training model is exported and deployed in the cold-rolled strip production line for online detection. The accumulated roll imprint images are added back to the training starting from step S2, and the number and diversity of samples are continuously increased to improve the model recognition rate, such as Figure 6 shown.

[0155] Configuration of the model deployment computer:

[0156] DELL Precision 5820

[0157] i7-9800X 3.8GHz, memory 32G, hard disk 4T, GPU Nvidia RTX 2080TI

[0158] There are 5 pictures of the actual test after deployment, and the test results are as follows Figure 7 .

[0159] Those skilled in the art know that, in addition to implementing the system, device and its various modules provided by the present invention in a purely computer-readable program code, it is entirely possible to implement the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers and embedded microcontrollers by logically programming the method steps. Therefore, the system, device and its various modules provided by the present invention can be considered as a hardware component, and the modules included therein for implementing various programs can also be considered as structures within the hardware component; the modules for implementing various functions can also be considered as both software programs for implementing the method and structures within the hardware component.

[0160] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A method for detecting roller marks on cold-rolled strip after grinding based on deep learning, characterized in that: include: Step M1: collecting defect images reflected by the cold-rolled steel strip after being polished by the automatic polishing system; Step M2: preprocessing the defect image, and dividing the preprocessed defect image into images of a preset size; Step M3: manually label the segmented images of preset size and create a data set; Step M4: Use the Hrnet pre-trained model based on the cityscapes data set to train the data set for a preset number of times and learning rate, so that the defect recognition rate marked in the data set reaches a preset value, and obtain a trained model; Step M5: deploy the trained model on the cold-rolled strip production line to monitor defect recognition results online; The Hrnet pre-training model is a parallel structure that maintains high-resolution representation at all times, not just by restoring high-resolution representation from low-resolution representation; The step M2 comprises: Step M2.1: Cut the invalid edge of the defect image and obtain the ROI area; Step M2.2: Use an equalization algorithm to equalize the brightness of the ROI area to obtain a ROI area with equalized brightness; Step M2.3: Segment the ROI area with balanced brightness into images of preset sizes; The step M3 comprises: Step M3.1: manually label the segmented images of preset size and generate PNG tags in the form of legends, representing the roller print and the background respectively; Step M3.2: Divide the images and PNG labels into training set, validation set and test set according to the preset ratio; Step M3.3: After classifying the images and PNG tags, create a dataset based on the training set, validation set, and test set; Step M3.4: Check the data set according to the preset requirements. If the check fails, repeat steps M3.2 to M3.3 until the data set meets the preset requirements. Among them, checking the data set includes: obtaining the maximum width and height of the image; calculating the maximum aspect ratio of the image; obtaining the number of channels of the image; determining whether the label is a grayscale image; verifying whether the size of the image and the annotation match; verifying the format of the annotated image, counting the categories and number of pixels of the annotated image; counting the format, category and number of pixels of each category on all annotated images; verifying the annotated image and outputting the verification results; determining the relationship between the image height and width of the set data and the validation set and the test set; verifying whether the image format is consistent with the setting; outputting the shape verification results; checking whether the delimiter meets the requirements; The step M4 comprises: Step M4.1: Load the Hrnet pre-trained model based on the cityscapes dataset; Step M4.2: The loss function uses the dice loss function and the bce loss function to evenly distribute the categories of the roller print and the background, and obtain a evenly distributed data set; Step M4.3: Use the Hrnet pre-trained model based on the cityscapes dataset to train the uniformly distributed dataset with a preset number of times and learning rate, so that the defect recognition rate marked in the test dataset reaches the preset value; The step M5 includes: after the test is completed, exporting the training model, deploying the training model on the cold-rolled strip production line to perform online detection of defect recognition results, and when the defect recognition result meets the preset value, outputting the defect position; when the defect recognition result does not meet the preset value, repeating steps M1 to M5 until the defect recognition result meets the preset value.

2. The method for detecting roller marks after grinding of cold-rolled strip based on deep learning according to claim 1, characterized in that: The step M1 includes: after the cold-rolled steel strip is polished by the matching automatic steel strip polishing system, the cold-rolled steel strip is irradiated with a light source at a preset angle, and the camera receives the defect image emitted by the cold-rolled steel strip at a preset angle.

3. A cold-rolled strip grinding roller mark defect detection system based on deep learning, characterized in that: include: Module M1: Collect defect images reflected by the cold-rolled steel strip after being polished by the automatic polishing system; Module M2: preprocessing the defect image and dividing the preprocessed defect image into images of a preset size; Module M3: Manually annotate the segmented images of preset size and create a data set; Module M4: Use the Hrnet pre-trained model based on the cityscapes dataset to train the dataset with a preset number of times and learning rate, so that the defect recognition rate marked in the dataset reaches the preset value, and obtain the trained model; Module M5: Deploy the trained model on the cold-rolled strip production line to monitor defect recognition results online; The Hrnet pre-training model is a parallel structure that maintains high-resolution representation at all times, not just by restoring high-resolution representation from low-resolution representation; The module M2 comprises: Module M2.1: Cut the invalid edges of the defect image and obtain the ROI area; Module M2.2: Use the equalization algorithm to equalize the brightness of the ROI area to obtain a ROI area with equalized brightness; Module M2.3: Split the ROI area with balanced brightness into images of preset sizes; The module M3 comprises: Module M3.1: Manually annotate the segmented images of preset sizes and generate PNG labels in the form of legends to represent the roller print and background respectively; Module M3.2: Divide the images and PNG labels into training set, validation set and test set according to the preset ratio; Module M3.3: After classifying the images and PNG tags, create datasets based on training, validation, and test sets; Module M3.4: Check the data set according to the preset requirements. If the check fails, repeat Modules M3.2 to M3.3 until the data set meets the preset requirements; create a data set based on the training set, validation set, and test set; Among them, checking the data set includes: obtaining the maximum width and height of the image; calculating the maximum aspect ratio of the image; obtaining the number of channels of the image; determining whether the label is a grayscale image; verifying whether the size of the image and the annotation match; verifying the format of the annotated image, counting the categories and number of pixels of the annotated image; counting the format, category and number of pixels of each category on all annotated images; verifying the annotated image and outputting the verification results; determining the relationship between the image height and width of the set data and the validation set and the test set; verifying whether the image format is consistent with the setting; outputting the shape verification results; checking whether the delimiter meets the requirements; The module M4 comprises: Module M4.1: Load the Hrnet pre-trained model based on the cityscapes dataset; Module M4.2: The loss function uses the dice loss function and the bce loss function to evenly distribute the categories of roller prints and backgrounds, thus obtaining a evenly distributed data set; Module M4.3: Use the Hrnet pre-trained model based on the cityscapes dataset to train the uniformly distributed dataset with a preset number of times and learning rate, so that the defect recognition rate marked in the test dataset reaches the preset value; The module M5 includes: after the test is completed, exporting the training model, deploying the training model on the cold-rolled strip production line to perform online detection of defect recognition results, and when the defect recognition result meets the preset value, outputting the defect location; when the defect recognition result does not meet the preset value, repeatedly executing modules M1 to M5 until the defect recognition result meets the preset value.

4. The cold-rolled strip post-grinding roller mark defect detection system based on deep learning according to claim 3 is characterized in that: The module M1 includes: after the cold-rolled steel strip is polished by the matching automatic steel strip polishing system, the cold-rolled steel strip is irradiated with a light source at a preset angle, and the camera receives the defect image emitted by the cold-rolled steel strip at a preset angle.

Images